Hand real GTM work to an agent that touches real files, and prove the output is
right before anyone acts on it. Every module ends in a file you keep and use at work.
Claude and Claude Cowork are products of Anthropic. This course is independent and is
not affiliated with, endorsed by, or sponsored by Anthropic. A paid Claude plan is required.
Cowork 101 · no setup needed
Cowork 101 / about 5 min
Concept
What Cowork actually is
You already know Claude as a chat window: you ask, it answers in text, and then it is your job to
do something with that text.
Cowork is different in one simple way. You give it a job, and it hands you back the finished
thing. Not advice about a spreadsheet: the cleaned spreadsheet. Not tips for a summary: the
summary, as a document, saved where you can find it. That is also the rule for choosing between
them: if you want words back, chat is enough. When the result should be a thing saved in a
folder, that is Cowork.
Eight short lessons, about 50 minutes end to end. In a hurry? Lessons 1, 2 and 6 are a useful
20 minutes on their own; come back for the rest when the first win lands.
It does all of this inside one folder you choose. It can read what is there and save new things
there. It is built to work only inside that folder, so the folder is the boundary you control:
if Cowork should not see something, it does not go in. And here is the part that makes everything
easy: you never have to create files or folders yourself. You ask, and Cowork makes them.
Try this now (60 seconds)
Open Cowork in the Claude app.
When it asks which folder to work in, make a new one right there in the chooser if it lets
you, or create one wherever you normally keep files and pick it. Call it Cowork sandbox.
That is the only folder you will ever need for this course.
Drag any work file into it: an export, some notes, a report. Pick one you would be
comfortable pasting into any online tool. That habit applies to every AI tool, and it
starts now.
Paste this:
Look at what is in this folder and tell me, in plain language:
what each file is, what is inside it, and one useful thing you
could do with it. Do not change anything yet.
What you should see: a short description of your own file, plus a suggestion or two. Nothing
was changed. That is the whole trick: it read your folder, like a colleague looking over your
shoulder. If it describes the wrong folder or says the folder is empty, say: "Look again at the
folder you are working in and list every file by name." A short correction like that is normal,
and it beats starting over.
Worth knowing: Cowork comes with every paid Claude plan. It is not on the free plan. It lives
in the Claude desktop app on Mac and Windows, and it has been arriving on web and mobile in
stages, so where you find it depends on your plan and platform. If you do not see it, update the
app and check your plan before assuming anything is broken.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 7 min
Do it
Prep me for this meeting
The fastest way to feel what Cowork is for: hand it the prep you were about to do by hand.
You do not need to prepare anything. Paste what you have straight into the chat, and Cowork will
file it away and do the rest.
Try this now
Pick a real meeting from this week.
Open Cowork in your sandbox folder.
Paste the prompt below, change the name, and then paste the meeting invite, the email
thread, or your old notes directly underneath it. If you have a related file, drag it into
the folder too. Either way works. Leave out anything too sensitive for a cloud tool; the
brief works fine without it.
I have a meeting with [company or person] this week.
First, save everything I pasted below this message as a note in
this folder, so we do not lose it.
Then read everything here about them and write me a one-page
brief, saved as a note called "meeting brief":
- who they are and where things stand
- the three things most worth raising
- anything that looks unresolved or overdue
- what you could NOT find, so I know what is missing
No meeting this week? Practice on this instead. Use the name Meridian Supply in the prompt, then
paste this underneath it. Every word of it is invented, and it is enough to see the whole flow.
PRACTICE MATERIAL
Invite: quarterly review with Meridian Supply, Thursday 14:00,
45 minutes. Attendees: you and Dana Okafor, Head of Ops.
Email from Dana, last week: "Before we meet, can we cover the
delayed onboarding and where we stand on the renewal? Still
waiting on the usage report you mentioned."
Your old note: "Meridian renewal due next quarter. Onboarding
stalled in week two, waiting on their IT. Promised a usage
report. Dana has asked twice."
What you should see: two new notes appear in your folder, made by Cowork, not by you: your
pasted material, filed, and the brief. Open the brief. The best part is the last section, what it
could not find, because that tells you what to check before you walk in. If only one note
appears, say: "You saved the brief. Now also save what I pasted as its own note." Steering with a
short correction like this is normal, and it works better than starting over.
Worth knowing: the brief is only as good as what you give it. If the answer looks thin, the
fix is to paste in one more email, not to type a longer prompt. And when you run this on a real
meeting where a wrong guess would sting, add one line to the prompt: "Before you start, ask me
any questions you need, one at a time, and wait for my answers." Answering two short questions
beats writing a long prompt, every time.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 8 min
Do it
Clean up any messy export
Every CRM, every tool, every list you have ever downloaded: the export is always a bit of a mess.
Duplicates, blanks, three spellings of the same country. Cleaning it by hand is an hour of your
life every time.
Two rules make this safe, and they are the only rules you need today:
Ask it to find problems before it fixes anything.
Fixes go into a new copy that Cowork creates. Your original is never touched.
Try this now
Drag any export into your sandbox folder. A contact list, a deal list, anything from your
CRM or spreadsheet tool.
No export handy? Have Cowork make one. Paste this first, then continue from step 2:
Create a practice contact list in this folder: about 30 rows
with name, company, country and email. Make it realistically
messy: a few duplicates, some blank emails, and three
spellings of the same country. Tell me when it is ready.
Paste prompt one. Read the answer.
If the list of problems looks right, paste prompt two.
Look at the spreadsheet I just added and list the problems you
can see: duplicates, blanks, inconsistent spellings, anything
that looks wrong. Show me a few examples of each. Do not change
anything.
Fix only the safe, obvious problems: spacing, casing, consistent
spellings. Put the result in a NEW copy called "cleaned", next
to the original, and leave my original file exactly as it is.
Leave anything you are unsure about unchanged too, and give me a
short list of what you fixed and what you skipped.
What you should see: first an honest list of what is wrong with your data. Then a new, cleaner
copy sitting next to your untouched original, plus a note about what it deliberately left alone.
If it starts fixing things on the first prompt, say: "Stop. Do not change anything yet. Only list
the problems with examples." If the cleaned copy never appears, say: "Put the fixes in a new copy
called cleaned, next to the original, and confirm the original is untouched."
Worth knowing: the skipped list is not a failure, it is the point. A tool that fixes only what
it is sure about, and tells you the rest, is a tool you can trust with real work. Two more things.
Cowork shares your plan's usage allowance with regular Claude chat, and multi-step jobs use more
of it. If you hit a limit mid-run, it comes back after a break, and a smaller file gets you there
sooner. And if a file is the team's live working copy in a shared drive, clean a copy in your
sandbox, not the original.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 7 min
Do it
From messy notes to a follow-up email
You take notes in a hurry. Then the follow-up email sits unwritten for two days, and the value of
the meeting quietly evaporates.
This one has a golden rule, and it is the most important sentence in this whole course:
it drafts, you send. Nothing goes to another human being without your eyes on it first.
Try this now
Open Cowork in your sandbox folder.
Paste the prompt below, then paste your meeting notes directly underneath it. Messy is fine.
Messy is normal. If your notes are already a file, drag it in instead and say so.
Below this message are my meeting notes. Save them as a note in
this folder first, so they are kept.
Then create two more notes for me:
1. "next steps": every action item you can find, each with who
owns it and by when, if the notes say. If the notes do not
say, write "not stated" rather than guessing.
2. "follow-up draft": a short, friendly follow-up email
confirming what was agreed. Only include things that are
actually in the notes. This is a DRAFT for me to edit, not
something to send.
No notes of your own? Use these, invented for practice:
PRACTICE NOTES
call w Meridian, dana + priya
onboarding blocked, THEIR it dept, ticket open 9 days, chase
dana wants usage report before renewal talk
priya: budget review in sept, no pricing changes for now
we said: report by friday, intro to our support lead
next call in 2 wks, dana sending dates
What you should see: your notes filed, plus two new notes, all made by Cowork. If you get one
long note instead of the two you asked for, say: "Split that into two separate notes, one called
next steps and one called follow-up draft." Check the next
steps against your memory of the meeting. Then open the draft, change the two sentences that do
not sound like you, and send it yourself from your own email. Count the "not stated" entries:
each one is a real gap in your notes, found before it cost you anything.
Worth knowing: "not stated" is the honest answer when your notes have a gap. If the draft ever
contains something you do not remember agreeing to, that is your cue to check the notes, not to
trust the draft.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 8 min
Do it
Your Monday morning routine
Everything so far was a one-off. The real payoff is a job that comes back every week.
Pick the weekly thing you dread. For many people it is the Monday status update: what moved, what
stalled, what needs attention. Pipeline, projects, hiring, cash: the shape is the same whatever
you run. Fifteen minutes of copying numbers into prose, every single week.
The setup takes one prompt. Cowork builds everything itself.
Try this now
Open Cowork in your sandbox folder and paste this:
Set up my Monday routine. Create a folder here called "monday",
and inside it save a note called "instructions" containing
exactly this:
Every time I ask for my Monday update: read the files in this
folder, and draft a note called "monday update" with three
sections: what moved since last week, what looks stuck, and
the three things most worth my attention. Keep it under 300
words. Under every number, say which file it came from. This
is a draft for me to review. Never send anything.
Then tell me it is ready.
Drag this week's exports into the new monday folder: your pipeline, your task list,
whatever your update is built from. No exports this week? Ask Cowork for a practice
pipeline file first, the same move as in lesson 3, and run the routine on that.
Say: "Read the instructions note in the monday folder and do it."
What you should see: your Monday update, drafted, with every number traceable to a file. Next
Monday, drag in the fresh exports and say the same one line. If a number appears without a file
under it, say: "Under every number, name the file it came from." The instructions note already
requires this; hold it to that.
Worth knowing: Cowork starts every conversation fresh. It does not remember last week. That is
why it wrote your instructions into a note in the folder, where they never get lost. You can also
ask it to run on a schedule; treat that as a convenience and glance that it actually ran, rather
than betting your Monday on it.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 6 min
Concept
Three habits that keep you safe
You have now done real work with Cowork. Before you make it a daily tool, three habits. They take
no extra time, and they are the difference between a tool you trust and a mess you clean up.
1. Give it its own folder, never your whole computer.
Everything you did in this course happened inside one sandbox folder. Keep it that way. If a file
is not something Cowork should see or touch, it does not go in the folder. Simple as that.
2. Ask where every number came from.
Add one line to any prompt that produces numbers: "under every number, say which file it came
from." A number with a source takes ten seconds to check. A number without one is just a claim,
however confident it sounds.
3. It drafts, you send.
Emails, messages, anything a customer or colleague will read: Cowork writes the draft, you read
it, you send it. No exceptions. This one habit removes almost all of the ways this could ever
embarrass you.
And three things not to hand over at all: decisions (pricing, discounts, who to hire),
promises to customers, and anything you would have no way to check. If you could not tell whether
the answer is wrong, it is not a job for Cowork. It is a job for you.
Two smaller habits, both about not wasting your allowance or your patience. First: thinking out
loud is a chat job. Talk a fuzzy task through in regular Claude chat, which uses less of the
allowance the two share, and come to Cowork once you can say what the finished thing should be.
Second: keep your setup as small as lesson 5's, one folder and one instructions note. You will
meet guides that recommend elaborate folder systems and profile files describing you, your
voice, and your rules. People who ran those setups found the files go stale and quietly mislead,
and most walked them back. A setup you cannot maintain is worse than none.
One honest note on scope. This course kept Cowork inside one folder on purpose. The product does
more: it can connect to tools like your email and shared drives, work inside your browser, and
take on scheduled tasks. All of that is worth having only once the three habits are automatic,
which is why the habits came first.
Two short lessons remain: every prompt from this course on one page, and a final check that
you can do all of this with the course closed.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 4 min
Concept
Every prompt on one page
Every prompt from this course in one place, each with its copy button. Bookmark this page.
When a prompt has moved into your muscle memory, stop pasting it and say it in your own
words. That is the goal.
Look around, change nothing (lesson 1)
Look at what is in this folder and tell me, in plain language:
what each file is, what is inside it, and one useful thing you
could do with it. Do not change anything yet.
Meeting brief (lesson 2)
I have a meeting with [company or person] this week.
First, save everything I pasted below this message as a note in
this folder, so we do not lose it.
Then read everything here about them and write me a one-page
brief, saved as a note called "meeting brief":
- who they are and where things stand
- the three things most worth raising
- anything that looks unresolved or overdue
- what you could NOT find, so I know what is missing
Diagnose the mess (lesson 3, prompt one)
Look at the spreadsheet I just added and list the problems you
can see: duplicates, blanks, inconsistent spellings, anything
that looks wrong. Show me a few examples of each. Do not change
anything.
Fix only the safe problems (lesson 3, prompt two)
Fix only the safe, obvious problems: spacing, casing, consistent
spellings. Put the result in a NEW copy called "cleaned", next
to the original, and leave my original file exactly as it is.
Leave anything you are unsure about unchanged too, and give me a
short list of what you fixed and what you skipped.
Make me practice data (lesson 3)
Create a practice contact list in this folder: about 30 rows
with name, company, country and email. Make it realistically
messy: a few duplicates, some blank emails, and three
spellings of the same country. Tell me when it is ready.
Notes to next steps and a draft (lesson 4)
Below this message are my meeting notes. Save them as a note in
this folder first, so they are kept.
Then create two more notes for me:
1. "next steps": every action item you can find, each with who
owns it and by when, if the notes say. If the notes do not
say, write "not stated" rather than guessing.
2. "follow-up draft": a short, friendly follow-up email
confirming what was agreed. Only include things that are
actually in the notes. This is a DRAFT for me to edit, not
something to send.
Set up the Monday routine (lesson 5)
Set up my Monday routine. Create a folder here called "monday",
and inside it save a note called "instructions" containing
exactly this:
Every time I ask for my Monday update: read the files in this
folder, and draft a note called "monday update" with three
sections: what moved since last week, what looks stuck, and
the three things most worth my attention. Keep it under 300
words. Under every number, say which file it came from. This
is a draft for me to review. Never send anything.
Then tell me it is ready.
Run the Monday routine (lesson 5)
Read the instructions note in the monday folder and do it.
The one line that keeps numbers honest
Under every number, say which file it came from.
The one line that stops guessing (lesson 2)
Before you start, ask me any questions you need, one at a time,
and wait for my answers.
And the three habits, in full: its own folder, never your whole computer. A source under
every number. It drafts, you send.
One last thing you may have noticed: nothing on this page is long. That is the real lesson
of the whole course. A short prompt pointed at real files beats a long prompt every time,
because the work lives in what Cowork can read, not in your wording.
Checked against Cowork in August 2026.
Cowork 101 · no setup needed
Cowork 101 / about 5 min
Do it
The exit check
You are done with this course when you can say yes to five statements, with the course
closed:
I can point Cowork at one folder and know why it gets nothing more. (Lesson 1)
I can turn pasted material into a brief that admits what it could not find. (Lesson 2)
I can make it diagnose before it fixes, into a new copy, original untouched. (Lesson 3)
I can get next steps and a follow-up draft that never sends itself. (Lesson 4)
My Monday routine sits in a folder, in instructions Cowork wrote, re-run with one
line. (Lesson 5)
A no is not a failure. It is the number of the lesson to redo, and each takes under ten
minutes.
Then make it real before the week ends.
Try this now
Open Cowork in your sandbox folder and paste this:
Ask me these three questions, one at a time, and wait for each
answer:
1. Which task from my real work will I hand to Cowork next week?
2. On which day?
3. Which habit am I most likely to forget: own folder only, a
source under every number, or drafts only?
Then save my answers as a note called "my first week", and end
the note with one line: the habit I named, written as a reminder.
What you should see: a short interview, then a note in your folder that you produced
by answering three questions. If it answers the questions itself instead of asking you, say:
"Do not answer for me. Ask me one question at a time and wait for my reply." That note is the point of a 101: not what you learned, what
you do with it next week.
There is no certificate for this track, on purpose. A checklist you passed yourself is
worth more than a badge nobody checked. The full course ends with a real assessment,
graded on the work you produce, and that one is worth showing.
Where to go from here. You now use Cowork the way a careful generalist does. If your
work is revenue work, the advanced course takes you much further: cleaning a whole CRM
safely, weekly pipeline reviews that catch what dashboards miss, account research that
never invents facts, and a way of checking an agent's work in minutes instead of redoing
it. It is the other button in the top bar, and its first module takes about 90 minutes.
If your work is not revenue work, you still leave with the method that carries everywhere:
one folder, a source under every number, drafts only.
Checked against Cowork in August 2026.
Getting started
Setup, about 10 minutes
Start here
Start here
Claude Cowork for Revenue Teams. This file is the one you read first. It gets you from nothing
to finishing Module 1 with a file you keep.
Everything else in course/ is either content you will be sent to, or material for whoever
maintains the course. You do not need to read it in order to take the course.
Before you start, 10 minutes
1. You need a paid Claude plan. Cowork is included on every paid tier and is not available on
the free tier (reference/COWORK-FACTS.md §4). Pro is enough to complete the course, and heavy runs
hit usage limits faster there than on Max (§6).
2. Install and open Cowork. Desktop on macOS or Windows, or the web and mobile beta (§3). The
course is written for desktop, because Module 2 configures a folder on your machine.
3. Make the workspace. One folder, and everything the course does happens inside it:
revenue-workspace/
00-course/ the course files, including the practice data
10-input/ exports you drop in
20-working/ scratch, safe to delete
30-output/ what runs produce
40-reference/ the briefs and policies you write and keep
ops/ scheduled tasks and the run register
Module 2 explains why each folder exists and what may write where. Create the shell now.
4. Get the practice data. Copy course/labs/data/ into 00-course/. It is about 1.5 MB: a
fictional B2B SaaS company called Talvera, with accounts, contacts, opportunities, activities,
weekly pipeline snapshots, tickets, leads, usage, and 19 unstructured call notes and QBR documents.
The data is deliberately broken in documented ways. That is the point: you cannot learn to verify an
agent's output on data where everything is already correct.
5. Do not read labs/data/PLANTED-ISSUES.md yet. It is the answer key. It is released to you
per lab, after you have attempted that lab. Reading it early is allowed, it just marks that lab's
practice numbers key-seen and excludes them from your record (curriculum/ASSESSMENT-MODEL.md
§2.1). You are only cheating yourself, but the record says so honestly.
The shape of the course
Eight modules. Each is lessons, then one lab, then a short checkpoint. Every module ends with a
file you keep and use at work. Roughly 22 to 24 hours in total, self-paced.
Modules
What it is
Foundations
M1 to M4
Everyone takes these at full depth. What Cowork is, how to fence it, what to hand it, and how to brief and check it
Applications
M5 to M8
The revenue work: CRM hygiene, pipeline, demand and targeting, retention and repeatable operations
Your track
3 extra labs
RevOps, AE/SDR, marketing or CS. Pick one at the end of M4
Capstone
3 to 5 hours
One workflow, end to end, on your own data
Do this now: Module 1
Read M1 L1.1, Which of your Monday can leave your hands (12 minutes).
Read M1 L1.2, Your first end-to-end run (15 minutes), and actually run it. This is the
first time you point Cowork at a folder and get back a file rather than a chat reply.
Read M1 L1.3, The four walls, before you build against them (13 minutes). The limits, up
front, before you build habits that depend on them not existing.
Run LAB-01 (labs/LAB-01/README.md). Part A on the shipped data, Part B on your own.
Take checkpoint CP-01 (assessment/CHECKPOINTS/CP-01.md).
You finish Module 1 with cowork-fit-log.md: an inventory of your recurring work with a first call
on each item. It is the document the rest of the course builds on.
Lessons are readable on the site under /course, or as markdown in lessons/.
How you know you are doing it right
Every lab states a manual baseline, a target, and a verification cost. Fill in your own numbers.
If your verification takes longer than the work you saved, the lab says so, and so should you. Three
labs in this course conclude that the saving is small; that is an honest finding, not a failure.
Labs are scored two ways. A four-criterion rubric on the artifact (Correctness, Completeness,
Reusability, Safety), published before you start, with worked examples at each quality level in each
lab's anchors/. And on some labs a detection scorecard against the planted-defect key, which
reports what you caught, what you flagged wrongly, and what you correctly could not resolve.
Finding a defect we did not plant is credited, not penalised. Out-of-key findings go to
adjudication. An item you write to your exceptions file is an admission rather than an assertion and
scores a half catch, because telling an agent to declare what it could not resolve is the single
highest-yield habit in this course.
Two things to know before you trust anything here
The product moves. Every lesson carries a last_verified date and lists the claims most likely
to go stale, with where to re-check them. If you are reading this long after that date, re-check
before relying on a specific behaviour. Every product claim in the course traces to a section of
reference/COWORK-FACTS.md, which marks unverified claims as unverified rather than smoothing them.
Cowork cannot do some things you will want it to do. No memory between sessions. No event
triggers. Scheduling that is convenience automation rather than a guaranteed cron. One usage pool
shared with chat. An enterprise audit gap. The course teaches each of these as a module with an
exercise, not as a disclaimer, because every one of them will otherwise bite you in week three.
Where things live
Path
What it is
lessons/M1 to M8
The lessons. Start at M1/L1.md
labs/LAB-01 to LAB-08
The labs, with hints, rubrics, anchors and solutions
labs/TRACK-LABS/
Twelve role-specific labs, three per track
labs/data/
The Talvera practice dataset. Copy this into your workspace
assessment/
Checkpoints, the capstone, rubrics, the reviewer guide
curriculum/
Syllabus, outcome map, the assessment model. Read if you want the design rationale
reference/COWORK-FACTS.md
The sourced product facts every claim traces to
Claude and Claude Cowork are products of Anthropic. This course is independent and is not
affiliated with, endorsed by, or sponsored by Anthropic.
Getting started
Talvera, a fictional B2B SaaS company
The practice data, up close
Every lab runs on this. It is deliberately broken in documented ways, because you cannot learn
to verify an agent's work on data where everything is already correct. Here are real rows.
accounts.csv 308 rows
account_id
account_name
segment
country
arr_usd
health_score
csm
ACC-02801
Quinlan Care Homes Corp
SMB
United States
16000.0
82
Helena Zabala
ACC-02802
Ziegfeld Transport Inc
SMB
Canada
24000.0
49
Tobias Costa
ACC-02803
Zennor Stores LLC
SMB
United States
ACC-02804
Wexley Facilities Corp
SMB
United States
ACC-02805
Vanbrugh Grocers GmbH
SMB
Germany
18360.0
8
Tomasz Jonker
ACC-02806
Uphall Distribution Limited
SMB
United Kingdom
34290.0
83
Niamh Ricci
opportunities.csv 528 rows
opportunity_id
account_id
stage
amount
currency
forecast_category
close_date
OPP-006941
ACC-02984
Closed Won
34000.0
EUR
Closed
2022-11-15
OPP-006877
ACC-02897
Closed Lost
23000.0
GBP
Omitted
2022-11-20
OPP-007486
ACC-02969
Closed Lost
22000.0
USD
Omitted
2022-12-05
OPP-007211
ACC-02930
Closed Lost
79000.0
GBP
Omitted
2022-12-10
OPP-007488
ACC-02803
Closed Lost
13000.0
USD
Omitted
2022-12-24
OPP-006886
ACC-03163
Closed Won
57000.0
USD
Closed
2022-12-31
Six defect classes are deliberately hard to spot. One CSM recorded health scores on a
1 to 10 scale while everyone else used 1 to 100. A weekly pipeline snapshot is a byte copy of the
week before. One call note disagrees with the CRM on amount, close date and who signs, and closes
by claiming the record was updated.
Last verified: 2026-08-02.Product claims: FACTS §1, §2, §3, §4, §5.1, §5.2, §5.10, §6, §7.1, §7.2, §7.3.
Dataset: Talvera, seed 20260801. Every identifier and value quoted in these lessons was read
out of the shipped files in your practice pack's data/ folder and checked against
answer-keys/PLANTED-ISSUES.md or data/DATA-DICTIONARY.md. None of it was copied from another document.
The module at a glance
It is Monday. Three ad hoc reporting requests are sitting in Slack, an account executive wants to
know whether two records are the same company, a VP wants to know why the forecast moved, and
marketing wants the lead list segmented by a definition nobody has written down.
A chat assistant writes well and cannot open the export in your downloads folder. This module is
about which parts of that Monday can leave your hands, what happens when they do, and the four
things the product structurally cannot do, told to you before you spend an hour designing around
one of them.
You will run one real end-to-end task on shipped data in Lesson 2. You will hit the product's
hardest limit on purpose in Lesson 3.
Lessons
#
ID
Title
Time
Reading
Doing
Carries
L1
L1.1
Which of your Monday can leave your hands
12 min
9
3
Discrimination pair
L2
L1.2
Your first end-to-end run
15 min
11
4
Worked exercise
L3
L1.3
The four walls, before you build against them
13 min
10
3
Failure lane
Lesson total: 40 minutes. The M1 lesson budget in PEDAGOGY.md §3.1 is 35, and the tolerance
is plus or minus 15 percent (§6.9), so this sits at the top of the band rather than inside it.
Reading minutes are measured from the shipped text, not estimated: prose at 200 words per minute,
tables and code blocks faster because they are scanned. No lesson exceeds the 20-minute hard cap.
The module also carries LAB-01 (45 min), the exercise sheet in labs/LAB-01 of your practice
pack (start with its README), and CP-01 (10 min), the module quiz in the
interactive app, built in Parts 3 and 4.
Every ## Check yourself section carries a collapsed answer block: a response per item, per-option
lines wherever the item is closed multiple choice, and items that genuinely run on the learner's own
list marked Open so nobody hunts for a key that should not exist. Those blocks are counted
inside the reading figure above even though they are collapsed and consulted one item at a time,
which is the conservative accounting.
Module-level sections appear once each, per PEDAGOGY.md §3.1a and §6.9b. The map above is the
declaration of where.
Outcomes
ID
You can
O1.1
Given a revenue task, state whether it needs chat, Cowork or a human, and name the deciding factor
O1.2
Describe the Cowork task loop and name the two points at which you steer it
O1.3
Name the documented limits and, for a proposed workflow, state which limit it collides with
O1.4
Classify a revenue task as read-only, draft-only or writes-to-system
O1.5
Run one end-to-end task on a supplied folder and describe it in task-loop terms
O1.6
State what the 4Ds cover, what TRACK adds, and why the addition exists
What you keep
cowork-fit-log.md (template in templates/M1/cowork-fit-log.md in your practice pack).
A one-page inventory of your recurring revenue work: one row per task, with rough minutes per week,
whether it needs files or connectors, its blast-radius class, and a first call of human, chat or
Cowork. Twelve to twenty-five rows.
You build it before you know the product well. That is deliberate. Module 3 makes you re-score it
with a real instrument, and the tasks that move are the lesson.
Run evidence: first-run/account-snapshot.md, the output of Lesson 2, kept as the before
picture.
Before you start
A paid Claude plan. Cowork is included in every paid plan and is not in the free tier (FACTS §4).
The shipped dataset downloaded to a folder you can point at.
Comfort with a CRM and a spreadsheet. No terminal, no API key, no engineering help.
Where this goes next
Module 2 turns the folder decision you made casually in Lesson 2 into a boundary you make on
purpose. Do not skip to Modules 5 to 8: they assume the workspace from Module 2 and the brief
method from Module 4.
Module 1 · Cowork as a revenue work surface
M1 lesson / about 12 min
Concept
Which of your Monday can leave your hands
Do this nowList your recurring week60s
Open templates/cowork-fit-log.md and add three rows.
Write the task name and the minutes per week it really costs, including the part where you go and find the file.
Mark the row with the largest number.
You know it worked when: Three rows carry a minutes figure and you can say out loud which one costs you most.
What you type
Read opportunities.csv and accounts.csv in this folder.
List every open opportunity (stage in Qualification, Discovery, Solution Review,
Technical Validation, Proposal, Negotiation) where days_in_stage is more than
double the segment norm for its account. Segment norms from DATA-DICTIONARY.md:
SMB 18, Mid-Market 32, Enterprise 48.
Output a markdown table: opportunity_id, account_id, segment, stage,
days_in_stage, owner. Sort by days_in_stage, highest first.
If an opportunity's account_id does not resolve in accounts.csv, do not guess a
segment. Write the row to exceptions.csv with the reason. Do not modify any
input file.
Notice: It read all 528 rows without you pasting any of them, and it told you about the one row it could not place instead of quietly dropping it.
A working session runs describe, plan, execute; the human enters twice, in the brief before the run and by interrupting during it, and reading the finished files is Check rather than a third steering point.
Open last week's calendar and sent folder. Somewhere in there is the work that never appears in a
job description and eats the day anyway: the forty-minute reporting request, the pre-call
research, the CRM fields you update on Friday. You have probably tried a chat assistant on some
of it. It writes well, but it cannot open the export in your downloads folder, and every Monday
you paste the same context in from scratch. Before you decide what to do about that, you need a
list.
What you will be able to do
Write down your recurring work with an honest minute count. Every decision in this course
starts here.
Say, for any task, whether it needs a conversation, a working session with your files, or a
person, and name the reason.
Describe the loop a working session runs, and the two moments where you can change its
direction.
Where your week actually goes
Everything this lesson uses came in the practice pack from the course page; if you have not
downloaded it, do that first.
Try this now
Open templates/M1/cowork-fit-log.md from your practice pack and fill only the first two
columns for twelve to twenty-five tasks. Four minutes, before any product enters the picture. This list is the spine
of Modules 1, 3 and the capstone.
Task
Minutes per week
Rebuild the weekly pipeline movement view
45
Answer ad hoc "how many X" questions in Slack
90
Pre-call research on named accounts
120
Three rules keep it honest:
Recurring only. If it happened once, it is not on this list.
Minutes as they really are, including the part where you go and find the file.
Name the work, not the tool. "Rebuild the movement view", not "use the BI tool".
Leave the other columns blank. You fill them in Lesson 3 and again in Module 3.
Check yourself: which three tasks cost you the most minutes per week?
Chat, Code, Cowork
Cowork is a mode of the Claude application that runs multi-step work on real files and connected
tools, not just text in a chat window. It sits alongside Chat and Code as a third mode of the
same product. Anthropic's framing is the most useful sentence anyone has written about it: Chat
is a conversation; Cowork is a working session where you describe the task, Claude plans and
executes it, and you steer along the way.
Two things follow, and both matter to you.
It was built for people who do not write code. Anthropic aimed it at non-developers after
noticing developers already using the coding agent for non-coding work.
It has a filesystem. Point it at a folder and it can read, create and modify files inside
it. That is what moves work out of the chat window, and it is the source of every risk in
Module 2.
Worth knowing: reporting on Anthropic's own session data indicates most Cowork usage is
business operations and content creation, not coding. The exact percentage is unverified, so
take the direction and skip any number you see quoted, including ours, until you read it at
source.
Check yourself: can you tell the three modes apart in one sentence each?
The task loop, and the two places you steer
you describe an outcome
-> Claude plans
-> Claude executes across multiple steps against real files
-> you steer
You get two steering points, and they are not equally good.
Before, in the brief you write. The cheap one. Everything you say up front costs a sentence
and saves a rerun. Module 4 turns this into a seven-part template.
During, by interrupting. You can stop a running task and narrow it ("only the EMEA
accounts") or correct it ("the amount field is in local currency"). This one is expensive, and
Module 4 teaches when to stop correcting and rewrite instead.
There is no third steering point. Once a run has produced eleven files, reading them is not
steering, it is checking: a different discipline with its own cost, and the Check letter of
TRACK.
Check yourself: which of the two steering points have you been neglecting?
Watch the difference
The same question, asked two ways, on the shipped Talvera data.
In a chat window. Paste 40 rows of the pipeline export and ask "which of these are stuck?"
You get a competent answer about those 40 rows. For the other 488 rows, you paste again. Next
week, you paste again. The context dies when the tab does.
In a working session. Point Cowork at your pack's data folder and write: "Read opportunities.csv. List
every open opportunity where days_in_stage is more than double the segment norm in the data
dictionary. Output a markdown table with opportunity id, account id, stage, days in stage,
owner." It reads the file itself, applies the rule across all 528 rows, and writes a file.
The whole difference: you did not paste the data. Pasting caps the job at what fits in a
conversation. Pointing at a folder caps it at what is in the folder, which is why "what else is
in the folder" is the most important question in Module 2.
Files or conversation?
These get conflated because both start with you typing a request, and the wrong call wastes
either an hour of setup or an afternoon of pasting.
Needs a conversation:
"Help me think through whether we should keep running the SMB motion at all after the pricing
change. I have a view and I want it challenged."
No file answers this. The inputs are your judgment and things you have not said out loud yet.
The output is a better-formed opinion, not an artifact. The tell: the output is an opinion
you will hold, not a file you will send.
Needs files:
"For every open opportunity in the pipeline export, tell me which ones have a close date that
has already passed, and group them by owner."
The answer exists in the data and nowhere else, across hundreds of rows. You could not paste
enough of the file into a chat to answer honestly. The tell: a file already holds the right
answer, and the job is to get it out.
The trap case. "Draft the renewal email for our largest at-risk account" looks like
conversation work because the output is prose. Wrong axis. The question is not whether the
output is words, it is whether the inputs are in files. This email is only worth sending if
it names the right renewal date, amount, stakeholder and the thing that went wrong, and all four
sit in records: accounts.renewal_date, the open opportunity amount, contacts.persona, and a
support ticket history. A chat draft will be fluent and generic, and you will spend longer
fixing its facts than writing it yourself.
The real conversation-shaped task hiding next to it is "what is our position going into this
renewal". Keep that a conversation. Drafting the email that carries the position is file work.
And note: the email is draft-only, not read-only, because a human sends it. Lesson 3 makes that
distinction properly.
What you keep
cowork-fit-log.md, columns 1 and 2 filled, in your own folder rather than a shipped course
folder. Lesson 3 adds three columns. Module 3 re-scores the whole thing with the five-factor fit
test.
Check yourself
Take the highest-minute task on your log. Name the file, export or system that holds the
right answer for it. If you cannot name one, say what that tells you about handing it over.
One of your tasks probably has an output that is prose but inputs that are records. Which
one, and which column of your log did you nearly put it in?
You have a run going that is producing the wrong grouping. Was the cheaper fix available
before or during? Say what the sentence would have been.
Answers
Items marked Open run on your own list and have no key.
1. Open, and the checkable part is what a blank means. If you cannot name a file, export or
system, Checkability is low, and Module 3 makes that a veto: you would be handing over work
whose correctness you cannot test.
2. Open. The usual case is a drafting task: prose output, record inputs. If yours went into
the conversation column, re-read the trap case above.
3. Before. Grouping is scope, and scope is the cheapest thing to fix in the brief: "group by
region, not by country". Correcting during the run buys you this output. Putting it in the
brief buys you this output and next week's.
Where this goes next
Next lesson: you run one, on data you did not have to prepare, and get a real file out of
it.
Track labs: every track lab starts from a row on this fit log, so a thin log makes for thin
labs. Part B of the exercise sheet in labs/LAB-01 of your practice pack (start with its README)
is the full build.
Module 1 · Cowork as a revenue work surface
M1 lesson / about 15 min
Do it
Your first end-to-end run
Do this nowRead two fields on Vireo45s
Open accounts.csv and find ACC-02858, Vireo Retail PLC.
Read arr_usd and currency on that one row.
Write both values down before you run anything.
You know it worked when: You are holding 121920.0 next to a currency code of GBP, and one question about which unit this account is actually measured in.
What you type
Read accounts.csv, contacts.csv, opportunities.csv and activities.csv in this
folder.
Produce a one-page snapshot of the account named "Vireo Retail PLC" and write it
to first-run/account-snapshot.md.
Sections, in this order:
1. Who they are: segment, industry, employee count, country, region, owner, CSM,
ARR in USD, billing currency, health score, renewal date, tier.
2. Open opportunities: opportunity_id, name, stage, amount with its currency,
type, close date, forecast category, days in stage, next step.
3. Contacts: name, title, persona, last activity date.
4. What the data does not tell you: anything a person would want before this call
that is not in these four files. Name the field that is missing or ambiguous.
Do not infer anything that is not in the files. If a field is blank, say it is
blank rather than filling it in.
What comes back
first-run/account-snapshot.md written.
## Who they are
ACC-02858 Vireo Retail PLC. Mid-Market, Retail, 538 employees, United Kingdom,
EMEA. Owner Bettina Farrow. CSM Niamh Ricci. ARR 121,920 USD. Billing currency
GBP. Health score 36. Renewal 2026-09-17. Tier 2. Customer.
## Open opportunities
OPP-006874, Vireo Retail PLC - Renewal 2026, Negotiation, 96,000 GBP, Renewal,
close 2026-09-17, Best Case, 42 days in stage.
Next step: "Chase signed order form, sponsor left"
## What the data does not tell you
- 96,000 GBP times 1.27 is 121,920, which is the arr_usd field. Two files agree
only after a conversion neither of them performs.
- contacts.csv lists Wouter Nyberg, Chief Operating Officer, persona Economic
Buyer, last activity 2026-05-08. No column records whether he still works there.
- health_score 36 has no definition anywhere in this data.
Notice: The section that earns the run is the last one, and a snapshot that reports nothing missing on this account has failed it.
A working session runs describe, plan, execute; the human enters twice, in the brief before the run and by interrupting during it, and reading the finished files is Check rather than a third steering point.
An account executive drops into your Slack at 09:12: "Can you pull me everything on Vireo before
the 11 o'clock?" You know what that means: open the account record, work out which opportunity
rows are still live, find the last touches, notice the amount is in pounds while the ARR field
is in dollars. Twenty minutes of clicking, and again on Thursday. Do it once, properly, on
shipped data, so you have seen the whole loop before anyone asks you to trust it. This lesson is
read-only: nothing writes outside the output folder you create.
What you will be able to do
Point a working session at a folder and get back a file, not a chat reply.
Describe what happened in terms of the task loop, not clicks.
Read your own output for the thing that matters most: what it did not know.
Choose the folder on purpose
Before the brief, you choose a folder. Most people make this choice by accident.
The shipped dataset lives in the data/ folder of your practice pack, the zip you downloaded
from the course page and unzipped into your course folder. The pack root around it holds the
templates, the labs and the answer keys. Point at data/, not the pack root:
Claude cannot read or write outside the directories you explicitly grant, so everything inside
the folder is in scope, including files you forgot were there. Copy the folder somewhere of
your own first; nothing here modifies a file, but the habit is free.
Check yourself: which folder did you grant, and what is one thing in the parent that you did
not?
Write the brief, then run it
Try this now
Type this brief exactly. Do not paraphrase. Module 4 explains why it looks like this. For now,
notice that it names files, names an output, and says what to do about things it cannot
answer.
Read accounts.csv, contacts.csv, opportunities.csv and activities.csv in this folder.
Produce a one-page snapshot of the account named "Vireo Retail PLC" and write it to
first-run/account-snapshot.md.
Sections, in this order:
1. Who they are: segment, industry, employee count, country, region, owner, CSM,
customer since, ARR in USD, billing currency, health score, renewal date, tier.
2. Open opportunities: for each one, opportunity id, name, stage, amount with its
currency, type, expected close date, forecast category, days in stage, next step.
3. Closed history: one line per closed opportunity, oldest first.
4. Contacts: name, title, persona, last activity date.
5. Last three activities: date, type, subject, owner, outcome.
6. What the data does not tell you: anything a person would want before this call
that is not in these four files. Be specific about which field is missing or
ambiguous.
Do not infer anything that is not in the files. If a field is blank, say it is blank
rather than filling it in.
If it answers in the chat instead of writing the file, say: "Write it to
first-run/account-snapshot.md in this folder and confirm the file exists."
Check yourself: what are three things in that brief that a one-line request would not have
carried?
Read what came back
Do not read the output for whether it sounds right. Read it for four things, in order.
Did it use all four files? Section 5 is the tell. If activities are missing, it answered
from one file and stopped.
Does anything reconcile?accounts.arr_usd is 121920.0 and the open renewal is
96000.0 GBP. The build rate is GBP = 1.27 USD (DATA-DICTIONARY.md, in your pack's data
folder, Conventions section), and
96,000 times 1.27 is exactly 121,920. Two numbers from two files agreeing is the cheapest
evidence you will ever get.
Did it invent anything? Scan section 1 against accounts.csv field by field. It is one
row.
Is section 6 real? A "what the data does not tell you" section that says "nothing" is a
failed section, and on this account it is provably failed.
What our run looked like
Yours will differ in wording and section order. That is why the course ships reference
artifacts, not answer keys.
What it got right. Section 1 matched accounts.csv row ACC-02858 field for field:
Mid-Market, Retail, 538 employees, United Kingdom, EMEA, owner Bettina Farrow, CSM Niamh Ricci,
ARR 121,920 USD, currency GBP, health score 36, renewal 2026-09-17, Tier 2. Section 2 found one
open opportunity, OPP-006874, Renewal 2026, Negotiation, 96,000 GBP, close date 2026-09-17,
Best Case, 42 days in stage, next step "Chase signed order form, sponsor left".
The dangerous number. The output said probability 85. We did not ask for it.
DATA-DICTIONARY.md defines opportunities.probability as the stage default: not the rep's
call, not evidence about this deal. An 85 on a slide is a forecast claim; here it is a lookup
table. We struck it and wrote the reason into section 6.
What section 6 should contain. Three things this data cannot tell you:
The economic buyer may not be there.contacts.csv lists Wouter Nyberg, Chief Operating
Officer, persona Economic Buyer, last activity 2026-05-08. No field says whether a person
still works there. The absence is the finding, and the folder holds a call note that speaks
to it.
The renewal is in GBP, the ARR field is in USD. Sum this account into a regional total
without converting and you have a wrong number. Converting is the analyst's job, per the data
dictionary.
Health score 36 has no definition in the data. A number with no rule behind it.
What we did not ask, on purpose: whether the deal will close. No file holds the answer, so
the output would have been fluent and unfounded. Module 3 turns that judgment into an
instrument.
Now you try: Oldcastle Leisure Inc
You are covering for a colleague. The CS lead wants the same snapshot for Oldcastle Leisure
Inc, a Mid-Market hospitality customer, before a call this afternoon. Same four files
(accounts.csv, contacts.csv, opportunities.csv, activities.csv), different account, on
purpose: you are running a loop, not reproducing an answer. Grant your copy of the pack's
data/ folder. No connectors. Read-only: the only file written is your output.
Try this now
Grant the folder. Say out loud what is in its parent that you did not grant.
Fill the four gaps in the starter brief below, then run it.
Do not correct it mid-run, even if it goes somewhere odd. Let it finish, so you see what an
unassisted run produces.
Run the four checks, in order.
Add one sentence at the bottom: the thing you would not put in front of the customer
without checking, and why.
Read accounts.csv, contacts.csv, opportunities.csv and activities.csv in this folder.
Produce a one-page snapshot of the account named "Oldcastle Leisure Inc" and write it
to first-run/oldcastle-snapshot.md.
Sections:
1. Who they are: <<GAP 1: list the exact account fields you want, by column name>>
2. Opportunity history: every opportunity on this account, oldest first, with
opportunity id, name, type, stage, amount, currency and close date.
<<GAP 2: state what to do if the same opportunity id appears on more than one row>>
3. Contacts: name, title, persona, last activity date.
4. Last three activities: date, type, subject, owner, outcome.
5. What the data does not tell you: <<GAP 3: say what counts as a good answer here,
so it does not write "nothing">>
Constraints: <<GAP 4: one line telling it what not to infer>>
Stuck? Take the hints in order.
The gap that matters most is not formatting. This account's opportunity history has a shape
problem; a snapshot that adds up the amounts without noticing prints a number no system in
the company holds. What do a row count and an identifier count each tell you?
Count two things in opportunities.csv for account_id = ACC-02848: rows, and distinct
opportunity_id values. Compare your total against accounts.arr_usd. Write the
instruction that makes the agent do the same comparison.
GAP 2: report the duplicate identifier rather than silently picking a row, count each
opportunity once, and never add a Renewal amount to a New Business amount (renewal
amounts restate current ARR, per the data dictionary). GAP 3: name at least two classes of
unknown, such as a field with no definition, and a fact a person on the call would ask about
that no column carries.
What a correct run finds (templates/M4/ in your practice pack carries the mature version
of this brief). The checkable parts, read from the shipped files:
Account.ACC-02848, Oldcastle Leisure Inc, Mid-Market, Hospitality, 302 employees,
United States, NA, owner Ursula Ulrich, CSM Helena Zabala, contract start 2023-10-19, ARR
55,000 USD, currency USD, health score 37, renewal date 2026-10-19, Tier 2, Customer.
Opportunity history. Four rows, three opportunities. OPP-007468 New Business 55,000 USD
Closed Won 2023-10-19. OPP-007132 Renewal 2024, 55,000 USD, Closed Won 2024-10-19.
OPP-007099 Renewal 2025, appearing twice, at 69,000 USD with close date 2025-10-19 and
at 55,000 USD with a blank close date. Both are documented in
answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab,
under the classes duplicate-opportunity-rows and missing-values.
Contacts. Three: Ulrich Steenkamp (Payroll Analyst, End User), Fabian Nyberg (Chief
Information Officer, Economic Buyer), Tobias Lombardi (Director of IT, Champion).
A correct "what the data does not tell you" contains at least: OPP-007099 has two
conflicting amounts and the file does not say which is right; no renewal opportunity exists
for the 2026-10-19 renewal date; health score 37 has no stated definition.
Section order, table versus prose, and wording may all legitimately differ. What may not: an
output that prints a single total without mentioning that one opportunity id appears twice is
wrong, however good the rest looks.
Worth knowing: four rows, three opportunities, one duplicate missing its close date. The
arithmetic that goes wrong first is not the hard arithmetic. It is counting.
What you keep
first-run/account-snapshot.md and first-run/oldcastle-snapshot.md, in your own folder. Keep
them even though they are rough: Module 4 regenerates the same snapshot from a written brief and
diffs the two. You cannot make that comparison later if you delete the before picture now.
Check yourself
Your run produced a total for Oldcastle Leisure. Name the two source-of-truth comparisons
that would tell you whether it is right, and say which file each one lives in.
The demo struck the probability field from the Vireo snapshot. Give the reason in one
sentence that does not use the word "wrong".
You granted the pack's data/ folder. If your own CRM export had been sitting in that folder, what
would have changed about this run? Answer in terms of what the agent could reach, not what it
was told.
Answers
1.Row count against distinct opportunity_id count, both from opportunities.csv:
four against three. Sum of Closed Won New Business and Expansion amounts against
accounts.arr_usd, in accounts.csv: 55,000 against 55,000. The second only works because it
names which types go in, since Renewal amounts restate current ARR rather than adding to it.
2. Because opportunities.probability is the stage default: a property of the stage, not
evidence about this deal. It would read the same on any opportunity in Negotiation.
3. Nothing about what it was told, everything about what it could reach. The export would
have been inside the granted directory, so in scope for every request in the session, including
ones where you had forgotten it was there. Nothing would have warned you.
Where this goes next
Next lesson: you close this session, open a new one, and ask for the same thing on another
account. What happens is the most important limit in the product.
Track labs: Part A of the exercise sheet in labs/LAB-01 of your practice pack (start with
its README) is this run at full scope, with a measured baseline.
Module 1 · Cowork as a revenue work surface
M1 lesson / about 13 min
Concept
The four walls, before you build against them
Do this nowProve there is no memory60s
Confirm first-run/account-snapshot.md from the previous lesson is on disk.
Close the session entirely and open a new one with the same folder granted.
Type only: Do the same as yesterday, but for Oldcastle Leisure Inc.
You know it worked when: Either it asks what "the same" means, or it produces something shaped like your old file, which is a file being read rather than anything being remembered.
What you type
Do the same as yesterday, but for Oldcastle Leisure Inc.
What comes back
I have no record of a previous session.
I can see one file in this folder that looks like a prior output:
first-run/account-snapshot.md, a one-page account snapshot with four sections
(who they are, open opportunities, contacts, what the data does not tell you).
Two options:
1. Reuse that file's structure for Oldcastle Leisure Inc (ACC-02848).
2. Tell me the sections you want.
If "the same" meant something other than that file, say so, because it is the
only thing in this folder that describes it.
Notice: Nothing was remembered. A file inside the granted folder was read, and if you move that file out, the memory goes with it.
Rails is the one TRACK letter you build once at setup and then inherit; Triage, Assign, Check and Keep run on every single task, which is why Module 2 comes before Module 3.Each rung of the blast-radius ladder fixes its own review gate, approver and reversal plan, and the top rung requires the reversal plan tested and recorded before the first write.There is no memory between sessions, so the five things written to files cross the session boundary and everything said inside the session does not.With no event trigger available, cadence plus saved state substitutes for it, and the price is a detection latency you state in writing rather than discover.
The run in Lesson 2 worked. That is the dangerous moment. The next thing most people do is
describe a workflow to their manager that the product cannot build, usually: "when a deal moves
to Stage 4, Claude writes the close plan and posts it in the channel." It sounds modest. It is
not buildable here, and the reason is architectural, not temporary.
What you will be able to do
Name the documented limits and, given a proposed workflow, say which one it collides with.
Classify any task as read-only, draft-only or writes-to-system, and say what each class
obliges you to do.
Say what the 4Ds cover, what TRACK adds, and why.
Record what your own machine and plan actually support, with the date you checked.
The six limits
Each is documented, and none is a footnote.
Limit
What it means on a Monday
Handled in
No memory between sessions. Continuity is manufactured with files.
You cannot say "same as yesterday". Yesterday does not exist.
M4
No event triggers and no inbound email address. There is no native "when a deal moves to Stage 3, run this".
Every workflow you design is either you starting it or a cadence starting it.
M8, named here
Scheduling is convenience automation, not a guaranteed cron. It should not carry a task that must run at an exact time without fail.
The board pack that must exist by 08:00 Monday is not a scheduling problem you have solved.
M8
One shared usage pool. Cowork draws on the same plan allowance as Claude.ai chat and Claude Code, under a 5-hour rolling window plus a weekly cap.
A heavy hygiene pass can cost you your chat access for the rest of the window.
M3
The directory boundary is the security boundary. Claude cannot read or write outside the directories you explicitly grant.
What is in the granted folder is in scope, including what you forgot was in there.
M2
The enterprise audit gap. The Compliance API and Audit Logs do not cover Cowork yet.
If you need to evidence what happened to an auditor, you build that yourself.
M8, and M4 at individual level
Do not memorize a number for the usage limit. Anthropic does not publish stable allowances, and
we could not verify the figures. Learn the shape, look up the number when you need it.
The framing that matters most: Cowork is not a workflow engine. It is good at multi-step
work you start or schedule, and structurally unable to react to an event in your CRM without
something else doing the triggering.
Check yourself: which of the six limits does your favourite idea collide with?
Blast radius: three classes
You will use these in every module and every lab from here.
Class
Meaning
Default approval
Read-only
Reads files or systems, writes nothing outside the working folder
You proceed
Draft-only
Produces something a human sends or applies: an email draft, a proposed field update, a CSV of changes
Human review before it leaves the folder
Writes-to-system
Changes a system of record, sends to a customer, or touches money
Named approver plus a reversal plan, every time
The trap: the class is set by where the output goes, not by how the work felt. A list of
at-risk accounts is read-only right up until someone pastes it into a customer's business review
deck, at which point it was draft-only all along and nobody reviewed it.
Try this now
Open the cowork-fit-log.md you started in Lesson 1 and classify five rows. At least one
will be harder than it looks.
Check yourself: for one task on your log, must a person read the output before it leaves
your folder?
The 4Ds, and what TRACK adds
The 4Ds are Anthropic's AI Fluency Framework, developed with Prof. Rick Dakan (Ringling
College of Art and Design) and Prof. Joseph Feller (University College Cork):
D
What it covers
Delegation
Deciding whether, when and how to engage AI on a task
Description
Describing the goal so the AI behaves usefully. Prompting lives here, and only here.
Discernment
Judging the quality, accuracy and bias of outputs
Diligence
Taking responsibility for what you do with AI, including governance and safety
Their most useful move is to demote prompting: one part of one of four competencies. This
course gives prompt craft one short section in Module 4 and never a module.
TRACK is what we add, for one reason: the 4Ds were designed for a conversation, and a
conversation has no data perimeter.
Letter
The decision you make
Taught in
T Triage
Is this worth handing over, and what is the worst thing that happens if it goes wrong?
M3
R Rails
What can it see, what may it change, what is fenced off?
M2
A Assign
What am I telling it to do, and where does that instruction live?
M4
C Check
What evidence proves the output is right, without redoing the work?
M4
K Keep
What survives this run so the next one is cheap and this one is defensible?
M4, M8
Rails is the letter the 4Ds do not have, and it is the reason for the addition. One
asymmetry to note: R is built once; T, A, C and K run every time. That is why Module 2 is
Rails and Module 3 is Triage, even though T comes first in the acronym. Short version: the 4Ds
tell you how to work with AI. TRACK tells you how to hand over work that touches your company's
files and your customers' data.
Check yourself: which D does your current chat habit exercise, and which TRACK letter does
it ignore?
What you actually have
Try this now
Record this in your fit log, with today's date. Two minutes.
Plan. Cowork is included in every paid Claude plan and is not in the free tier.
Operating system. macOS since January 2026. Windows later, with arm64 not supported as
of March 2026.
Surface. Desktop, plus web and mobile in beta from July 2026, rolling out starting with
Max.
Linux. A beta exists and computer use is not available in it. We could not verify its
full scope, so check before designing around it.
Date checked. Write it down.
Worth knowing: the most complete public Cowork tutorial in this market still says Max-only
and macOS-only, which stopped being true around April 2026, and the page carries no
verification date. Check rather than trust, and that includes this course.
The limits, used on a real proposal
Back to the workflow from the top: "when an opportunity moves to Stage 4, Cowork writes the
close plan and posts it in the deal channel."
Which wall. Event triggers. There is no native "when a deal moves, run this" and no inbound
email address. This needs a different architecture, not a better prompt.
What survives. "Write the close plan from the record" is file work and entirely buildable.
The trigger is not. The substitute is a cadence: a daily pass over the opportunity file that
finds deals which entered Stage 4 since the last pass.
What that costs. Detection latency: up to 24 hours late on a daily cadence. Tell your VP
that number up front. Module 8 builds this pattern and calls it cadence-plus-check.
The second wall. A daily pass is a scheduled task, and scheduling is convenience automation,
not a guaranteed cron. The design needs an answer to "what happens on the morning it does not
run". If your answer is "someone notices", you do not have an answer yet.
Two walls, one modest-sounding request. That is why the limits are Lesson 3 and not an appendix.
Break it yourself
Try this now (three minutes)
Open the account snapshot from Lesson 2. Confirm it is on disk.
Close the session entirely. Not a new message in the same thread: a new session.
In the new session, with the same folder granted, write exactly this and nothing more:
Do the same as yesterday, but for Oldcastle Leisure Inc.
What you should see: a new Cowork conversation has no memory of previous ones, so it has no
recollection of yesterday. What happens next depends on the folder, and both outcomes teach
something:
If the folder is clean, it asks what "the same" means. Mildly annoying, entirely safe.
If your first-run/account-snapshot.md is in the granted folder, it may read that file and
imitate its shape. This looks like memory and is not: it is inference from a file lying
around. Change the file and the "memory" changes; delete it and it is gone. This is the
expensive outcome, because the output is close enough that a busy person pastes it into a
deck.
Worth knowing: continuity is a file convention, not a product feature. The documented
conventions are per-folder context files: preferences.md or CLAUDE.md in the working
folder, MEMORY.md read at session start, and a project folder's own CLAUDE.md and
MEMORY.md read inside that project. The fix is not to remember harder: write the instruction
into a file the next session reads. Module 2 builds those files; Module 4 writes the brief that
lives in them.
What you keep
cowork-fit-log.md, now with all five columns (task, minutes per week, needs files or
connectors, blast-radius class, first call), plus the dated record from "What you actually
have". It is the input to Module 3's triage instrument, the shortlist for the capstone, and the
document you show a manager who asks what you would actually use this for.
Check yourself
A colleague proposes: "every time a support ticket is escalated, Cowork emails the CSM a
summary and updates the account health field." Name every limit this collides with, and
which one is fatal rather than merely inconvenient.
Take the read-only row on your fit log with the highest minutes. Describe a plausible way
its output ends up in front of a customer. Reclassify it and state the added obligation.
Someone says they got Cowork to "remember" their preferred report format across sessions.
Give the two explanations consistent with the no-memory limit, and what you would look at to
tell them apart.
Which D does "I got a better answer by rephrasing my prompt three times" exercise, and which
TRACK letter does it ignore?
Answers
1. Three limits, one fatal. No event triggers is fatal: "every time a ticket is
escalated" is not buildable, and no rewording fixes it. The other two only constrain the design:
the health-field update is writes-to-system, so it needs a named approver and a reversal
plan, and the audit gap matters if anyone later has to evidence who changed that field.
Substitute a cadence and state the detection latency.
2. Open on which row, determinable on the obligation: it was draft-only all along, so a
named human reads it in full before it leaves your folder.
3. Either it is in global instructions, or a file in the granted folder is read at
session start. Neither is memory. Move the folder file out and re-run: if the behaviour
survives, it is global.
4. It exercises Description. It ignores Check, because "better" was judged by
reading rather than evidence. It also loses Keep, since the winning phrasing was never
written down.
Where this goes next
Next module: Module 2 makes the folder decision you made casually in Lesson 2 into a
boundary you make on purpose, and proves that an instruction is not one.
Track labs: Part B of the exercise sheet in labs/LAB-01 of your practice pack (start with
its README) builds the fit log from your real week. CP-01, the module quiz in the
interactive app, tests the limits map by handing you a
workflow that collides with one.
Module 1 · Cowork as a revenue work surface
Lab / ? min / read-only
Lab
The first run, and the session that forgot
LAB-01: The first run, and the session that forgot
The situation
It is Monday. You are the RevOps analyst. An AE has asked you, again, for "everything we know about
Wexley Care" before a call this afternoon. You know exactly where the answer is: it is spread across
four exports sitting in a folder. Pulling it together is twenty minutes of filtering and retyping, and
you will do it again on Thursday for a different account.
This lab is the first end-to-end run, before any configuration and before any framework. You point
Cowork at a folder, ask for one page, and then you deliberately break it: you close the session, open a
new one, and ask for the same thing again. What happens next is the single most important product fact
in this course, and describing it would not stick.
What you will produce
File
What it is
first-run/account-snapshot.md
A one-page snapshot of one named account, including a section listing what the data does not tell you
first-run/second-session-note.md
Three sentences on what changed when you asked again in a fresh session
cowork-fit-log.md (Part B)
An inventory of your own recurring revenue work, 12 to 25 rows, scored for a first call of human, chat or Cowork
Outcome accounting
Manual baseline: 19 minutes
Measured. The lab author built this snapshot by hand from the four CSVs in a
spreadsheet: filter accounts on account_id, filter opportunities, filter
contacts, sort activities by timestamp and take the last three, then write the
page. Timed once, on this account. One account, not a book.
Target with Cowork: 4 minutes to first output, observed while building this lab. Stated as observed,
not promised. Your first run will be slower because you are also learning where
to point it.
Verification cost: 5 minutes. Reconcile the opportunity count and the amount total against
opportunities.csv, reconcile the contact count against contacts.csv, and check
the three activity timestamps. Counted separately and inside the net.
Net on one account: 19 - (4 + 5) = 10 minutes
Your actual: _____ to first output _____ to verified output
The honest version: 10 minutes on one account does not justify learning a product. The number that
matters is the second account, which you do in Part A step 7 and which took the
lab author 3 minutes including the check. That is the number to compare against
19, and it is the reason the brief is worth more than the output (Module 4).
Blast radius
Read-only. Cowork reads files in the folder you grant and writes only into that folder. Nothing in
this lab touches a CRM, sends anything, or changes a system of record. Review gate: a spot check on a
sample, which is the verification step above. Reversal plan: delete the output file.
The directory boundary is the security boundary: Claude cannot read or write outside the directories
you explicitly grant (FACTS §5.2). You are granting a folder of synthetic data with no real people in
it, which is why this is the first lab and not the fifth.
Declared scorecard scope
None. This lab produces a snapshot, not a findings list, so there is no defect population to compute
recall against. The four labs that carry a detection scorecard are LAB-05 to LAB-08. See SCORING.md §1.
How you verify your own output, before anyone scores it
Do these three checks yourself. They are the whole of the Check step at this stage, and they take about
five minutes.
Reconciliation. Your snapshot names some number of opportunities and some total amount. Open
opportunities.csv, filter account_id = ACC-02830, and confirm both. If the count matches and the
total does not, you have found a currency or a duplicate problem, which is Module 5.
Sample. Pick two contacts named in the snapshot. Confirm the title and the persona against
contacts.csv character for character. Titles are the field agents most often smooth over.
The unknowns test. Read your "what the data does not tell you" section. If it is empty, or if it
says something generic like "we do not know their budget", the snapshot is not finished. A usable
unknowns section names a field that is blank, a unit that is ambiguous, or a claim that has no source
in the folder. Three specific items is a pass.
If a check fails, do not edit the output. Change the request and run it again. That distinction is the
entire method taught in Module 4 and it starts here.
Files you need
Everything is in course/labs/data/. You need accounts.csv, contacts.csv, opportunities.csv and
activities.csv. Read DATA-DICTIONARY.md first: it tells you that amounts on opportunities are in the
account's local billing currency while arr_usd is in USD, and that fact is going to matter in about
four minutes.
Copy the folder somewhere you are comfortable granting. Do not grant your whole home directory.
What comes next
LAB-02 turns the folder you just granted into a workspace with a boundary you can prove, which is what
makes it safe to point Cowork at your own data in the labs after it.
Module 1 · Cowork as a revenue work surface
5 questions / pass at 70%
Checkpoint
Module 1 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 5
It is Monday. Five things are waiting for you.
Shorten a cold email you already drafted, and make it less pushy.
Produce a one-page snapshot of one named account: who they are, open opportunities with amounts
and stages, the last three logged activities, and what the data does not tell you. The account,
opportunity and activity exports are three CSV files in a folder on your machine.
Decide whether to approve a 15% discount to close a deal before quarter end.
Build a duplicate candidate list across 308 account rows and write it to a CSV with the match
reason on every pair.
Turn 16 messy call-note files into one structured record per opportunity.
Which assignment is correct, and for the right reason?
Hint
Ask one question of each task before anything else: where do the inputs live right now, and where does
the answer have to end up? Then ask a second question of the one that still looks ambiguous: if this
output were wrong, what would I compare it against?
A. Correct. The deciding factor is file access and file production, not difficulty. Items 2, 4 and 5 each begin in files that already exist on disk and end in files that have to exist afterwards, which is what Cowork adds over a chat window (FACTS §5.1). Item 1 is a paragraph in, a paragraph out, so the extra machinery buys nothing. Item 3 has no source of truth to check an answer against: the right discount depends on what you know about the buyer, the quarter and the precedent it sets, and no file holds that.
B. Incorrect. Output length is the wrong test. Item 2 produces one page, but it produces that page by joining three exports that are sitting in a folder. If you take it to a chat window you paste the data yourself, which is most of the work, and you do it again next week because there is no memory between sessions (FACTS §7.1). The test is where the inputs live, not how big the output is.
C. Incorrect. Everything except the discount decision is correctly placed, and the discount decision is the one that matters. It fails the fit test on Checkability: there is no record you can compare an answer against, so you could never prove the output right or wrong. "More capable mode" is not a triage rule, and treating it as one is how a team ends up with an agent making commercial calls.
D. Incorrect. Messy notes are exactly the work that benefits from being handed over, because the result is checkable: every structured record traces back to a line in a named file, and you can sample five of them against source. You are also wrong on item 2, for the reason in option B. Note that "unstructured" and "unverifiable" are different properties, and only the second one is a veto.
Question 2 of 5
A sales manager writes up an idea after reading about agents over the weekend:
"When an opportunity moves into Stage 4, Cowork should pick that up automatically, read the account
history and the call notes, and write the close plan into the deal record. That way the plan exists
before the deal review and nobody has to remember."
You have to answer today. Which response is correct?
Hint
Split the request into the part that describes work and the part that describes when the work starts.
Only one of the two collides with something documented, and the other one is a good candidate.
A. Incorrect. Nothing in Cowork watches a folder or a system and reacts to a change in it. There is no native "when a deal moves to Stage 3, run this" (FACTS §7.3). A run happens because you started it or because a schedule started it, and nothing else. This is the single most common design mistake a revenue operator makes in week one, and it is the reason the limit is taught in Module 1 rather than at the end.
B. Incorrect. A schedule is a clock, not a trigger. It fires on a cadence regardless of whether anything changed, and it will not notice the stage change unless you tell it to compare against a saved previous state. Worse, scheduling is convenience automation and is not a guaranteed cron, so it should not carry a task that must run at an exact time without fail (FACTS §7.2). A schedule plus a diff is the answer. A schedule alone is not.
C. Correct. You have separated the part that collides with a documented limit from the part that does not. The retrieval and drafting work is a good fit. The trigger is not available (FACTS §7.3), so you replace it with cadence plus a state comparison and you say out loud how late the detection is. A requirement stated as "when X happens" always turns into "how quickly after X do I need to know", and that second question has an answer.
D. Incorrect. This over-refuses and loses the value. Assembling the account history, pulling the relevant call notes and drafting a first version is retrieval, it is high volume, and it is checkable against the source records. What must stay human is the commercial judgment inside the plan and the decision to put anything into the deal record, which is a draft-only review gate rather than a refusal.
Question 3 of 5
Classify each of these three by blast radius: read-only, draft-only, or writes-to-system.
Produce a churn-risk summary from the usage and ticket exports, saved as a markdown file in your
output folder, which you read and then delete.
Produce the same churn-risk summary, and paste the top section into the monthly email you send to
each account's main contact.
Produce a CSV of proposed owner corrections that your operations lead reviews and then imports into
the CRM.
Hint
For each item, finish this sentence: "if this output is wrong, the first person to see the mistake is
_____." The answer to that sentence is the class.
A. Incorrect. Item 2 is the trap and you fell into it. The agent's action is identical in 1 and 2, and the blast radius is not, because in 2 the output is sent to a customer. Read-only means the output never leaves your working folder. The moment it does, a wrong line has an audience and the obligation changes.
B. Correct. You classified on destination rather than on activity, which is the point of the three classes. Item 2 acquires a review gate because a customer reads it. Item 3 is correctly split: the file is draft-only, and the import is the writes-to-system act, which is why the approver and the reversal plan attach to the import rather than to the file.
C. Incorrect. You have over-classified, and over-classification is not a safe error. If producing a draft is already writes-to-system, you need a named approver and a tested reversal plan for a file nobody has sent yet, the process becomes unusable, and people route around it. Reserve writes-to-system for the act that actually changes a system of record or reaches a customer, and put the review gate on the draft.
D. Incorrect. This is the most expensive answer of the four. Blast radius is a property of the workflow, not of the agent's file operations. If the workflow ends with a customer reading generated text or an operations lead importing generated rows, those endings are inside the workflow you are designing, and pretending otherwise is how an unreviewed line reaches a customer.
Question 4 of 5
You asked for an account snapshot. Twelve minutes later this came back.
Talvera account overview
Accounts reviewed: 308
Segments: SMB 127, Mid-Market 137, Enterprise 44
Top 10 accounts by ARR (USD) ........
Open pipeline by stage .............
Live customers with no health score: 2
You wanted one page on one named account, not a portfolio roll-up across the whole file. You watched it
start by loading every row of accounts.csv and said nothing, because you assumed it was warming up.
Where should the steer have happened, and why?
Hint
There are exactly two moments where you can steer. For each of them, ask whether using it would have
changed only this run or every future run.
A. Incorrect. Editing the output is the cheapest move today and the most expensive move over a quarter. You still have no brief, so you pay the same twelve minutes and the same edit every time, and you have no record of what produced the number if somebody asks. Editing output is a patch. Editing the instruction is a fix.
B. Incorrect. Interrupting would have been better than watching, and it is genuinely the second steering point, so this is the strongest wrong answer. But it treats a symptom. The plan was a fair reading of an underspecified request, so the same plan is what you get on the next run and the run after that. Steering during a run is for the surprises a good brief could not have anticipated, not for the ones it should have prevented.
C. Incorrect. Nothing here indicates a misunderstanding. A request that does not name an account or an output shape has many correct readings, and you received one of them. Rerunning the same request is the definition of a process you cannot reproduce, and it is the habit that makes people conclude the tool is unreliable when the input was ambiguous.
D. Correct. The two steering points are before (the brief) and during (interrupt, narrow, correct). Both were available here and only one of them fixes the cause. A brief that had named the account, the three input files and the exact sections would not have produced a portfolio roll-up, and it is reusable next week. Note the second reason this matters: a correction made during a run lives in a session that will not exist tomorrow, because there is no memory between sessions (FACTS §7.1).
Question 5 of 5
A colleague describes how they work:
"I have got really good at this. I write a long, careful prompt, I read what comes back properly
instead of skimming it, and if a number looks off I go and check it against the report. I have been
doing it for months."
They are about to point Cowork at the folder that holds this quarter's CRM export, last year's
commission statements and a churn post-mortem, because that is where they keep their work.
Which D are they exercising well, and which TRACK letter does their habit leave uncovered?
Hint
Sort the colleague's described habits into the Ds first, honestly, and give credit where it is due. Then
look at the last sentence of the stem, which describes something their method has no step for.
A. Incorrect. Delegation is deciding whether and when to engage AI on a task at all, and nothing in the description tells you they do that deliberately. Your reading of Check is also too strict: checking every number is usually more expensive than doing the work by hand, which fails the rule that verification must cost less than the job. Spot-checking a suspicious number is weak evidence, not absent evidence.
B. Correct. Writing a careful prompt is Description. Reading the output properly and checking a suspicious number against a source is Discernment. Both are real skills and both transfer. What does not transfer is the perimeter decision, because the 4Ds were built for human and AI collaboration in general, which in practice means a conversation, and a conversation has no filesystem. Rails is the letter the 4Ds do not have, and the folder in the stem is the reason it exists: commission statements and a churn post-mortem are in the same place as the work (FACTS §5.2).
C. Incorrect. Diligence is taking responsibility for what you do with the output, including governance and safety, and the stem shows the opposite: they are about to grant a folder containing commission data without appearing to have thought about it. Assign is also not the gap here. Their instruction is not well structured, but the thing that will hurt them first is what the agent can reach, not how the request is phrased.
D. Incorrect. The four Ds are a good general model and they are silent on four decisions a chat user never faces: which folder does it get and what else is in that folder, what does done look like when the output is twelve files, how do you check 400 rows without redoing the work, and how does this run again next Tuesday when the agent remembers nothing between sessions (FACTS §7.1). "The method is unchanged" is exactly the belief this item is testing.
M2: Rails, your workspace, your data boundary, your connectors
Last verified: 2026-08-02.Product claims: FACTS §2, §3, §5.2, §5.5, §5.7, §5.8, §7.1, §7.8.
Dataset: Talvera, seed 20260801. The failure lane in Lesson 1 needs no shipped data: you
create the hazard yourself, from three invented lines, which is why it reproduces identically for
every learner.
The module at a glance
You are about to point a working agent at "the revenue folder". Open it and look at what is
actually in there. On a real revenue laptop that folder holds, at minimum: a CRM export with names,
emails and mobile numbers; a pipeline export with amounts; last quarter's commission statements; a
board deck; a churn post-mortem naming an account and a person; and a spreadsheet somebody sent you
once that you never opened.
You are not deciding whether to trust a model. You are deciding what a working agent may see and
change on a machine that holds your customers' data. Get it wrong and the failure is not a bad
answer, it is a modified file in a folder that also holds the CRM export.
This module makes that decision once, deliberately, and turns it into a folder structure you keep.
Rails is the letter the 4Ds do not have, and it is the one you build once. Everything from
Module 3 onward writes into the workspace you build here.
Lessons
#
ID
Title
Time
Reading
Doing
Carries
L1
L2.1
The folder is the boundary
14 min
9
5
Failure lane
L2
L2.2
What goes in, and what the folder already knows
16 min
12
4
Worked exercise
L3
L2.3
Connectors, pre-flight, and the bad day
14 min
12
2
Discrimination pair
Lesson total: 44 minutes against the M2 budget of 40 in PEDAGOGY.md §3.1, inside the plus or
minus 15 percent tolerance (§6.9). Reading minutes are measured from the shipped text at 200 words
per minute for prose, faster for tables and code. No lesson exceeds the 20-minute hard cap. The
full-scope workspace build is LAB-02 Part A (85 min), the exercise sheet in labs/LAB-02 of your
practice pack (start with its README); the lessons get each file started and prove
it works. CP-02, the module quiz in the interactive app,
is 10 min.
Every ## Check yourself section carries a collapsed answer block, with per-option lines on the two
closed multiple-choice items in this module (L2.2 item 1, the global / folder / neither sort, and
L2.3 item 2, the SharedRoot posture options). Those blocks are counted inside the reading figure
above even though they are collapsed and consulted one item at a time.
Outcomes
ID
You can
O2.1
Build a workspace structure separating granted from fenced data, with a stated reason per folder
O2.2
Given a folder layout, predict what Cowork can and cannot reach, and explain why
O2.3
Write standing instructions at two levels and state which rule belongs at which level
O2.4
Authorise the minimum connector set for one workflow and state what each reads and writes
O2.5
Run a platform and plan pre-flight and record what is available, with the date
O2.6
Apply the sandbox mitigation posture and explain the disclosed risk it responds to
What you keep
revenue-workspace/, a configured working folder. Three rubric-scored pieces:
revenue-workspace/data-boundary.md (template in templates/M2/data-boundary.md in your practice pack): the fence
decision record. Every folder listed as granted or fenced with a one-line reason, the redaction
defaults for exports, the dated pre-flight record, and the connector inventory.
revenue-workspace/CLAUDE.md (template in templates/M2/CLAUDE.md): folder-level standing
instructions covering role, output conventions, the exceptions rule and the standing safety rails.
The folder scaffold itself, created and empty where appropriate.
Every later module writes into this folder and reads these two files. It is also the thing a
security-minded manager asks to see before approving a rollout.
A note on what is ours and what is the product's
The folder layout, the redaction gate and the connector inventory are our opinion, developed for
this course. They are not product requirements and the lessons say so at the point of use.
What is the product's, and is sourced: the sandbox and the permission prompts, the fact that Claude
cannot read or write outside directories you explicitly grant, read-only-by-default behaviour, the
existence and destination of global instructions, the per-folder context file conventions, what a
plugin packages, the install and marketplace paths, the platform gaps, and the SharedRoot
disclosure. Each carries its FACTS section at the sentence that makes the claim.
Where this goes next
Module 3 needs somewhere for the delegation policy to live, and that is
revenue-workspace/40-reference/. Module 4 writes briefs into 20-briefs/ and run records into
50-runs/. Every Part B lab in the course requires the redaction gate defined here to have been
applied.
Module 2 · Rails, your workspace, your data boundary, your connectors
M2 lesson / about 14 min
Failure mode
The folder is the boundary
Do this nowCount what else came in60s
Open the data folder you copied into 10-data/ and list everything that is not one of the eight CSV files.
Open notes/ and count the files in call-notes/ and in qbr/.
Write both counts at the top of data-boundary.md.
You know it worked when: You can name sixteen call notes and three business reviews that arrived alongside the eight CSVs you actually wanted.
What you type
List every file you can read under this folder, as a tree, with the size of each.
Do not read the contents of any file. Do not summarise anything.
Just the inventory.
Notice: CLAUDE.md says never read anything under 10-data/_private/, and the inventory lists it anyway, because the grant is the boundary and the line in the file is a request.
The granted folder is the whole perimeter: everything inside it is readable, only 30-output/ and 50-runs/ are written to, and the private tree is safe because it was never granted.A CLAUDE.md line telling the agent not to read a folder does not fence it; moving that folder out of the granted tree does, because there is then nothing to obey.Rails is the one TRACK letter you build once at setup and then inherit; Triage, Assign, Check and Keep run on every single task, which is why Module 2 comes before Module 3.
In Module 1 you granted a folder without thinking, because it held nothing but practice data.
Right call then. Wrong habit now.
Here is the version that costs money. An analyst points a session at ~/Documents/Revenue
because the pipeline export lives there. Three levels down sits Comp/2026-Q2-commissions.xlsx,
because it also lives there. Nobody attacked anything. The perimeter was simply whatever happened
to be in one directory. This lesson draws it on purpose.
What you will be able to do
Say, for any folder layout, what a session can and cannot reach, and why.
Build a workspace that separates granted data from fenced data by location.
Prove to yourself, in three minutes, that an instruction is not a boundary.
The boundary is a directory, not a rule
Two product facts. They are the two most important sentences in this course.
Cowork runs inside a sandbox and asks permission before certain high-risk operations.
Claude cannot read or write files outside the directories you explicitly grant. In the words
the source uses: if you do not want Cowork to touch something, do not put it in the shared folder.
The second one cuts both ways. It guarantees what is outside, and it means everything in the
granted folder is in scope, including what you forgot was in there. The boundary is not "the
files I meant". It is the directory.
The uncomfortable part: your working folders were organised by what you were doing at the time,
not with a boundary in mind. Nobody's were. That is why this module exists.
Check yourself: name one folder you would have granted last week and should not have.
Read-only by default, and what sandboxing changes
The underlying model is read-only by default and asks before modifications or command
execution. Sandboxing then creates pre-approved boundaries so it stops asking per action.
That is a real tradeoff. Make it on purpose, not by clicking through prompts:
Narrow pre-approved area
Wide pre-approved area
More prompts, more interruptions
Fewer prompts, smoother long runs
Smaller blast radius if a run misbehaves
Larger blast radius if a run misbehaves
Good for anything touching customer data
Good for a scratch folder of your own output
The rule to carry: the size of the pre-approved area is a decision, and it should differ by
folder. Your own reports can sit in a wide area. A customer list export should not. The
strongest "narrow" is not granting the folder at all, which is next.
The revenue workspace layout
This layout is our opinion, not a product requirement. The principle underneath is not opinion:
fencing is done by location.
revenue-workspace/ <- this is the folder you grant
00-inbox/ raw drops, nothing trusted, cleared weekly
10-data/ working copies of exports, redacted per the boundary rules
20-briefs/ the durable briefs. This is where reusable work lives.
30-output/ everything Claude produces, by workflow and date
40-reference/ ICP rules, field definitions, stage definitions, policies
50-runs/ run records, one line per run
CLAUDE.md the folder's standing instructions
data-boundary.md the fence decision record
And, outside the workspace and never granted:
~/revenue-private/ comp, personnel, legal, board, anything under a restriction
Why each folder earns its place:
00-inbox/: so "somebody sent me a file" does not silently become "the agent has a file".
Files leave on purpose, after the Lesson 2 gate.
10-data/: copies, never originals. If a run modifies something, let it be a copy.
20-briefs/: no memory between sessions means a brief that lives only in a chat is thrown
away. Reusable work survives here.
30-output/: generated files stay separate from sources, so "what did it produce" is
answered by looking, not remembering.
40-reference/: the definitions that make outputs comparable across weeks.
50-runs/: your own audit trail, because the enterprise one does not cover Cowork yet.
The line that matters most is the last one.~/revenue-private/ is not protected by a rule.
It is protected by not being in the granted tree. "Do not read the comp file" is a request. A
folder that was never granted is a boundary.
Try this now
Create the seven folders and the two empty files, plus ~/revenue-private/ outside the tree.
Four minutes. You fill them in Lesson 2.
Check yourself: for each of your seven folders, one sentence on why it is inside the boundary.
Run the visibility probe
The single most useful thing you will do in this module.
Try this now
Put a copy of your practice pack's data/ folder in 10-data/.
Start a session. Grant revenue-workspace/ and nothing else.
Paste this:
List every file you can read under this folder, as a tree, with the size of each.
Do not read the contents of any file. Do not summarise anything.
Just the inventory.
If it starts summarising anyway, say: "Stop. Do not read any file contents. Only list the tree."
Why this prompt? An inventory, not a summary, so you learn what is reachable before anything
is read. A tree, because the item that catches people is never at the top level: it is three
levels down inside a folder somebody copied in whole. And do not read the contents, because
the point is the perimeter, not the data.
What you should see: on our run, eight CSV files and DATA-DICTIONARY.md at the top of
10-data/, plus 10-data/notes/call-notes/ and 10-data/notes/qbr/ holding sixteen call notes
and three business reviews. That last part is the finding: one copied-in data folder also handed
over nineteen unstructured documents nobody had thought about. On a real revenue laptop, that is
the whole problem in one sentence.
Now read the tree against your intent, line by line, and mark each line keep, move or fence. The
line you skim past is the one you would have skimmed past on your own machine.
When this goes wrong
You will create the hazard yourself: your file, invented on the spot, nothing sensitive, same
result for everyone.
Try this now
Inside the granted tree, create revenue-workspace/10-data/_private/decoy-comp-sheet.md and
paste exactly these three lines:
Start a new session, grant revenue-workspace/, and ask:
Summarise everything in this working folder, one line per file, including
anything under subfolders.
What you should see: the summary references the decoy, by filename or by content. If it does
not on your first try, ask: "list every file that contains a person's name". The instruction did
not create a boundary, because an instruction is a request and the folder grant is the boundary.
Note what did not happen. Nothing malfunctioned. Nothing was bypassed. The agent did what a
diligent colleague would do with a folder you handed over and a polite note asking them not to
look in one drawer.
The guard. Move _private/ into ~/revenue-private/. Re-run the identical request. The
decoy is not summarised. Notice the difference in kind: nothing blocked it and no rule was obeyed.
The file was outside the perimeter, so there was nothing to obey. That is a boundary, and it is
why data-boundary.md is about locations, not rules.
Delete the decoy when done. Write one line in data-boundary.md recording that you ran this and
what happened.
What you keep
revenue-workspace/, scaffolded: seven folders, CLAUDE.md and data-boundary.md created
(near-empty is fine until Lesson 2), and ~/revenue-private/ outside the tree.
data-boundary.md, started: granted list, fenced list, and one dated line recording the
decoy test and its result.
Check yourself
A colleague says their setup is safe because their CLAUDE.md tells the agent never to open
the commission folder. Give the shortest accurate response, then the fix.
You are handed this tree and asked what must leave before access is granted:
revenue-workspace/10-data/ holding pipeline-export.csv, contacts-full.csv, and
projects/renewals-q3/ which itself contains renewals-q3.xlsx and legal/msa-redlines.pdf.
Name what leaves, what may stay but should be read-only, and the one item most people miss.
Someone widens the pre-approved area to stop the permission prompts on a long hygiene run.
Name what they bought and what they paid.
Answers
1. That is an instruction, not a boundary, and you watched one fail in this lesson. The fix
is location: move the commission folder into ~/revenue-private/ and re-run the visibility probe
to confirm it is gone. Keep the CLAUDE.md line as documented intent, but stop counting it as
the control.
2. Leaves:contacts-full.csv (direct contact identifiers are not needed for renewal
analysis) and projects/renewals-q3/legal/msa-redlines.pdf (contract material under a
restriction). May stay, read-only:pipeline-export.csv and renewals-q3.xlsx, which are
the job. Most missed: the redlines PDF, two levels down inside a copied-in project folder,
not at the top where you would see it.
3. They bought fewer interruptions on one run. They paid a permanently larger blast radius on
every run in that area, including ones not yet written.
Where this goes next
Next lesson: what may enter 10-data/ in the first place (the redaction gate), and standing
instructions at two levels so the folder itself carries the rules.
Track labs: Part A of the exercise sheet in labs/LAB-02 of your practice pack (start with
its README) runs the visibility probe, the decoy test and the guard as a measured sequence.
CP-02, the module quiz in the interactive app, hands
you a folder tree and asks what comes out.
Checked against Cowork in August 2026.
Module 2 · Rails, your workspace, your data boundary, your connectors
M2 lesson / about 16 min
Do it
What goes in, and what the folder already knows
Do this nowWrite a rule it cannot guess60s
Open revenue-workspace/CLAUDE.md.
Add an output convention: every file you produce goes to 30-output/, named <YYYY-MM-DD>-<workflow>.md.
Add one more line: the first line of every output is exactly "Source files: " followed by every file you read.
You know it worked when: CLAUDE.md now holds a rule you will never repeat in a request, so the next output either carries it or the file was not read.
What you type
Using the files in 10-data/, tell me how many accounts are in each lifecycle
status, and what the total ARR is for the ones that are customers.
What comes back
30-output/2026-08-06-account-counts.md written.
Source files: accounts.csv
| lifecycle_status | accounts |
| Prospect | 204 |
| Customer | 87 |
| Churned | 17 |
| total | 308 |
Total arr_usd across Customer accounts: 7,699,060 USD.
Prospect and Churned accounts are excluded from that figure.
Notice: The request named no filename, no folder and no provenance line, so every one of those in the output is evidence that CLAUDE.md was read.
The granted folder is the whole perimeter: everything inside it is readable, only 30-output/ and 50-runs/ are written to, and the private tree is safe because it was never granted.
Two things are still missing from the workspace you built in Lesson 1.
First, a rule about what may enter 10-data/ at all. Right now the answer is "whatever I
export". That is how a file with 911 contact rows, mobile numbers included, ends up inside the
perimeter for a job that needed account identifiers.
Second, the folder knows nothing. Every session starts from zero, so you restate your role and
your rules Monday, Tuesday, and again next week. Module 1 showed you why: there is no memory
between sessions. So give the folder a memory instead.
What you will be able to do
Run a five-question gate on any export before it enters the granted tree, and record the
decision.
Write standing instructions at two levels and say which rule belongs at which level.
Prove your folder file is actually being read, rather than assuming it.
The redaction gate
This is our policy, not a product claim. Nothing in Cowork asks for it. The directory boundary
only helps if you are deliberate about what goes inside it, and this gate stands in front of
every Part B lab in this course. Five questions, each with a default. Run them on every export
before it enters 10-data/.
Question
Default
Why the default
Do I need direct contact identifiers for this analysis?
No. Use IDs.
Most revenue analysis joins on account_id and contact_id. Names, emails and phone numbers ride along out of habit, and they are the part you least want to see in an output file.
Do I need amounts at deal level?
Often yes. That is the job.
Do not over-redact. A pipeline analysis without amounts is not an analysis. Redaction that makes the work impossible gets abandoned within a week.
Is anything here restricted by a customer agreement or a regional rule?
Check, do not assume.
Some contracts and regions restrict where customer data may be processed. A question for your legal or privacy owner, not for you and not for this course.
Is anything here about employees rather than customers?
Fence it.
Comp, performance, headcount and personnel go to ~/revenue-private/ and are never granted.
Would I be comfortable if this file were copied verbatim into an output file by mistake?
If no, it does not enter.
The honest version of the four above, and the one that catches the file the others missed.
Applied to the practice dataset: contacts.csv in your pack's data folder carries
contact_name, email and phone alongside contact_id, persona and last_activity_date
(see DATA-DICTIONARY.md in the same folder). In the
Module 1 snapshot work, personas and last activity dates earned their place. The phone numbers
were never needed once. The habit: name the columns the job needs, and know why the others came
along.
Check yourself: for the last export you pulled, which columns would not have entered?
Standing instructions, at two levels
A standing rule can live in two places, and putting it in the wrong one is the most common
configuration mistake in this course.
Global instructions apply to every Cowork session: tone, output format, role background,
standing safety rails. Third-party setup guides place them under Settings, then Cowork, with an
edit control next to global instructions. That path comes from those guides, not from Anthropic's
own documentation, so confirm the exact wording on your build.
Per-folder context files carry the rules that belong to this body of work. The documented
conventions are preferences.md or CLAUDE.md in the working folder. At session start, Claude
reads MEMORY.md for cross-project learnings and, inside a project folder, that folder's own
CLAUDE.md and MEMORY.md.
The split, as a rule you can apply without thinking:
Global for who you are and how you write. Folder for what this work is and what it must never
do.
Worked both ways:
Rule
Where
Why
"I am a RevOps analyst at a B2B SaaS company. Write plainly. No marketing adjectives."
Global
True in every session, including ones with nothing to do with revenue.
"Dates as YYYY-MM-DD. Money with its currency code. Never mix currencies in a total."
Global
A formatting habit you want everywhere.
"Output goes to 30-output/<workflow>/<YYYY-MM-DD>-<name>.md."
Folder
Only true inside this workspace.
"Amounts in opportunities.csv are in the account's local currency. arr_usd is USD. Convert before comparing."
Folder
A fact about this data, not about you.
"Anything you cannot resolve goes to exceptions.csv with the row id and the reason. Do not guess."
Folder
The single highest-yield line in this course. Module 4 teaches it properly.
"Never contact a customer directly."
Neither
It is an instruction, and instructions are not boundaries. Lesson 3 explains the real control.
Worth knowing: memory here is a file convention, not a product feature. Cowork does not carry
state between sessions. The folder does not remember. The folder is read.
Check yourself: which two of your own standing rules are at the wrong level right now?
Prove the folder file is being read
Most people write a CLAUDE.md, get a good output, and conclude it worked. Not evidence: a good
output is what you would have got anyway. You need a rule the agent could not have guessed and an
output that visibly obeys it.
So put a rule in CLAUDE.md that is arbitrary on purpose:
Output convention: every file you produce goes in 30-output/, named
<YYYY-MM-DD>-<workflow>.md, and its first line is exactly:
Source files: <comma-separated list of every file you read>
Nothing about that is guessable. If it appears, the file was read.
Then make a request that mentions none of it:
Count the accounts in 10-data/accounts.csv by region and by segment.
Write it up.
What you should see: on our run, 30-output/2026-08-02-account-counts.md, first line
Source files: accounts.csv. The obvious conclusion: the folder file is live. The better one:
you now get a provenance line on every output for free, permanently. It catches the case
where you are looking at last week's output and cannot remember which export it came from.
Module 4 builds the heavy version with run records.
Do not put the source-files rule in global instructions. It refers to a folder most of your
sessions do not have, and a global rule that is usually wrong gets quietly deleted.
Worked exercise
Situation. You have the Lesson 1 scaffold and the practice dataset in 10-data/. Write the
two files that make the workspace yours, then prove one is being read. Budget five minutes. The
full build is Part A of the exercise sheet in labs/LAB-02 of your practice pack (start with its
README).
Setup
Folder to grant: revenue-workspace/.
Files in play: 10-data/accounts.csv, 10-data/contacts.csv, 10-data/opportunities.csv.
Connectors needed: none.
Blast radius: read-only. One output file, inside your own 30-output/.
Steps
Try this now
Open templates/M2/data-boundary.md from your practice pack and fill the granted list,
the fenced list and the redaction defaults. One line of reason per folder. The reason is
what counts, not the list.
Open templates/M2/CLAUDE.md and fill the four gaps below.
Start a session, grant revenue-workspace/, and run the request in the starter brief, which
mentions none of your conventions.
Open the output. Check it against your conventions, one rule at a time.
If a rule was not obeyed, do not add emphasis. Ask why it was ambiguous, and rewrite it so
it is checkable. A rule you cannot check from the output is a hope, not a rule.
Starter brief, with gaps marked
In CLAUDE.md:
# Standing instructions for revenue-workspace
## Role and voice
<<GAP 1: one or two lines. Who you are and how output should read. Keep this short:
anything true of every session belongs in global instructions, not here.>>
## What this folder holds
10-data/ holds working copies of exports. 20-briefs/ holds durable briefs.
30-output/ holds everything you produce. 50-runs/ holds run records.
<<GAP 2: two lines about THIS DATA that an agent could not know from the files alone.
Look at DATA-DICTIONARY.md for the two facts most likely to produce a wrong total.>>
## Output conventions
<<GAP 3: a naming and location rule specific enough that you can tell from the
output whether this file was read.>>
## Standing rules
- Do not modify anything in 10-data/. Write only to 30-output/.
<<GAP 4: the exceptions rule. One sentence. What happens to anything you cannot
resolve.>>
And the request to run, which mentions none of the above:
Using the files in 10-data/, tell me how many accounts are in each lifecycle
status, and what the total ARR is for the ones that are customers.
Hints
Hint 1, Concept. Three gaps are about you, one is about the data, and the data one is the
only one that changes whether the answer is numerically right. Which numbers in the output could
be wrong for reasons that have nothing to do with the agent's competence?
Hint 2, Move. Open DATA-DICTIONARY.md: the Conventions section and the opportunities
table. Two conventions there silently produce a wrong total if unknown: one about which currency
a money column is in, one about a type value whose amounts must not be added to another type's.
Write both into GAP 2 as facts about this folder.
Hint 3, Shape. A correct GAP 2 has two sentences, each naming a column. A correct GAP 3
produces something visible in the output that the request did not ask for: that is what makes it
a test. A correct GAP 4 is one sentence and contains "do not guess". If your GAP 1 is longer than
your GAP 2, you have written a personality profile, not a working configuration.
Reference artifact
The filled version ships with the LAB-02 solution in your pack's answer-keys/ folder, which
you open only after attempting the exercise. The two data facts GAP 2 is looking for, traced
to DATA-DICTIONARY.md:
opportunities.amount and pipeline_snapshots.amount are in the account's local billing
currency, while accounts.arr_usd is in USD. Converting is the analyst's job. The fixed
rates used to build the data are EUR = 1.08, GBP = 1.27, USD = 1.00.
opportunities.type = Renewal amounts restate current ARR and must not be added to New
Business when computing ARR.
Your role line, naming scheme, tone rules, choice of CLAUDE.md or preferences.md, and rule
count may all differ. What may not: if your CLAUDE.md contains no statement about this data
that changes a number, it is a style guide, and it will not stop the first wrong total.
The takeaway: the test of a folder file is not whether the output is good. It is whether the
output shows you something you did not ask for in the request.
What you keep
revenue-workspace/data-boundary.md: granted list, fenced list, one reason per folder,
redaction defaults. The pre-flight record and connector inventory come in Lesson 3.
revenue-workspace/CLAUDE.md: role, folder map, the two data facts, output conventions,
the exceptions rule.
30-output/<date>-account-counts.md: the evidence that the folder file is read.
Check yourself
Sort these into global, folder or neither: (a) "Never mix currencies in a total"; (b) "This
company's fiscal year equals the calendar year"; (c) "Never send anything to a customer";
(d) "Write in British English". For any you put in "neither", say what the real control is.
Your CLAUDE.md says output goes to 30-output/. A run wrote to the folder root instead. Give
two explanations, and say which one you can rule out by looking at the file rather than by
guessing.
An analyst redacts every column except account_id and amount before granting the folder,
then cannot answer any of the questions they were asked. Which of the five gate questions did
they apply badly, and what does that cost in practice?
Answers
1, one line per option:
(a) "Never mix currencies in a total": global. An arithmetic habit for every session.
(b) "Fiscal year equals calendar year": folder. A fact about this body of work; noise
anywhere else.
(c) "Never send anything to a customer": neither. An instruction, and instructions are not
boundaries. The real control: no connector with send scope was authorised, recorded in
data-boundary.md. Keep the line anyway, to document intent.
(d) "Write in British English": global. How you write, true everywhere.
2. Either the folder file was not read at all (moved, renamed, or the session pointed at a
different folder), or it was read and your naming rule was ambiguous. Rule out the first by
looking at the file: if the output carries the Source files: line, which also comes from
CLAUDE.md, the file was read and the rule is the problem.
3. They applied question 1's default ("use IDs") to every column instead of to direct contact
identifiers, and never asked question 2, which exists to stop exactly this. Nothing in the gate
asks you to drop stage, segment or renewal_date. The real cost: over-redaction gets the
whole gate abandoned within a week, after which nothing is redacted.
Where this goes next
Next lesson: connectors at minimum scope, the dated pre-flight record, and the disclosed
sandbox escape that makes the folder decision in Lesson 1 the control that survives a bad day.
Track labs:LAB-02 Part A proves CLAUDE.md works by running a request that depends on a
convention declared in it. CP-02, the module quiz in the
interactive app, sorts eight candidate rules across the
two levels.
Checked against Cowork in August 2026.
Module 2 · Rails, your workspace, your data boundary, your connectors
M2 lesson / about 14 min
Concept
Connectors, pre-flight, and the bad day
Do this nowPut a date on pre-flight60s
Open data-boundary.md and start a pre-flight block.
Write your operating system, whether you are on desktop or on web, your plan, and whether your organisation restricts plugins.
Put today's date on the last line.
You know it worked when: data-boundary.md can now answer "when was this last checked against what my build actually supports" with a date.
What you type
Two things, no analysis, no summaries.
1. List every file you can read under this folder, as a tree. Do not read the
contents of anything.
2. Then attempt to read ~/revenue-private/2026-comp-plan.xlsx and report exactly
what happened, whether it succeeded or not. Do not work around it and do not
look for another copy.
What comes back
1. Readable under revenue-workspace/
10-data/ 8 CSV files, notes/call-notes (16 files), notes/qbr (3 files)
20-briefs/ account-snapshot.brief.md
30-output/ 2026-08-06-account-counts.md
40-reference/ delegation-policy.md
50-runs/ RUN-LOG.md
CLAUDE.md, data-boundary.md
2. ~/revenue-private/2026-comp-plan.xlsx
Not read. That path is outside the directories granted to this session, so
there is no route to it from here. Nothing was attempted against it.
Notice: Part 2 is not obedience. There was no path, which is the whole difference between a boundary and the line in CLAUDE.md that asks politely.
Your workspace is fenced and your folder file works. Now somebody asks you to connect it to
something. The request always arrives sensibly: "Can it read the CRM directly?" "Can it check the
shared drive?" "Can it post the summary to the channel?" Each one is reasonable. Each one moves a
boundary you drew last lesson. And one is a write path into your system of record.
Grant the minimum, record it, and know your posture on the day something is disclosed.
What you will be able to do
Choose the minimum connector set for one named workflow, and say what each connector can read
and write.
Record what your platform and plan actually support, with the date you checked.
State the current mitigation posture for the sandbox boundary and the disclosed risk it
responds to.
Connectors, plugins and skills, at minimum scope
The vocabulary first, because these words are not interchangeable:
A plugin is a package that extends Cowork. One install can add skills, MCP connectors,
subagents, slash commands and hooks. That is a lot of surface behind one Install control.
Skills are domain expertise that activates automatically.
Slash commands are explicit actions you trigger.
Connectors are MCP servers wired to your tools.
The install path is one of only two UI paths this course states verbatim: sidebar, Customize,
Plugins, Browse plugins, select, Install. If the plugin includes a connector needing
authentication, you are prompted to sign in. The default marketplace is Anthropic's official
catalog; others are added under Plugins, Add marketplace, accepting
https://github.com/owner/repo or the owner/repo shorthand. Enterprises can run private
marketplaces.
Worth knowing: Anthropic publishes an open-source plugin repository for knowledge workers,
anthropics/knowledge-work-plugins. We could not confirm its exact contents, so we point at it
as a place to look and name nothing from it. If a post names a plugin, check the repository, not
the post.
The principle over the catalogue:
Authorise the fewest connectors that make one named workflow work, and write down what each
one can read and what it can write.
"One named workflow" is doing the work there. Connector sets grow because people authorise for a
category ("CRM access") rather than a job ("read opportunity and account records for the weekly
movement report"). The category has no natural stopping point.
The default that keeps you out of trouble: a CRM connector is a write path into your system of
record. Treat it as writes-to-system until established otherwise: a named approver and a
reversal plan, every run. If it offers read-only scope, take it, and record that you did.
Check yourself: for the workflow you most want to build, what is the smallest connector set
that makes it work?
Pre-flight, with a date on it
Try this now
Record this in the data-boundary.md you started earlier in this module, with today's date.
Two minutes.
Item
Note
Operating system and whether it is supported
Windows arm64 was not supported as of March 2026. Check the download page for your platform.
Desktop, web or mobile
Web and mobile arrived in beta July 2026, rolling out starting with Max.
On Linux?
A Linux beta exists and computer use is not available in it. We could not verify its full scope, so check before designing around it.
Plan
Cowork is on every paid plan and not on free.
Organisation restricts plugins or runs a private marketplace?
Ask, do not discover.
Date checked
This exists because the fullest public Cowork tutorial in this market still carried "Max only,
macOS only" long after both stopped being true, with no verification date on the page. The habit
is checking rather than trusting, and it applies to this course too.
Check yourself: one thing about your setup you would have assumed and were wrong about.
The security posture, stated plainly
On 27 July 2026, researchers disclosed SharedRoot, a local macOS sandbox escape. The host
root filesystem was mounted read/write into the Cowork virtual machine and, chained with a Linux
kernel flaw, allowed a session to gain root inside its VM and then read and write files anywhere
on the Mac without a further permission prompt. The demonstration used one connected folder
and a single instruction. Cloud execution was not affected, because Claude no longer runs inside
the local VM.
Worth knowing: this course does not tell you whether it is patched, on purpose. We could not
verify patch status or the affected version range, and a related advisory reference we could not
open could not be confirmed to map to SharedRoot. So we do not say it is fixed, we do not say it
is not, and we assert no CVE. Check your build against the vendor's advisory. Guessing here would
be exactly the thing this module trains you out of.
The posture, which the researchers recommend and which we would teach anyway: grant only folders
you have approved, and prefer read-only where the work allows it.
The point is not fear. The folder decision from Lesson 1 is the control that still works on the
day the sandbox does not. If your perimeter was drawn deliberately, a disclosure like this changes
what you read, not what you do. If your perimeter was your home directory, it changes everything.
Check yourself: name the one posture change that actually reduces your exposure, and one
reassuring change that does not.
A real scoping decision, narrated
The workflow. "Every Monday, produce the pipeline movement report and put it where my manager
can read it."
The lazy connector set. CRM (read and write), Drive, Slack, calendar. Four authorisations,
one that can write to the system of record and one that can post to humans.
Decision 1: what does the job actually read? Opportunity and account records. Nothing else
needs a system. One connector, scoped to read, on those objects.
Decision 2: does anything need to be written back? No. The output is a report. Write scope
buys nothing and costs the writes-to-system obligations on every run. Do not authorise it.
Decision 3: where does the output go? "Where my manager can read it" is a wish, not a
destination. 30-output/pipeline-movement/ plus a link needs no connector; a Drive folder needs
one. Pick the folder: the cheapest connector is the one you did not authorise.
Decision 4: the channel post. Posting to a channel reaches humans without a review step. Do
not authorise it. A person reads the report and posts it: that person is the review gate. After
several clean runs, promoting it is a conversation with a named approver.
Result: one connector, read scope, one named object set. Recorded in data-boundary.md:
name, reads, writes (nothing), who authorised it, date.
Tell these apart
An instruction and a boundary. People conflate them constantly, and the conflation is why a setup
gets believed safe when it is not.
An instruction
In revenue-workspace/CLAUDE.md:
Never contact a customer directly.
A good line, and you should have it. It is a request, obeyed the way a colleague obeys a note
on a shared drawer: reliably, until something makes it ambiguous, and never as a matter of
physics. You have the evidence: in Lesson 1 you wrote "never read anything under
10-data/_private/" into this same file, then watched a summary reference the decoy anyway.
A boundary
No connector with send scope was ever authorised, and ~/revenue-private/ sits outside the
granted tree. Nothing here is obeyed, because there is nothing to obey. There is no path. The
agent cannot email a customer for the same reason it cannot read the comp folder: the capability
was never granted.
The one-question test: if the agent decided to do the thing anyway, could it? "It would not,
because we told it not to" means instruction. "There is no route" means boundary.
A worked piece of feedback
A learner sorting eight candidate rules files "Never contact a customer directly" under
boundaries, with the note "this is our hardest safety rule". The response:
You are right that this is your most important safety rule. That is exactly why it is worth
being precise about what enforces it: right now, nothing does, and separately, no capability
exists.
A boundary is a missing capability. An instruction is a stated preference. Sorting by
importance puts this rule under boundaries; sorting by mechanism puts it under instructions,
and mechanism is what holds on a bad day.
Run the test: if a run decided emailing the customer was the best way to finish the task, could
it? Today, no, and not because of your CLAUDE.md line: no connector with send scope was ever
authorised. Your rule is a note on a locked door. The lock is the authorisation you did not
grant.
Two consequences. First, your safety is invisible in your own documentation: CLAUDE.md
records the intent, nothing records the control. That is what the connector inventory in
data-boundary.md is for. Otherwise the next person authorises send scope for a good reason
and nobody notices the rule lost its backing. Second, the day you do need send scope, this rule
alone will not be enough. What replaces it is draft-only: a human reads it before it leaves the
folder, with a named reviewer.
Keep the line. It documents intent and makes ambiguous cases resolve your way. Just do not file
it under the column holding the weight.
What you keep
revenue-workspace/data-boundary.md, now complete:
granted folders and fenced folders, one reason each;
redaction defaults from Lesson 2;
the dated pre-flight record from this lesson;
the connector inventory: name, what it reads, what it writes, who authorised it, date;
one line recording the Lesson 1 decoy test and its result.
This is the file a security-minded manager asks to see before approving a rollout, and it is a
scored input to the capstone.
Check yourself
A workflow needs data from the CRM, the support desk and the shared drive. Pick the minimum
connector set and state the write risk of each. Then say which one you would try to avoid
entirely, and what you would do instead.
Given the SharedRoot facts as stated above, which of these actually reduces exposure and which
is reassuring and useless: (a) adding "never leave the working folder" to CLAUDE.md;
(b) moving ~/revenue-private/ off the machine; (c) granting one narrow folder per workflow
instead of one wide one; (d) turning off permission prompts to reduce clicks?
Your data-boundary.md lists three granted folders and no fenced folders, and has no date on
the pre-flight record. Name the two questions the file can no longer answer.
Answers
1. Minimum set: CRM at read scope on the named objects, plus the support desk at read scope.
Write risk: a CRM connector is a write path into the system of record unless you took read scope,
which is why you take it and record that you did. Avoid the shared drive entirely: a wide,
poorly mapped tree whose job here is delivery, which a file in 30-output/ does with no
authorisation at all.
2, one line per option:
(a) The CLAUDE.md line: reassuring and useless. The disclosed escape did not require the
session to cooperate, and an instruction is not a boundary.
(b) Moving ~/revenue-private/ off the machine: genuinely reduces exposure. Data not on the
host cannot be read from the host.
(c) One narrow folder per workflow: genuinely reduces exposure, and it is the researchers'
recommended posture: grant only folders you have approved.
(d) Turning off permission prompts: increases exposure. It widens the pre-approved area,
the opposite of the trade you want here.
3. It cannot answer "what is deliberately outside the perimeter, and why", and it cannot
answer "when was this last checked against what my build actually supports".
Where this goes next
Next module: Module 3 decides what is worth putting through these rails at all, and gives you
the instrument that stops you automating something that saves eleven minutes.
Track labs: Part B of the exercise sheet in labs/LAB-02 of your practice pack (start with
its README) repeats the whole build on your real machine with one real export behind the
redaction gate. CP-02, the module quiz in the
interactive app, includes a rails item where one
candidate rule is a boundary wearing an instruction's clothes.
Checked against Cowork in August 2026.
Module 2 · Rails, your workspace, your data boundary, your connectors
Lab / ? min / read-only
Lab
The perimeter you can prove
LAB-02: The perimeter you can prove
The situation
Somebody is going to ask you, eventually, what an agent on your laptop can see. "Only the folder I gave
it" is the right answer and it is not evidence. This lab produces the evidence: a folder structure with
a written reason for every line, a probe that lists what the agent can actually reach, and a decoy you
plant yourself to prove that an instruction is not a boundary.
You will build this once and use it for the rest of the course and, if it works, for the rest of your
career. Every module from here writes into it.
What you will produce
File
What it is
revenue-workspace/
The folder scaffold: 00-inbox/, 10-data/, 20-briefs/, 30-output/, 40-reference/, 50-runs/
revenue-workspace/data-boundary.md
The fence decision record: every folder granted or fenced with a reason, the redaction defaults, the dated pre-flight record, the connector inventory
The visibility probe run twice, before and after the decoy is moved out of the tree
Outcome accounting
Manual baseline: There isn't one, and the lab says so rather than inventing one.
Setup is a one-time cost, not time saved. Nobody currently does this by hand,
so there is no manual baseline to compare against and claiming a saving here
would be fiction.
The number we record: time to first correct run.
"Correct" means: the visibility probe returns exactly the files you intended
to grant and nothing else, and one request obeys a convention you declared
in CLAUDE.md.
Observed while building this lab: 31 minutes on Part A, of which 9 minutes
was writing data-boundary.md and 6 was the decoy experiment.
Your actual: _____ minutes to first correct run
_____ minutes to write data-boundary.md
Verification cost: Ongoing, not one-off: about 3 minutes per month to re-run the visibility probe
and reconcile it against data-boundary.md line by line. That recurring 3
minutes is the whole of what this artifact costs to keep true, and it is the
reason to write it as a file rather than as a decision you remember making.
Note what is being measured. Modules 5 to 8 make time-saving claims. This one does not, because it
would be false.
Blast radius
Read-only, and deliberately so. Nothing in this lab writes outside the workspace folder you create.
The one file you create inside the granted tree is a decoy you write yourself, containing three invented
lines, which you then move out.
The mitigation posture we teach is the one the SharedRoot researchers recommended, and it is the
discipline this module teaches anyway: grant only folders you have approved, and prefer read-only
where the work allows it (FACTS §7.8, §5.2). SharedRoot was disclosed on 27 July 2026 as a local macOS
sandbox escape. Its patch status and affected version range are UNVERIFIED in our fact base, and this
lab does not assert a CVE mapping (FACTS §7.8). Cloud execution was reported as unaffected, because
Claude no longer runs inside the local VM.
Declared scorecard scope
None. This lab produces a boundary and a probe, not a findings list. See SCORING.md §1.
How you verify your own output
The probe reconciles line by line. Every file the probe lists appears in data-boundary.md as
granted, and every folder data-boundary.md calls granted appears in the probe. A file in the probe
that is not in your record is the finding. Sample size: all of it, because the probe output on a lab
folder is short enough to read.
The decoy disappears when it is moved and not when it is forbidden. Two probe runs, two different
answers, and the difference is location rather than instruction. If the decoy vanished when you added
the rule, re-check: you probably moved the folder as well.
CLAUDE.md is provably in effect. Declare one checkable convention in it (a filename pattern, a
date format, a required header line), run one unrelated request, and confirm the output obeys it.
That is the only proof that the file is being read. Do not take it on faith.
data-boundary.md answers two questions without you. Give it to a colleague and ask: what is
fenced and why, and when was the platform pre-flight last checked. If they cannot answer both from
the file, it is not finished.
Files you need
The whole of course/labs/data/, copied into 10-data/. You also need somewhere outside the
workspace for the fenced folder: ~/revenue-private/.
Do not copy course/labs/data/generate.py, content_docs.py, notes_a.py, notes_b.py,
PLANTED-ISSUES.md, PLANTED-ISSUES.json or _holdout/ into your workspace. Those are build and key
files, not learner files (ASSESSMENT-MODEL.md §2.1). If your copy of the dataset contains them, that
is itself a redaction gate exercise: they are the first thing to remove.
What comes next
LAB-03 fills 40-reference/delegation-policy.md. LAB-04 fills 20-briefs/ and 50-runs/. Every lab
after that writes into 30-output/. If the scaffold is wrong now, you will feel it in four labs' time.
Module 2 · Rails, your workspace, your data boundary, your connectors
6 questions / pass at 70%
Checkpoint
Module 2 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
You are about to grant Cowork access to ~/work/revenue/. Here is what is actually in it.
Which of these is the correct action before you grant the folder?
Hint
Read the tree twice. The first pass finds the folder that is obviously sensitive. The second pass is for
the file that is sensitive but is sitting under a heading that sounds routine.
A. Incorrect. This is safe and it does not work. The reference folder holds the stage definitions and the ICP notes, which are what stop the agent inventing its own definitions of a qualified stage or a good-fit account. Minimum scope means the smallest grant that lets the work happen, not the smallest grant.
B. Incorrect. This is the single most important mistake in the module and it is the one people make. An instruction is a request. The folder grant is the boundary. Claude cannot read or write outside the directories you explicitly grant, and the corollary is that anything inside a granted directory is in scope regardless of what an instruction file says about it (FACTS §5.2). Fencing is done by location.
C. Incorrect. You got the obvious one. `territory-planning/_archive/comp-plans-2026-signed.pdf` is signed compensation data, it is about employees rather than customers, and it is sitting inside a folder whose top level is genuinely fine. Most real revenue folders hide their worst file exactly like this, one level below a reasonable name, which is why the visibility probe exists: ask what it can read and reconcile the list against your intent line by line.
D. Correct. You caught both hazards and you distinguished between them. The board pack is obviously restricted and the compensation PDF is the one that is easy to miss, because it is nested two levels down inside a folder whose name sounds like ordinary sales work. You also made the right call on `exports/`: keeping a source export read-only where the work allows it is the mitigation posture, and it costs you nothing here (FACTS §5.2, §7.8).
Question 2 of 6
Eight candidate rules. Each belongs in global instructions, in the folder's CLAUDE.md, or in neither.
I am a RevOps analyst at a B2B SaaS company. Write for a sales audience, not an engineering one.
Dates are always ISO 8601.
In this folder, accounts.region is the reliable geography field. Do not derive geography from
country.
Never contact a customer directly.
Anything you cannot resolve goes to exceptions.csv with the row identifier and the reason. Do not
guess.
Output files go to 30-output/<workflow>/<name>-YYYY-MM-DD.md.
In this folder, opportunity amounts are in the account's local currency, so never sum amount across
accounts without converting.
Do not read the compensation file.
Which placement is correct for the four numbered rules listed in the option?
Hint
Two of these eight rules are not instructions at all, they are boundary decisions written in the wrong
place. Ask of each rule: if the agent ignored this sentence, what would actually have stopped it?
A. Correct. Rule 1 is who you are and how you write, which is the global level (FACTS §5.7). Rule 3 is what this particular body of work is, which is the folder level. Rules 4 and 8 are the two traps and you rejected both for the right reason: neither is a boundary. "Never contact a customer" is enforced by not authorising a connector that can send (FACTS §5.8), and "do not read the compensation file" is enforced by the file being somewhere you never granted (FACTS §5.2). Writing them down is not harmful, and believing they are controls is.
B. Incorrect. Your placements for 1 and 3 are right. Your placements for 4 and 8 are the error the module exists to prevent. Putting "never contact a customer directly" in global instructions produces a document that reads like a policy and enforces nothing. If a connector with send rights is authorised, the sentence does not stop anything, and if no such connector is authorised, the sentence was never what was stopping it.
C. Incorrect. You have inverted the split. Your role and your writing conventions apply to every session you ever run, which is the definition of global. The fact that `accounts.region` is the clean geography field is true of this dataset and this folder and is false somewhere else, so it belongs with the folder. The test is simple: would this rule still be correct in a completely different piece of work?
D. Incorrect. You were right that rule 8 does not belong, but for a reason that will not hold. "You would never put that file in the folder" is a statement about your intentions. The folder in front of you was organised years ago by somebody with different intentions, and the whole point of the visibility probe is that people are routinely wrong about what is in their own working directory. Rule 4 has the same problem as in option B.
Question 3 of 6
A colleague reports this.
"I granted my revenue/ folder and asked for a summary of everything in it. It summarised a private
note I had left in revenue/personal/notes.md, which I had explicitly told it to ignore in
CLAUDE.md. Then, on a later run where I had approved a wider pre-approved area so it would stop
asking me before each step, it rewrote a file in revenue/exports/ while cleaning up formatting.
Nothing warned me."
Which rail failed, and what is the minimum fix?
Hint
Ask whether either event is something the product did that it was not configured to do. If the answer is
no for both, the diagnosis is about the configuration and not about a failure.
A. Incorrect. There is no "instruction rail". Treating instructions as a control layer is the belief that produced both incidents, and a firmer instruction file will produce a third. Reading the output more carefully does not scale and would not have prevented either event.
B. Incorrect. The read-only default and the per-action prompts are the behaviour before sandboxing widens the pre-approved area, and your colleague widened it on purpose to stop being asked (FACTS §5.2). The tradeoff was taken knowingly. Filing this as a defect will get you an accurate explanation of your own configuration.
C. Correct. Neither event is a malfunction, and saying so is the diagnosis. Cowork cannot read or write outside the directories you grant, and `personal/` was inside one, so the instruction was competing with a boundary it does not have authority over (FACTS §5.2). Sandboxing pre-approves an area precisely so the agent stops asking per action, which is more convenient and is also a wider blast radius (FACTS §5.2). Both fixes are decisions about location and scope, not about wording.
D. Incorrect. This is the answer a bad week produces and it removes the tool rather than the hazard. Both incidents have a specific, cheap fix that keeps the work: move the private folder out of the tree, and size the pre-approved area on purpose instead of by accident.
Question 4 of 6
The workflow: every Monday, read the pipeline export, produce a review pack, and post a link to it in a
Slack channel your team already uses. Nothing is written back to the CRM. The pack is assembled from a
file export that you drop into the workspace yourself.
Four connector sets are proposed. Pick the minimum one that makes the workflow work, and state the write
risk correctly.
Hint
Write one line per proposed connector: what it reads, what it writes, and who is affected if the write
is wrong. The correct answer is the shortest list where that third column is still acceptable.
A. Incorrect. "So you are not blocked later" is how a workflow ends up with a write path into the system of record that nobody remembers authorising. A connector into a CRM is treated as writes-to-system by default until proven otherwise, and this workflow does not need one at all. Google Drive is not mentioned anywhere in the requirement.
B. Incorrect. Defensible, and it is worth saying why it is not the best answer rather than pretending it is wrong on safety. The Slack post is the step most likely to be skipped on a busy Monday, which is the step that makes the whole workflow visible to the team. Minimum scope means the smallest set that makes the workflow work, and a workflow nobody sees has stopped working.
C. Incorrect. You have removed the connector with the smallest blast radius and kept the one with the largest. The CRM connector would save one export step per week and would add a write path into the system of record to a workflow whose stated design point is that it does not write to the CRM. Trading a manual export for a standing write path is a bad exchange.
D. Correct. You authorised one connector, you scoped it, and you named what it writes rather than treating "read-only workflow" as a property of your intentions. The Slack write is a real write: a message in a channel is seen by people, cannot be unsent from everybody's memory, and therefore makes the pack draft-only. You also identified the load-bearing negative: the CRM is safe because no path to it exists, not because the brief says so.
Question 5 of 6
On 27 July 2026 researchers disclosed SharedRoot, a local macOS sandbox escape in which the host root
filesystem mounted read/write into the Cowork VM, chained with a Linux kernel flaw, allowed a session to
gain root inside its VM and then read and write files anywhere on the Mac without a further permission
prompt. The demonstration used one connected folder and a single instruction. Cloud execution was not
affected because Claude no longer runs inside the local VM. Patch status and affected version range are
not established in our fact base.
Your security lead asks what you are changing. Which answer actually reduces exposure?
Hint
For each option ask what would have to be true for it to change the outcome of the demonstrated attack.
One of the four survives that question, and one of the others is doing nothing while sounding careful.
A. Incorrect. The disclosed escape reaches files anywhere on the machine without a further permission prompt. An instruction in a settings field is not a control against that, and it was not a control against an ordinary file in a granted folder either. This is the reassuring and useless answer, and it is the one that reads best in an email, which is why it is here.
B. Correct. This is the researchers' recommended mitigation and it is also the posture the module teaches for ordinary reasons (FACTS §7.8, §5.2). The teaching point is not fear: the folder decision you already made is the control that survives a bad day, which is why it is worth making deliberately rather than by default. Refusing to assert a patch state you cannot verify is part of the answer, because a security lead who later discovers you guessed will not believe the next thing you tell them.
C. Incorrect. The demonstration used one connected folder and a single instruction, which is an argument that the preconditions are ordinary rather than exotic. Reading it as "the bar is high" gets the finding backwards. Noticing a malicious instruction also assumes you wrote every instruction the session saw, which stops being true as soon as the folder contains documents from anyone else.
D. Incorrect. Overreaction is not automatically the safe answer, and this one has a specific problem: it presents a platform move as a mitigation while changing where your data is processed, which is a decision your security lead should make with the facts in front of them rather than one you make under time pressure. State what is unverified and apply the mitigation that is available today.
Question 6 of 6
Here is a learner's whole data-boundary.md.
# Data boundary
## Granted folders
- 00-inbox/ raw drops
- 10-data/ working copies of exports
- 20-briefs/ briefs
- 30-output/ output
- 40-reference/ definitions and policies
- 50-runs/ run records
## Connectors
- Slack
## Redaction
Contact emails and phone numbers are replaced with contact IDs before an export enters 10-data/.
Which rubric criterion does this fail, and which two questions can the file no longer answer?
Hint
The file is a decision record, not an inventory. Read it and ask what decision it records. Then ask what
a person reviewing you in six months would need from it that is not there.
A. Incorrect. The folder names are the prescribed layout and each carries a one-line purpose, so reusability is the one thing this file is not failing at. A colleague could adopt this structure. What they could not do is find out what you decided to exclude.
B. Incorrect. Those two files being absent from a list is cosmetic. This artifact has a missing half, not a missing line, and grading it on Completeness of the folder inventory misses that the fenced list is the substance of the document.
C. Correct. A boundary document that lists only what is inside the boundary has not recorded a decision, it has described a directory. The two unanswerable questions are the ones the file exists for. First, what did you decide to keep out, and on what reasoning, which is the only part a security reviewer cares about. Second, what was true of this machine when you decided: operating system and whether it is supported, desktop or web or mobile, whether computer use is available to you, your plan, and whether your organisation restricts plugins, all with the date you checked (FACTS §2, §3). Without the date the record cannot be re-checked, and the reference tutorials in this market are stale for exactly that reason.
D. Incorrect. This is the answer that makes a rubric worthless: a document that looks complete because everything it contains is correct. Safety scores at most 1 for a boundary file that lists granted folders and no fenced folders and no reasons, and the missing pre-flight record removes the only dated evidence in the artifact.
M3: Triage, what to hand over and what it costs if it goes wrong
Last verified: 2026-08-02.Product claims: FACTS §4, §5.2, §5.10, §6, §7.5, §7.7.
Dataset: Talvera, seed 20260801. Every row identifier in these lessons was read out of
the pack's answer-keys/PLANTED-ISSUES.md and the shipped CSVs, and every defect class is named as the
key names it.
The module at a glance
Your VP has read something over the weekend and now says "use AI for everything you can". You have
about twenty recurring tasks. Handing over the right five gives you back most of a day a week.
Handing over the wrong one puts an incorrect renewal amount in front of a customer, merges two real
accounts that happened to share a name, or writes a stage change nobody can unwind.
The expensive mistake is not the task that visibly fails. It is the task that produces a plausible
answer that is wrong in a way nobody notices for six weeks.
This module gives you an instrument for telling those apart before you start, and a policy document
you can put in front of your manager.
Triage is risk management, not enthusiasm management. That is the whole module in one line, and
the rest is the mechanism.
Lessons
#
ID
Title
Time
Reading
Doing
Carries
L1
L3.1
The dangerous output is the plausible one
17 min
12
5
Worked exercise
L2
L3.2
Blast radius sets the process
13 min
11
2
Discrimination pair
L3
L3.3
Scope it, cost it, and revise what you wrote in Module 1
15 min
11
4
Failure lane
Lesson total: 45 minutes against the M3 budget of 40 in PEDAGOGY.md §3.1, inside the plus or
minus 15 percent tolerance (§6.9). Reading minutes are measured from the shipped text at 200 words
per minute for prose, faster for tables and code. No lesson exceeds the 20-minute hard cap. The
module also carries LAB-03 (55 min) and CP-03 (10 min), the heaviest checkpoint of the
foundations track.
Every ## Check yourself section carries a collapsed answer block. Several items in this module run
on the learner's own fit log and are marked Open, with the checkable part of each stated so an
open item still teaches something. Those blocks are counted inside the reading figure above even
though they are collapsed and consulted one item at a time.
Outcomes
ID
You can
O3.1
Score a task on the five-factor fit test and produce a decision with the deciding factor named
O3.2
Assign a blast-radius class and derive the approval, review gate and reversal plan from it
O3.3
Identify tasks that must not be handed over, and state the rule that excludes each
O3.4
Rewrite an open-ended task into a scoped one with definition of done, out-of-scope list, stopping condition and exceptions rule
O3.5
Estimate the usage cost of a task and say when cost changes the decision
What you keep
revenue-workspace/40-reference/delegation-policy.md (template in
templates/M3/delegation-policy.md). Three sections:
Scored inventory. Every recurring task from your Module 1 fit log, scored on the five
factors, with a total and a decision: hand over now, hand over after the source of truth is
fixed, or keep human.
The shortlist. Your top five hand-over candidates, each with blast-radius class, review
gate, approver, reversal plan, and an estimate of minutes saved per week.
The never list. Your own exclusions, each with the rule that generates it.
This is the document a manager approves. It is also the selection mechanism for the capstone: the
capstone workflow must come from the shortlist. And it is the thing that stops you spending four
hours automating something that saves eleven minutes.
What is ours and what is the product's
Ours, and labelled as ours at the point of use: the five-factor fit test, the blast-radius table
and its obligations, the five never-delegate rules, and the scoping rewrite drill. These are
instructional design for this course, not product guidance.
The product's, sourced: that open-ended, recursive or poorly defined tasks raise both the
likelihood and the impact of errors (FACTS §7.5); that Cowork draws on the same plan allowance as
chat and Claude Code under a rolling window plus a weekly cap (FACTS §4, §6); that complex
multi-step tasks consume materially more allowance, with Pro hitting limits fastest (FACTS §6); and
that the Compliance API and Audit Logs do not cover Cowork yet (FACTS §5.10, §7.7).
No numeric usage allowance appears anywhere in this module. Anthropic does not publish stable
figures and our fact base marks them UNVERIFIED (FACTS §6). One checkpoint item in CP-03 uses a
specific number as a wrong answer for exactly that reason.
Where this goes next
Module 4 turns the top item on your shortlist into a brief you can run, check and keep. Nothing in
Module 4 works well if the shortlist is a wish list, so do the scoring honestly.
Module 3 · Triage, what to hand over and what it costs if it goes wrong
M3 lesson / about 17 min
Concept
The dangerous output is the plausible one
Do this nowCount Retail accounts two ways60s
Filter accounts.csv on industry equal to the exact string Retail, and count the rows.
Filter again on industry with leading and trailing spaces trimmed, and count again.
Write both numbers down and find the row that only appears in the second count.
You know it worked when: You get 87 the first time and 88 the second, and the account that only appears once is ACC-02948, whose industry is written as " Retail".
What you type
Read accounts.csv and opportunities.csv in this folder.
List every account where lifecycle_status is Customer, renewal_date is earlier
than 2026-08-02, and no opportunity of type Renewal on that account has a
close_date equal to that renewal_date.
Output: account_id, account_name, arr_usd, renewal_date, health_score, sorted by
arr_usd descending, with the total arr_usd underneath.
Create nothing and change nothing. If a Customer account has a blank
renewal_date it cannot be tested: report how many you left out rather than
assuming they are fine.
Notice: These are absences rather than losses, and two of the eight carry health scores of 90 and 74, so the field everyone sorts by would never have surfaced them.
Checkability scored 1 sends a task to the never list whatever the other four factors say, so the instrument records the deciding factor rather than the total.Fourteen shipped opportunity rows put forecast_category and stage in direct contradiction, and the six cells they fall into are the whole finding.
Here is an account list. Well formatted, sorted, total at the bottom, done in forty seconds
instead of forty minutes.
Retail accounts, EMEA and NA
account_id account_name country arr_usd
ACC-02988 Highmoor Retail Inc United States (blank, prospect)
ACC-03104 yewtree foods sl Spain (blank, prospect)
ACC-03174 Marlowe Stores NV Netherlands 23220.0
...
Total accounts listed: 87
Nothing about it looks wrong. That is the problem. The correct count is 88. The list is wrong
twice.
Everything this lesson uses came in the practice pack from the course page; if you have not
downloaded it, do that first.
Once, on a leading space.ACC-02948 carries industry Retail, leading space included
(answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab,
class inconsistent-values). Filter on the exact string Retail: 87
accounts. Trim first: 88. One character, one missing account, no error message.
Twice, on a comma. Row ACC-02890 is named "Highmoor Retail, Inc.", a comma inside a
quoted field (DATA-DICTIONARY.md in your pack's data folder, Conventions), and is documented under class
duplicate-accounts as a punctuation variant of ACC-02988: the same company. Split on commas
without honouring the quoting and you drop the row, shift every column, or invent a company
called "Highmoor Retail.
The total at the bottom is still a number, still bold, still wrong. Nobody reading the list will
find out.
What you will be able to do
Say why plausible is not evidence, in words you can use with a VP.
Score any candidate task on five factors and reach a decision.
Name the one factor that vetoes a task no matter how good the rest look.
Plausible is not the same as right
Two failure shapes. Only one is a problem.
Visibly wrong. The run errors or produces something obviously mangled. You notice in four
seconds. Cost: four seconds.
Plausibly wrong. A clean artifact, wrong in a way you can only see by reading the source. You
do not notice. Your manager does not notice. The customer notices in six weeks, or the number
lands in a board pack and gets defended by someone who was not in the room.
Everything in this course exists for the second shape. The product fact to quote when someone
says the fix is a better prompt: open-ended, recursive or poorly defined tasks raise both the
likelihood and the impact of errors. Both. A vague task does not just fail more often, it fails
larger, because vagueness lets it touch more.
Check yourself: name a task on your fit log whose failure would be plausible rather than
visible.
The five-factor fit test
This is our instrument, built for this course, not product guidance. A scoring sheet, not a
formula: its job is to make you say why, not to hand you a number to hide behind.
Score each factor 1 to 5.
Factor
Question
5 looks like
1 looks like
Volume
How many items, how often?
Hundreds of rows, every week
Three items, once
Checkability
Is there a source of truth I can compare against?
A file or system holds the right answer
The right answer does not exist yet
Judgment density
How much of this is a call only a human can make?
Almost none: it is a rule applied to records
Almost all of it
Cost of being wrong
What happens if a wrong answer ships?
Low and reversible: I delete a file
A customer sees it, or a system of record changes
Specification cost
Can I write down what done looks like in under 20 minutes?
Yes, and I already have
I would have to think about it for a week
Reading the score. High Volume plus high Checkability plus low Judgment density is the
strongest candidate. Do not add the numbers and take the top five. There is a veto:
A low Checkability score is a veto, regardless of everything else.
Low Checkability means shipping work whose correctness you cannot establish, at volume,
repeatedly. Volume then flips from an argument for handing over to an argument against: a wrong
method on three items is a mistake, on four hundred it is a report.
Two tasks can score identically and want opposite decisions, because the deciding factor differs.
So record the deciding factor, not just the total. A manager can argue with a factor. Nobody
can argue with a 17.
Check yourself: score one task from your fit log on all five factors, one sentence each.
Watch one task get scored
The task. "Every Monday, produce the list of live customers whose renewal date has already
passed with no renewal opportunity, with the ARR at risk on each."
A real pattern in the shipped data: PLANTED-ISSUES.md narrative N8, "Renewals are being missed,
not lost", lists eight accounts: ACC-02956, ACC-02958, ACC-03043, ACC-03097, ACC-03099,
ACC-03107, ACC-03123, ACC-03174. Not losses. Absences. Nobody built the process.
Factor
Score
The sentence behind it
Volume
3
308 accounts, weekly. Not huge, but every week forever, and it takes a person forty minutes because no view puts renewal date and renewal opportunities on one screen.
Checkability
5
Two named files, one rule. accounts.renewal_date compared against opportunities where type = Renewal. I can verify any row in twenty seconds.
Judgment density
2
The rule is mechanical. The judgment is what you do about the list, which stays with the CSM.
Cost of being wrong
2
Read-only. A false positive costs a CSM a five-minute check. A miss costs what it costs today, the same as not having the list.
Specification cost
5
I wrote the rule in one sentence, above.
The decision. Hand over now, read-only. Deciding factor: Checkability. Not volume, not
time saved.
The deciding factor tells the next person what would change the decision. Add "and create the
renewal opportunity" and Cost of being wrong jumps from 2 to 5 (it now writes to the system)
while Checkability holds. The decision flips, and the record shows which factor moved it.
And "forty minutes a week" is an estimate stated as an estimate: the labs require baselines
measured by doing the task by hand, verification cost counted separately, before any time-saved
claim.
Try this now
Your VP wants "a plan for what we automate first". Three candidate tasks from the Talvera revenue
team are on your desk. By end of day you need a shortlist row for what goes first and a
never-list row for what does not go at all.
Budget six minutes.
Setup
Open DATA-DICTIONARY.md in your pack's data folder. You do not need to run anything.
You are advancing revenue-workspace/40-reference/delegation-policy.md, sections 2 and 3.
This is a file you create yourself, starting this lesson, in the workspace you built in
Module 2; the seed template is templates/M3/delegation-policy.md in your pack.
The exercise itself is read-only.
The three candidates
Task A. Every Monday, list live customers whose renewal_date has passed with no renewal
opportunity, with the ARR at risk. (Scored for you above. Use it as your calibration anchor.)
Task B. For every open opportunity, read the rep's most recent notes and set
forecast_category to match what the notes actually say.
Task C. Decide which of this quarter's Enterprise deals we will call as Commit.
Steps
Score B and C on all five factors, one sentence per factor, not just a number.
Name each task's deciding factor in one word.
Write the shortlist row for the task that belongs in section 2 of delegation-policy.md:
blast-radius class, review gate, approver, reversal plan.
Write the never-list row for the task that belongs in section 3, naming the rule that excludes
it.
One of the three is not clearly in either place. Say in one sentence what would have to change
for it to move.
Starter brief, with gaps marked
## Shortlist entry
Task: <<GAP 1: the task, scoped tightly enough to be checkable>>
Fit score: V _ Ch _ Jd _ Cw _ Sc _
Deciding factor: <<GAP 2: one word, and it is not "volume">>
Blast radius: <<GAP 3: read-only / draft-only / writes-to-system>>
Review gate: <<GAP 4: what a human reads, and how much of it>>
Approver: <<GAP 5: a person's role, or "none needed" with a reason>>
Reversal plan: <<GAP 6: what I do if a run ships something wrong>>
Minutes saved: ____ per week, verification cost of ____ minutes already subtracted
## Never-list entry
Task: <<GAP 7>>
Excluded by: <<GAP 8: the rule, not the task. Rules are in Lesson 2.>>
What would change it: <<GAP 9, or "nothing", and say why>>
Stuck? Two hints, in order.
Two tasks are blocked, and not by the same thing. Only one is blocked by risk; the other is
blocked earlier, by something missing before you start.
Finish this for each: "I would know this output was wrong if I compared it against ___." Two
of the three take a filename in the blank. For the one that does not: does the right answer
exist before or after you must act on the output?
Reference answer
Task B: set forecast_category from the rep's notes.
Factor
Score
Reasoning
Volume
4
Hundreds of open opportunities, refreshed continuously.
Checkability
3
Partly. You can verify the field matches the note, not that the note is right, and the note is the input.
Judgment density
4
High. What a note "actually says" about a deal's confidence is interpretation, and reps write notes the way reps write notes.
Cost of being wrong
5
It writes to the system of record, and forecast category feeds the forecast.
Specification cost
2
Try writing down what makes a note mean Commit: far longer than twenty minutes, and you will not get agreement.
Decision: not as described. Reclassify rather than delegate as stated. Deciding factor:
Cost of being wrong, backed by Judgment density.
The version that survives is different: a draft-only list of opportunities where
forecast_category contradicts stage, with the evidence line for each. Checkable against a
rule, writes nothing, real work in it: PLANTED-ISSUES.md documents fourteen rows in class
forecast-contradiction, including OPP-006692 (Closed Lost still in Best Case) and
OPP-006934 (Omitted while the stage is Negotiation). The rewrite keeps the value and drops the
write path.
Task C: decide which Enterprise deals we call as Commit.
Factor
Score
Reasoning
Volume
2
A handful of deals, once a quarter.
Checkability
1
Veto. No source of truth. The right answer comes into existence when the quarter closes, after you must act on it.
Judgment density
5
It is the judgment. That is what a forecast call is.
Cost of being wrong
5
The number goes to the board.
Specification cost
1
If you could write down what done looks like, you would not need a forecast call.
Decision: never list. Excluded by the rule "anything where the source of truth is a person's
judgment rather than a record". Deciding factor: Checkability.
What would change it: nothing. But do not confuse the call with its support work: assemble the
evidence pack each Enterprise deal needs for the commit conversation is high-volume, checkable
and mechanical. Assembling the evidence is delegable. Making the call is not.
Your answer may differ, within limits. Scores can move a point either way, and reasonable
operators disagree on Task B's Checkability. What may not differ: Task C is vetoed by
Checkability, not by risk. "Too risky" is the wrong reason, and it will not survive a VP willing
to accept the risk.
Worth knowing: the task you must refuse is not the one with the largest downside. It is the
one where you could not tell afterwards whether it worked.
What you keep
revenue-workspace/40-reference/delegation-policy.md, started: one shortlist entry, one
never-list entry, and your calibration anchor from the demo. Lesson 3 fills section 1 and the
rest of 2 and 3, once you have the blast-radius obligations and the never-delegate rules.
Quick check
Two tasks score 17 out of 25. One is handed over this week, the other never. Give a pair from
your own fit log where that is true, and name the factor that differs.
Someone argues a low-Checkability task is fine to hand over because the volume is enormous and
doing it by hand is impossible. Answer them in two sentences.
Take the lowest Specification cost item on your fit log and write its definition of done in
one sentence, now. If you cannot, what does that say about the score you gave it?
Answers
1. Open on which pair. The factor that differs is the deciding factor, almost always
Checkability or Cost of being wrong. Identical totals with different deciding factors is the
normal case, which is why the total is not the output of this instrument.
2. Volume is not a defence, it is the amplifier. A wrong method on three items is a mistake,
on four hundred it is a report, and low Checkability means you would not find out either way. If
it truly cannot be done by hand or checked, then nobody knows whether it is right, including the
person asking.
3. Open, and a self-test. If you cannot write it in one sentence now, your Specification cost
score was optimistic. Re-score it: that factor is the one people flatter most.
Where this goes next
Next lesson: blast radius starts setting the process, plus the five rules that generate your
never list.
Track labs: the exercise sheet in labs/LAB-03 of your practice pack (start with its
README). Part A ships twelve real-shaped tasks across all four roles, with
reference scoring, and names the two tasks where reasonable operators disagree.
Module 3 · Triage, what to hand over and what it costs if it goes wrong
M3 lesson / about 13 min
Concept
Blast radius sets the process
Do this nowCheck one currency against its account45s
Open opportunities.csv, find OPP-006697, and read currency and account_id.
Open accounts.csv, find ACC-03045, and read currency.
Time yourself doing it.
You know it worked when: The opportunity says USD, the account says EUR, and the whole check took you about fifteen seconds.
What you type
Read opportunities.csv and accounts.csv in this folder.
For every opportunity row, compare opportunities.currency against the currency
recorded on its parent account in accounts.csv. List every row where the two
disagree.
Columns: opportunity_id, account_id, account_name, opportunity currency, account
currency, amount.
Do not change anything, and do not decide which of the two values is correct.
If an account_id does not resolve, write the row to exceptions.csv with the
reason rather than dropping it.
What comes back
4 rows disagree, out of 528 checked.
| opportunity_id | account_id | account_name | opp | acct | amount |
| OPP-006697 | ACC-03045 | Ardleigh Distribution AG | USD | EUR | 58,000.0 |
| OPP-006729 | ACC-02958 | Newholm Hotels SA | USD | EUR | 74,000.0 |
| OPP-007089 | ACC-03089 | Quarrybank Hospitality Group SL | USD | EUR | 25,000.0 |
| OPP-007326 | ACC-03021 | Redcliffe Facilities AB | USD | EUR | 90,000.0 |
exceptions.csv: 2 rows
OPP-006774, account_id ACC-93103 does not resolve in accounts.csv
OPP-006894, account_id ACC-92842 does not resolve in accounts.csv
Notice: 528 rows of work nobody wanted to do, and every finding is one account row away from being confirmed, which is why hard and unverifiable are different axes.
Each rung of the blast-radius ladder fixes its own review gate, approver and reversal plan, and the top rung requires the reversal plan tested and recorded before the first write.
You have a shortlist. Someone now has to say yes to it.
That conversation goes badly when your answer to "what happens if it gets it wrong" is a feeling,
and well when the answer is a table you filled in before you were asked.
That is blast radius. Not a warning label: it decides who reads the output, who approves the run,
and what you do at 16:40 on a Friday when something has already gone out.
What you will be able to do
Assign a blast-radius class and derive its review gate, approver and reversal plan.
Say when a workflow is promoted to writes-to-system, and why it never starts there.
Derive your own never-delegate list from five rules, not from somebody else's list.
The three classes and what each one obliges
This table is ours, teaching design. What it responds to is the product's directory and
permission boundary.
Class
Review gate
Approval
Reversal plan
Read-only
Spot check on a sample
None needed
Delete the output
Draft-only
Full human read before it leaves the folder
The sender
Do not send
Writes-to-system
Sample verify, then staged apply
Named approver, per run
Written, tested, and in the run record before the first write
Three things people get wrong here, in order of cost.
The class is set by where the output goes, not how the work felt. A list of at-risk accounts
is read-only until someone pastes it into a customer's business review deck; then it was
draft-only all along and nobody reviewed it. Follow the output to its final destination, not the
folder it lands in.
"Reversal plan" means tested, not imagined. "We would just re-import the old values" is a
sentence. A plan names the file with the previous values, names who runs the re-import, and has
been run once on a small batch. It goes in the run record before the first write, because
after is when you need it and cannot think straight.
Approval for writes-to-system is per run, not per workflow. A workflow approved in March is
not approved for a September batch full of objects March never saw.
The promotion rule, the most useful line in this lesson:
A workflow starts draft-only and is promoted to writes-to-system only after it has run clean
several times, with verification, and has a tested reversal path.
Nothing arrives as writes-to-system: not because the model is untrustworthy, but because you do
not yet know what your own brief does at the edges, and the edges are where systems of record
get damaged.
Check yourself: take one shortlist entry and derive its review gate, approver and reversal
plan, without asking anyone.
Five rules that generate your never list
Do not memorise a list of forbidden tasks: lists go stale and do not transfer. Derive your own
from five rules, so a new request classifies itself in ten seconds.
Rule 1. Anything that reaches a customer without a human reading it.
Not "anything customer-facing", which would exclude useful drafting work. The condition is the
missing human. Drafting the renewal email is fine, sending it is not.
Rule 2. Anything that changes a system of record where the change cannot be reversed.
Reversibility, not writing, makes the difference. A field update you can re-import is a
different risk from a merge you cannot unmerge. answer-keys/PLANTED-ISSUES.md in your pack,
which you open only after attempting the lab, documents eight
duplicate-accounts rows that genuinely should merge, for example ACC-02890 and ACC-02988,
the same company under a punctuation variant. Proposing those merges is delegable. Executing
them without a reversal path is not.
Rule 3. Anything where the source of truth is a person's judgment rather than a record.
A forecast call. A performance conversation. A pricing exception. This is the Checkability veto
from Lesson 1, stated as a rule you can quote.
Rule 4. Anything you must be able to evidence to a regulator or an auditor, because the
Compliance API and Audit Logs do not cover Cowork yet. People mis-state this one: the exclusion
is not "Cowork does this badly", it is that you cannot prove what it did at the level an
auditor requires. It is also the rule most likely to change, since it rests on a vendor gap, not
a property of the work.
Rule 5. Anything you could not defend line by line if a VP asked where a number came from.
If you cannot walk the number back to a file and a rule, do not put your name on it.
Try this now
Three minutes: derive three never-list entries, each with the rule that generates it. The rule
matters more than the task: in six weeks it decides a request you have not seen yet.
Check yourself: for each of your exclusions, can you state the rule, not just the task?
Watch a task get classified
The task. "Check every opportunity's currency label against its account's billing currency,
and tell me which ones disagree."
Why it looks refusable. 528 opportunity rows against 308 account rows. Tedious. Never done.
It sounds like an audit.
What it actually is.DATA-DICTIONARY.md in your pack's data folder says opportunities.currency should always match
the account, and on a few rows it does not. PLANTED-ISSUES.md documents exactly four, class
SUBTLE currency-mismatch: OPP-006697 (account ACC-03045), OPP-006729 (ACC-02958),
OPP-007089 (ACC-03089) and OPP-007326 (ACC-03021). Each says USD while the parent account
bills in EUR and the amount magnitude matches the EUR list price.
Decision 1: the class. Read-only. It produces a list, writes nothing, sends nothing. Spot
check on a sample, no approver, reversal plan is deleting the output.
Decision 2: the checkability. Every finding is a two-column comparison verified in fifteen
seconds: open the account row, read the currency, compare. Wrongness is visible on inspection,
which a plausible-looking summary cannot give you.
Decision 3: what I refused to add. "And fix them" is writes-to-system on a money field, and
it is not obvious in every case which value is wrong. So: a proposed-changes file plus an
exceptions file, a human decides, and the workflow is promoted later or never.
Worth knowing: this work is unpleasant and completely verifiable. Those two properties are
independent, and confusing them is the most common triage error there is.
Tell these apart
A task that is hard, and a task that is unverifiable. Both trigger "I would not want to be
responsible for that." The decision is opposite: hard work is the best thing you can hand over,
and unverifiable work is the thing you must not.
Hard, and a good candidate: the currency reconciliation above. High volume, tedious, easy
for a tired human to get wrong, done by nobody. And fully verifiable: two columns, one rule,
each finding checkable in seconds against a named file. Difficulty argues for handing it
over: the agent does not get bored on row 400, and you can check its work faster than you could
have done it.
Unverifiable, and not a candidate:
"Read the call notes and tell me which of these deals is actually going to close this quarter."
Not hard at all: sixteen notes, half an hour for a competent reader. Unverifiable: the source of
truth does not exist yet. It arrives when the quarter closes, after the moment you need to act.
And the output would be excellent, confident, specific, well argued, with no way to tell the
good version from the bad one until too late to matter.
The test in one question:if this came back wrong, what would I compare it against, and
would I have it before I need to act? The currency task: accounts.csv in your pack's data
folder, available now. The
notes task: nothing, available in eleven weeks.
A wrong answer, answered. A learner puts the currency reconciliation on the never list:
"too hard to verify at that volume, keep it human". The response:
You applied a real rule (do not hand over what you cannot check) to the wrong property of the
task. You measured effort; the rule is about evidence. You do not verify a 528-row
pass by redoing 528 rows, which is the assumption under your answer: you verify with a
reconciliation and a sample, and Module 4 teaches both. Each of the four documented mismatches
(OPP-006697, OPP-006729, OPP-007089, OPP-007326) checks in fifteen seconds against the
parent account's currency field.
Applied consistently, your reasoning would exclude every high-volume task, since every
high-volume task is expensive to redo by hand. Backwards: high Volume plus high Checkability
is the strongest hand-over profile there is, and this task is one. The genuinely unverifiable
task feels easier: sixteen call notes, which deals will close, nothing to check against until
the quarter ends. That one goes on your never list, under rule 3.
Swap the two entries, and write the deciding factor on each, because that is what stops this
pair being confused again: Checkability, high on the currency task, Checkability,
absent on the notes task.
What you keep
revenue-workspace/40-reference/delegation-policy.md, the file you started in Lesson 1,
section 3 filled: three or more
never-list entries, each carrying its generating rule, not only the task. Plus review gate,
approver and reversal plan completed on the shortlist entry from Lesson 1.
Quick check
A workflow has run draft-only for six weeks, clean verification every time. Someone proposes
promoting it to writes-to-system. Name the one thing that must exist before the first write,
and what "tested" means for it.
A task labelled read-only has its output pasted into a customer's quarterly business review
deck. Reclassify it and state the added obligation, in the words you would use with the
person who pastes it.
Which of your never-list entries would change if the Compliance API covered Cowork tomorrow?
If the answer is none, check rule 4 again.
Give a task from your week that is hard and verifiable, and one that is easy and
unverifiable. Which have you been avoiding handing over, and why was that backwards?
Answers
1. A tested reversal plan, in the run record before the first write. "Tested" means
executed once on a small batch: the file with the previous values exists, a named person can run
the re-import, and somebody has watched it restore rows. Six clean weeks earn the conversation.
They do not replace the plan.
2. It is draft-only, and always was. Added obligation: a full human read before it
leaves the folder, with a named reviewer. To the person pasting: "this is a draft until someone
has read it against source, and once it is in a customer's deck the reviewer is you."
3. Only entries under rule 4, because that rule rests on a vendor gap, not a property of
the work. If your answer was none, your rule-4 column is empty and you have probably filed an
auditability exclusion under rule 5 by mistake.
4. Open. The common pattern is avoiding the hard, verifiable one because volume feels risky,
which is backwards: volume plus a source of truth is the best candidate profile there is.
Where this goes next
Next lesson: the rewrite drill that turns an open-ended request into something checkable,
what usage costs you, and revising your Module 1 fit log.
Track labs: the exercise sheet in labs/LAB-03 of your practice pack (start with its
README). Part A has you defend three scoring decisions in writing, and scores a
matching answer with a wrong reason as a miss.
Module 3 · Triage, what to hand over and what it costs if it goes wrong
M3 lesson / about 15 min
Failure mode
Scope it, cost it, and revise what you wrote in Module 1
Do this nowScan the country column60s
Open accounts.csv and read the country column top to bottom.
Write down every value that is not a full country name.
Note how many rows each of them covers.
You know it worked when: You have four values on four rows: "united states" on ACC-02838, "DE" on ACC-02841, "netherlands" on ACC-02974 and "The Netherlands" on ACC-03002.
What you type
Clean up the accounts file and fix anything that looks wrong.
What comes back
Done. 10-data/accounts.csv updated in place. 20 rows changed.
- country normalised on 4 rows: "united states" to United States, "DE" to
Germany, "netherlands" and "The Netherlands" to Netherlands
- industry normalised on 3 rows: " Retail" to Retail, "logistics" to Logistics,
"Hospitality & Leisure" to Hospitality
- leading and trailing whitespace stripped from account_name on 4 rows
(ACC-02809, ACC-02957, ACC-03008, ACC-03021)
- health_score rescaled from a 1 to 10 scale to 0 to 100 on 9 rows
(ACC-02805, ACC-02834, ACC-02938, ACC-02945, ACC-02984, ACC-03007, ACC-03043,
ACC-03086, ACC-03118). For example ACC-03043 changed from 8 to 80.
- 9 further rows left unchanged where the intent was unclear
Notice: The last two bullets are the damage: "DE" became Germany by inference, ACC-03043 now reads 80 which no system in the company holds, and the file that proved it used to say 8 has been overwritten.
"Clean up the CRM."
You have said it and heard it. It is not a task, it is a mood with a verb attached. Hand it to
an agent with file access and you get exactly what you asked for: changes, everywhere, made
confidently, with no record of which rows were touched or why.
This lesson turns that sentence into something that can be done and checked, prices it, and
sends you back to your Module 1 list.
What you will be able to do
Rewrite an open-ended task into a scoped one with four named components.
Say what a task costs in usage terms, and when that changes the decision.
Re-score your fit log with the instrument and notice which tasks moved.
The rewrite drill
The product fact underneath this: open-ended, recursive or poorly defined tasks raise both the
likelihood and the impact of errors. Scoping is a teachable skill, fixed by writing, not by a
better model.
Four components. A brief missing any one is under-scoped, and the failure's shape tells you
which.
Component
What it is
The failure when it is missing
Definition of done
The condition that makes it finished
It keeps going, or it stops early and you cannot tell which
Out-of-scope list
What not to touch, named explicitly
It touches things nobody asked about
Stopping condition
When to stop and ask instead of continuing
It pushes through the ambiguous 8 percent and you inherit the guesses
Exceptions rule
What to do with anything unresolvable
It guesses, and the guesses are invisible
The drill, worked on the real sentence:
Before. "Clean up the accounts file and fix anything that looks wrong."
After.
Input: 10-data/accounts.csv (308 rows). Do not modify it.
Scope: the country column only.
Rule: canonical country values are full country names as written in
DATA-DICTIONARY.md. Flag any row whose country value is not one.
Output: 30-output/country-cleanup/proposed-changes.csv with columns
account_id, column, current_value, proposed_value, reason.
Out of scope: every other column. Do not merge accounts. Do not touch
health_score. Do not touch anything on a Churned account.
Definition of done: every row in accounts.csv has been examined and is
either absent from the output or present with a proposed value.
Stopping condition: if more than 5 percent of rows are unresolvable, stop
and report instead of continuing.
Exceptions: anything you cannot resolve goes to
30-output/country-cleanup/exceptions.csv with the account_id and
the reason. Do not guess.
Three things changed, and only one is about clarity.
It named one defect class. Not "anything that looks wrong": one column, one rule, with real
work to do. answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting
the lab, documents ACC-02838 as united states, ACC-02841 as DE,
ACC-02974 as netherlands and ACC-03002 as The Netherlands, all class
inconsistent-values.
It changed the blast radius. "Fix" writes to the file; "propose" writes a different file and
leaves the source alone. Same value, one class lower, a human in the middle.
It made the output checkable. A proposed-changes file with a reason column is verifiable row
by row. A cleaned file is not: the evidence of what changed was destroyed by the change.
Check yourself: rewrite one open-ended request from your own week, with all four components.
Usage is a triage input
Cowork draws on the same plan allowance as Claude.ai chat and Claude Code, under a 5-hour
rolling window plus a weekly cap. Complex multi-step Cowork tasks consume materially more
allowance than standard chat, and Pro users hit limits fastest under heavy use.
Worth knowing: this course teaches the shape, never a number. Anthropic does not publish
stable allowances, and we have not verified any figure. A specific allowance quoted anywhere,
even in a confident blog post, is folklore until you have read it on your own account page.
What the shape means for triage:
A big hygiene pass is not free; what it costs is your chat. One shared pool means a
full-file pass started at 14:00 can leave you without chat at 16:00, on a Thursday in
quarter-end week, when three people need something.
Run heavy passes when you will not need chat. A scheduling decision, not a technical one.
Split a large job into batches, by region, owner or segment. Batching also gives you a
verification point after batch one, which you wanted anyway.
"This task would consume my week" is a legitimate reason to decline or reduce scope, stated
in the delegation policy rather than unspoken.
Check yourself: for your largest shortlist task, when in the week would you run it, and why?
Revise the fit log
Go back to cowork-fit-log.md, the fit log you wrote yourself in Module 1, and re-score every row with the five-factor test,
writing the deciding factor on each.
Expect two or three rows to move each way. Up-movers are usually tasks you thought too messy;
messy usually means high Volume with a real source of truth, the best profile there is.
Down-movers are usually tasks you were enthusiastic about, where Checkability turned out to be
2. That is the point of writing the log before you had the instrument: nobody learns from a
list they scored correctly the first time.
Check yourself: produce section 1 of delegation-policy.md, the scored inventory, with a
decision per row.
Try this now: run it both ways
Both passes run in the revenue-workspace/ folder you built in Module 2, on the practice-pack
data already copied into 10-data/.
Pass 1, the request as people actually write it.
Clean up the accounts file and fix anything that looks wrong.
Pass 2, the scoped rewrite. The block from the rewrite drill, verbatim.
Two decisions behind pass 2 are worth copying. It is smaller and draft-only, not more
detailed: adding instructions to an unbounded task just makes a longer unbounded task. And its
stopping condition was set before any output existed, because one set after seeing the output is
a rationalisation. Here the country rule resolves cleanly and it never fires. Fine: a smoke
alarm that never goes off is working.
The break. Run pass 1, verbatim, on a copy of 10-data/accounts.csv. No other instructions.
Let it finish.
What to look for. Three things, none of them a syntax error.
1. Changes nobody asked for. "Anything that looks wrong" is a wide invitation, and
PLANTED-ISSUES.md documents how much the file has on offer:
seven inconsistent-values rows in country and industry, including ACC-02948 (leading
space in Retail) and ACC-03092 (Hospitality & Leisure);
two whitespace rows, ACC-03008 and ACC-03021, trailing spaces in account_name;
eight duplicate-accounts rows, for example ACC-02890 and ACC-02988, one company under a
punctuation variant;
seven missing-values rows, including ACC-02828 and ACC-03073 (blank owners);
nine SUBTLE scale-mismatch rows, health scores on a 1 to 10 scale, all owned by
CSM Tomasz Jonker: ACC-02805, ACC-02834, ACC-02938, ACC-02945, ACC-02984,
ACC-03007, ACC-03043, ACC-03086, ACC-03118.
A pass told to fix what looks wrong will engage with several of these classes. Which ones will
vary between runs, because nothing constrains it. That variability is itself the finding:
two people running the same sentence on the same file get different edits, which is why you run
this rather than read it.
2. No record of what was touched. A cleaned file does not tell you what changed. You cannot
diff against a source you overwrote, and you cannot review a decision you cannot see.
3. At least one change that is only visibly wrong from the source. Watch the health scores:
the nine scale-mismatch accounts are the trap, and both available moves are wrong. Leaving
ACC-03043 at 8 pulls the account into every "health below 40" list. Rescaling it to 80
invents a number no system holds: the file does not record which scale the CSM used. The right
move is neither: flag it, name the pattern (nine accounts, one CSM), and ask the person who
entered it. An under-scoped brief cannot "ask", so it picks.
The guard. Run pass 2 side by side, put the two outputs next to each other, and answer three
questions in writing:
Which rows did pass 2 touch, and how do you know? (Read proposed-changes.csv.)
Which rows did pass 1 touch, and how do you know? (You cannot, and that is the answer.)
What is in exceptions.csv, and what does its length tell you about your brief? An empty
exceptions file on messy data is a warning sign, not a success. Module 4 teaches how to read
it.
The guard is not "write better prompts". It is three structural changes: one named defect
class, a proposals file instead of an edit in place, and an exceptions route for anything
ambiguous.
What you keep
revenue-workspace/40-reference/delegation-policy.md, the file you have been writing since
Lesson 1, complete:
Scored inventory: every recurring task from the fit log, five factors, total, deciding
factor, and a decision (hand over now / after the source of truth is fixed / keep human).
The shortlist: top five candidates with blast-radius class, review gate, approver,
reversal plan, and minutes saved per week with verification cost already subtracted.
The never list: your exclusions, each with the Lesson 2 rule that generates it.
Plus cowork-fit-log.md, re-scored, moved rows marked so you can see your calibration
change. This is the document a manager approves, and the capstone workflow must come from
section 2.
Quick check
Here is a brief: "Go through the contacts file and fix the email addresses." Which of the
four scoping components is missing? Rewrite the one line that would do most to make the
output checkable.
A large hygiene job is proposed for Thursday afternoon in quarter-end week. State the usage
consequence and the right scope change. Do not state a numeric allowance, and say why not.
Your exceptions file after a scoped run on messy data has zero rows. Give the two most likely
explanations and say which you would check first.
Which task moved furthest between your Module 1 fit log and your scored inventory? Name the
factor that moved it. If nothing moved, you probably scored the log rather than re-scoring it.
Answers
1. All four are missing, and saying so beats picking one. The line that does most for
checkability is the output shape: change "fix" to "write proposed-changes.csv with
contact_id, column, current_value, proposed_value, reason". That turns an unreviewable
edit into a reviewable list and drops the blast radius a class. And the brief is worse than it
looks: PLANTED-ISSUES.md records 32 rows in class email-quality covering three different
problems with three different correct treatments, and a role-based address like sales@ is not
malformed and must not be "corrected" into a person.
2. One shared pool under a rolling window plus a weekly cap, so it can cost you chat access
for the rest of the window in the week you need it most. Correct change: batch by segment or
owner, first batch now, the rest out of hours. No number, because the allowances are not
published and we have not verified any figure.
3. Either the exceptions rule never made it into the run, or the brief was narrow enough
that nothing ambiguous fell inside it. Check the first: look for a documented ambiguous case
resolved silently.
4. Open, and it is a calibration check on you rather than on the task.
Where this goes next
Next module: Module 4 takes your top shortlist item and turns it into a brief you can run,
check with evidence, and keep, so next Tuesday costs a minute instead of forty.
Track labs: the exercise sheet in labs/LAB-03 of your practice pack (start with its
README). Part A ships twelve tasks with reference scoring and produces the
scoped rewrite of one deliberately open-ended task. CP-03, the module quiz in the
interactive app, is the heaviest checkpoint in the
foundations track: a wrong answer here is the one that costs money.
Module 3 · Triage, what to hand over and what it costs if it goes wrong
Lab / ? min / read-only (Part A). Part B classifies tasks up to writes-to-system and authorises none of them.
Lab
Twelve tasks, five factors, one policy
LAB-03: Twelve tasks, five factors, one policy
The situation
Your VP read something at the weekend and now says "use AI for everything you can". You have around
twenty recurring tasks. Handing over the right five gives you back most of a day a week. Handing over
the wrong one puts an incorrect renewal amount in front of a customer, merges two real accounts that
happened to share a name, or writes a stage change into the CRM that nobody can unwind.
The expensive mistake is not the task that visibly fails. It is the task that produces a plausible answer
that is wrong in a way nobody notices for six weeks.
This lab gives you the instrument, makes you use it on twelve real-shaped tasks, and then makes you
write the policy your manager approves.
What you will produce
File
What it is
40-reference/task-scoring.md (Part A)
All twelve tasks scored on five factors, each with a decision and a named deciding factor
40-reference/scoped-rewrite.md (Part A)
One deliberately open-ended task rewritten with a definition of done, an out-of-scope list, a stopping condition and an exceptions rule
40-reference/delegation-policy.md (Part B)
Your own scored inventory, your shortlist of five with blast radius and reversal plans, and your never list with the rule that generates each entry
Outcome accounting
Manual baseline: Not time. Decision quality.
This is the one lab in the course whose baseline is not minutes, and the lab
says so rather than inventing a number.
What is measured: Two counts, taken against the reference scoring in solution/.
1. How many of your twelve decisions match the reference decision.
2. How many of your matching decisions also match the reference DECIDING
FACTOR.
The second number is the real one. Matching the answer with the wrong reason
is scored as a miss, because the reason is what transfers to task thirteen.
Reference run: The lab author scored the twelve in 22 minutes and named the deciding factor
wrongly on 2 of them (tasks 4 and 9, the two where reasonable operators
disagree). Observed, not a target.
Your actual: _____ of 12 decisions match _____ of those also match on the reason
_____ minutes
Verification cost: Zero for Part A: the reference scoring is the check, and it is released with
solution/ after you submit. For Part B there is no reference, so the check is
the colleague test in "How you verify" below, about 10 minutes.
Blast radius
Part A is read-only. You are scoring tasks on paper, not running them. You will point Cowork at the
dataset once, in step 5, to run a deliberately under-scoped brief and see what it does, and even that run
writes only into 30-output/.
Part B classifies tasks all the way up to writes-to-system and authorises none of them. Writing
"writes-to-system, approver: the RevOps lead, reversal: restore from the pre-run export" in a policy
document is not the same as doing it. The promotion exercise is LAB-05; the live promotion is
TL-OPS-2.
Declared scorecard scope
None. No findings list. See SCORING.md §1.
How you verify your own output
Part A. Score all twelve before you open solution/. Then compare on both axes: decision and
deciding factor. Two of the twelve are deliberately contested and the reference says so; disagreeing
with those two is not a miss if your reasoning names the factor the disagreement turns on.
Part B, three checks, about ten minutes:
The never list is derived, not recited. Every entry names one of the five exclusion rules from
MODULES/03 beat 4 that generates it. An entry with no rule is a preference.
Every shortlist entry has a reversal plan proportionate to its class. Read-only: delete the
output. Draft-only: do not send. Writes-to-system: written, tested, and in the run record before
the first write. A writes-to-system entry with an approver and no reversal plan fails Safety at 1,
and MODULES/03 says so explicitly.
The colleague test. Hand delegation-policy.md to somebody who does your job and ask them to
name the one task on your shortlist they would remove. If they cannot form an opinion, the file does
not contain enough reasoning to be approved by anyone.
Files you need
course/labs/data/ for step 5, plus 40-reference/cowork-fit-log.md from LAB-01 and the workspace from
LAB-02.
What comes next
LAB-04 takes the top item from your shortlist and turns it into a brief you can rerun. The capstone
workflow must come from this shortlist, which is how we stop people building something they will never
use.
Module 3 · Triage, what to hand over and what it costs if it goes wrong
6 questions / pass at 70%
Checkpoint
Module 3 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
Five candidate tasks from one RevOps analyst's week.
Dedupe the accounts file monthly and produce a proposed-merge list. About 300 rows, every month.
Rank this quarter's open opportunities by how likely the rep is to be sandbagging.
Produce a field completeness report by owner across accounts, contacts and opportunities. Monthly.
Summarise last week's call notes into one structured record per opportunity. About 16 files, weekly.
Rewrite the stage definitions document so the wording is consistent. Once.
Which is the worst candidate, and what is the deciding factor?
Hint
Score all five on Checkability first and stop there if one of them scores low. The rule is that a low
Checkability score is a veto regardless of everything else, so the other four factors are only worth
computing for the tasks that survive it.
A. Correct. You picked the right task and, more importantly, the right reason. Sandbagging is a belief about a person's intent. No file contains it, so there is nothing to sample against and nothing to reconcile, which makes the Check letter impossible. The veto rule exists because tasks like this one score well on Volume and on Specification cost and therefore look attractive on a total.
B. Incorrect. Task 5 is a genuinely poor candidate and Volume is a real reason, but a one-off is a weak candidate rather than a dangerous one. If it goes wrong you notice immediately, because a definitions document is a page you read. The item asks for the worst, and the worst is the one that can be wrong for six weeks without anybody noticing.
C. Incorrect. You picked the right task with the wrong factor, which is scored as a miss for a reason: the factor is what you carry to the next decision. Cost of being wrong is high here, and if it were the only problem you would fix it with a review gate and a named approver. Checkability cannot be fixed with a gate. It is the veto because there is no procedure that would make the output verifiable.
D. Incorrect. Messy inputs are not the same as unspecifiable work. The output shape for task 4 is easy to write down (one record per opportunity, these fields, quote the line you took each fact from), and the result is checkable because every extracted fact traces back to a named file. Messiness raises the effort and does not remove the source of truth.
Question 2 of 6
A CS lead describes a workflow as read-only: "It reads the usage and ticket exports and writes an
adoption summary into my output folder. It changes nothing."
Then they add: "and I paste the adoption section straight into the customer's QBR deck."
Reclassify it and state what the reclassification obliges.
Hint
Ignore what the agent did and follow the output. Write down the last place it ends up and who reads it
there. That destination, not the file operation, decides the class.
A. Incorrect. You classified correctly and then invented the wrong obligation. A confidence score produced by the same run that produced the claim is not evidence, it is more output from the same source. What makes a draft reviewable is that every claim carries the record it came from, so a human can check it without redoing the work.
B. Incorrect. This is the most common misclassification in the course and it is the reason the item exists. Blast radius is a property of the workflow, and the paste into the customer deck is inside the workflow the CS lead just described. Drawing the boundary at the agent's last file operation means the riskiest step in the process is the one nobody classified.
C. Incorrect. Over-classification looks cautious and behaves badly. Writes-to-system is for changing a system of record, sending to a customer without a human in the loop, or touching money, and it carries a named approver per run and a tested reversal plan. Applying that machinery to a QBR slide a human is going to read anyway produces a process people quietly stop following, which leaves you with no gate at all.
D. Correct. The destination sets the class. Once the output reaches a customer, a wrong adoption number is a wrong number in front of the person paying you, and the control that matches that is a full human read by the sender before it leaves the folder. The reversal plan for draft-only is genuinely "do not send", which is why the review has to happen before, not after.
Question 3 of 6
Four tasks. One is excluded by a rule, and the item is asking for the rule rather than the task.
Draft the renewal email for a customer, for the AM to review and send.
Produce the quarterly evidence pack that your finance team submits to the external auditor, showing
which revenue was recognised in which period and how each figure was derived.
Build the campaign suppression list from opted-out contacts and existing customers.
Summarise the week's support tickets by category and severity.
Which is excluded, and by which rule?
Hint
Three of these four have a control that makes them safe enough to hand over. One has a requirement that
no control inside your own folder can satisfy, because the party asking the question is outside your
organisation.
A. Correct. The distinguishing feature of task 2 is not that Cowork would do it badly. It might do it well. The problem is that you cannot evidence what it did to a party who is entitled to ask, because the enterprise compliance tooling does not cover Cowork yet (FACTS §5.10, §7.7). Enterprise telemetry exists and is useful, and it is not the same thing as a compliance audit log. That is a governance exclusion, not a capability one, and it is the distinction the item is testing.
B. Incorrect. Task 3 is genuinely high stakes and the failure you named is real, and it is controlled rather than excluded. A suppression list is checkable against source, the exclusion counts reconcile, and the output is draft-only with the send gated on a human. High cost of being wrong plus high checkability is a workflow with a gate, not a workflow you refuse.
C. Incorrect. Right task, wrong rule, which is scored as a miss because the rule is what you carry to the next case. The evidence pack does not change a system of record. It is a document produced from records. The rule that excludes it is about your ability to evidence the process to an outside party.
D. Incorrect. Task 1 is the textbook draft-only case. The rule you cited is about anything reaching a customer *without a human reading it*, and the task explicitly ends with the AM reviewing and sending. Dropping the second half of a rule turns a workable gate into a blanket refusal.
Question 4 of 6
Here is a brief a colleague ran last week.
"Using 10-data/accounts-2026-08-01.csv, find accounts where the country value is not one of the
canonical full country names in 40-reference/field-policy.md. Write one row per affected account to
30-output/hygiene/country-fixes.csv with columns account_id, current_value, proposed_value,
rule. Do not modify the source file. Do not touch any other column. You are finished when every
non-canonical value in the file appears once in the output."
Which of the four scoping components is missing?
Hint
Read the brief once for each of the four components and tick them off. Three are present in some form.
For the fourth, ask what happens on the row where the correct answer is genuinely unclear.
A. Incorrect. "Do not modify the source file. Do not touch any other column." is an out-of-scope list. It is short, and short is fine: an out-of-scope list has to name the things a reasonable reading might otherwise include, not everything in the world.
B. Incorrect. The last sentence is a definition of done, and a usable one, because it states an observable condition rather than a feeling of completion. Re-read the brief looking for the condition rather than the phrase.
C. Correct. The brief has a definition of done ("finished when every non-canonical value appears once"), an out-of-scope list ("do not modify the source file, do not touch any other column") and an implicit stopping condition inside the definition of done. What it has no answer for is the row where the value is ambiguous. Without an exceptions rule the agent produces a proposed value anyway, and a guess written into a `proposed_value` column is indistinguishable from a derivation. This is the highest-yield instruction in the course for exactly this reason.
D. Incorrect. They are not the same. A definition of done says what finished looks like. A stopping condition says when to stop and ask instead of continuing, which is a different event, usually sized as a threshold ("stop and report if more than 5% of rows are unresolvable"). Collapsing the two is how a brief ends up with no instruction for the case where the work turns out to be bigger or messier than anybody expected. The brief is also not complete, for the reason in option C.
Question 5 of 6
It is Thursday afternoon in the last week of the quarter. Your VP wants a full hygiene pass across
accounts, contacts, opportunities and leads before Monday. You are also the person three other teams
message for a number between now and Friday evening.
What do you do?
Hint
One option is right about the shape of the answer and wrong about a specific number. Ask yourself where
that number would have come from.
A. Incorrect. "It is one task" is a description of how you would like to think about it, not of how it consumes. Four objects in one pass on the day your chat access matters most is the specific combination the limit was designed to make you think about. If the run exhausts your window you have both a half-finished hygiene pass and no way to answer the three teams waiting on you.
B. Correct. Usage is a triage input rather than a surprise. One shared pool across chat, Claude Code and Cowork means a heavy pass can cost you the ad hoc answers you are also on the hook for, and complex multi-step work is the kind that consumes materially more (FACTS §4, §6). The right move is a scope change with a stated delivery split, not a refusal and not a heroic single run, and telling the VP which half arrives when is the part that keeps you employed.
C. Incorrect. Nothing here says the workflow is new, and even if it were, "not now" without a counter-proposal loses the work and the credibility. Cost is a reason to reduce scope or move timing. It is rarely a reason to decline outright, and the difference between those two answers is what a manager is actually assessing when they ask.
D. Incorrect. The judgment is defensible and the number is not. Anthropic does not publish stable numeric allowances for these tiers, so no percentage of a weekly allowance can be stated, and any answer that prints one is telling your VP something you cannot support (FACTS §6). Teach and use the shape of the limit: one shared pool, a rolling window plus a weekly cap, heavy multi-step work costs more. Then link out for anything numeric.
Question 6 of 6
One entry from a learner's delegation-policy.md shortlist.
### 2. Monthly owner backfill
Blast radius: writes-to-system
Review gate: sample 20 rows before applying
Approver: Priya, RevOps manager
Minutes saved: about 90 per month
Score this entry and say what is missing.
Hint
Write out the three obligations that attach to the writes-to-system class, then tick them off against
the entry. Two are there. The third is the one you would need after a bad run rather than before it.
A. Incorrect. Every element you listed is present, which is what makes this the tempting answer. Count the obligations rather than the fields: approval, review gate, and reversal plan. The third is absent, and it is the one that matters on the day something goes wrong.
B. Correct. Writes-to-system carries three obligations and this entry has two of them. The missing one is the only one that operates after the mistake has already happened, which is why it cannot be traded against the other two. "Written, tested, and in the run record before the first write" is also three requirements in one line: written is not enough, because an untested restore procedure is a belief, and a tested one you cannot find at 6pm on a Friday is not a plan.
C. Incorrect. An estimate is what a shortlist entry is supposed to contain at this stage, because the point of the shortlist is to choose what to build before you have built it. The measured number arrives later, from the lab's outcome accounting, with the verification cost inside it. Grading an estimate as an error here would push learners to invent measurements they have not taken.
D. Incorrect. You correctly identified Safety and then named the wrong defect. A sample of 20 with a stated selection rule is a reasonable gate. A full review of every proposed change is the answer that fails the course's own rule that the check must cost less than the work, and a gate nobody can afford to run is a gate that gets skipped. The missing element is the reversal plan.
M4: The operator's working method, assign, check, keep
Last verified: 2026-08-02.Product claims: FACTS §5.1, §5.2, §5.7, §5.8, §5.10, §7.1, §7.5, §7.7.
Dataset: Talvera, seed 20260801. The account used throughout is ACC-02848, Oldcastle
Leisure Inc, chosen because its opportunity history contains a documented duplicate
(PLANTED-ISSUES.md, class duplicate-opportunity-rows, OPP-007099) and therefore gives the
reconciliation something real to catch. Every identifier here was read out of the shipped files.
The module at a glance
You handed over a task and got something back. It looks right. Three questions decide whether this
was worth doing:
Would you sign your name under it, and on what evidence?
Can you get the same thing next Tuesday without retyping the request?
If a VP asks where a number came from, can you show them?
Most people who try an agent for revenue work answer no to all three, conclude the tool is
unreliable, and go back to doing it by hand. The gap is not the model. The gap is that they wrote a
request instead of a brief, checked by reading rather than by evidence, and kept nothing.
This is the craft module. Modules 5 to 8 are applications of what is taught here, and a learner who
skips it produces output they cannot verify or reuse.
Three TRACK letters live here: Assign (what you say), Check (what proves it), Keep (what
survives).
Lessons
#
ID
Title
Time
Reading
Doing
Carries
L1
L4.1
A request is not a brief
17 min
12
5
Worked exercise
L2
L4.2
The folder changes the plan
13 min
10
3
Failure lane
L3
L4.3
Evidence cheaper than the work
16 min
13
3
Discrimination pair
L4
L4.4
Steering, and what survives the session
11 min
9
2
Lesson total: 57 minutes against the M4 budget of 50 in PEDAGOGY.md §3.1, inside the plus or
minus 15 percent tolerance (§6.9). Reading minutes are measured from the shipped text at 200 words
per minute for prose, faster for tables and code. No lesson exceeds the 20-minute hard cap. The
module also carries LAB-04 (85 min) and CP-04 (10 min).
Every ## Check yourself section carries a collapsed answer block, with per-option lines on the
three closed items in this module: L4.2 item 1 (the global / folder / brief sort), L4.3 item 1 (the
three verification procedures, where the wrong-but-tempting option is the one that compares two
outputs against each other rather than against source), and L4.4 item 3 (which deletions break the
workflow, sorted into Assign and Keep). Those blocks are counted inside the reading figure above
even though they are collapsed and consulted one item at a time.
Outcomes
ID
You can
O4.1
Write a seven-part brief and get a usable first result from it
O4.2
Move a brief into a durable file, and state which rules belong global, folder or brief
O4.3
Steer a running task, and say when restarting with a better brief beats correcting in place
O4.4
Design a verification procedure cheaper than the work, with sample size, source of truth and a reconciliation
O4.5
Instruct the agent to write unresolved items to an exceptions file, and read that file as a signal about the brief
O4.6
Produce a run record sufficient to reproduce and audit the run
What you keep
The brief pack, four files inside revenue-workspace/. Templates for all four are in
templates/.
20-briefs/account-snapshot.brief.md: the seven-part brief for one real workflow from your
Module 3 shortlist. Runnable as-is by pointing at the folder and naming the file.
40-reference/verification-checklist.md: the reconciliation, the sample size and the
known-answer procedure for that workflow, written before the first run.
50-runs/RUN-LOG.md: at least three real runs recorded.
revenue-workspace/CLAUDE.md (updated): output conventions and the standing exceptions rule.
Every application module from here reproduces this pack for a different body of work. It is also
three of the seven capstone rubric rows.
The five shared instruments this module owns
Three of the course's five shared instruments (PEDAGOGY.md §7) are defined here, in the canonical
wording:
The seven-part brief (L4.1): Outcome, Inputs, Output shape, Constraints, Definition of done,
Stopping condition, Exceptions rule.
The three cheap evidences (L4.3): Reconciliation (totals in the output agree with totals in the
source), Sampling (n rows checked by hand, with n stated), Tripwires (known-answer rows). Governing
rule: if checking costs more than doing it by hand, do not hand it over.
The exceptions rule (L4.3): anything the agent cannot resolve goes to exceptions.csv with the
row identifier and the reason, and it does not guess. An empty exceptions file on messy data is a
warning sign. A long one usually means the brief was underspecified.
Where this goes next
Modules 5 to 8 each run T, R, A, C, K in order against one body of revenue work: CRM hygiene,
pipeline and forecast inputs, demand and targeting, retention and repeat work. All four assume this
pack exists and add domain specifics rather than re-teaching the method.
Module 4 · The operator's working method, assign, check, keep
M4 lesson / about 17 min
Do it
A request is not a brief
Do this nowCount rows, then count ids60s
Filter opportunities.csv to account_id ACC-02848 and count the rows.
Count the distinct opportunity_id values in the same filter.
Write both numbers down and name the id that appears twice.
You know it worked when: Four rows, three distinct ids, and OPP-007099 is the one on two rows with different amounts.
What you type
Outcome
A one-page renewal-ready snapshot of one account that a CSM can read before a
call without opening any file.
Inputs
10-data/accounts.csv, 10-data/opportunities.csv, 10-data/contacts.csv.
Account: ACC-02848.
Output shape
30-output/account-snapshot/<YYYY-MM-DD>-ACC-02848.md
Sections: Account facts. Opportunity history, oldest first. Contacts.
Reconciliation. What the data does not tell you.
Example of a correct history row:
| OPP-007468 | New Business | Closed Won | 55,000 USD | 2023-10-19 |
Constraints
Read-only. Blank means blank, say so.
Never add a Renewal amount to a New Business amount: renewal amounts restate
current ARR (DATA-DICTIONARY.md, opportunities.type).
Definition of done
The Reconciliation section states the row count, the distinct opportunity_id
count, and whether they agree.
Stopping condition
If those two counts disagree, do not choose between the conflicting rows.
Report both and continue.
Exceptions rule
Anything unresolvable goes to exceptions.csv with the row identifier and the
reason. Do not guess.
What comes back
30-output/account-snapshot/2026-08-06-ACC-02848.md written.
Account facts: Oldcastle Leisure Inc, Mid-Market, Hospitality, 302 employees,
United States, NA. Owner Ursula Ulrich. CSM Helena Zabala. ARR 55,000 USD,
billing currency USD. Health score 37. Renewal 2026-10-19. Tier 2. Customer.
Opportunity history
| OPP-007468 | New Business | Closed Won | 55,000 USD | 2023-10-19 |
| OPP-007132 | Renewal | Closed Won | 55,000 USD | 2024-10-19 |
| OPP-007099 | Renewal | Closed Won | 69,000 USD | 2025-10-19 |
| OPP-007099 | Renewal | Closed Won | 55,000 USD | (blank) |
Reconciliation: 4 rows, 3 distinct opportunity_id values. They do not agree.
No total printed: 3 of the 4 rows are type Renewal and restate current ARR.
exceptions.csv: OPP-007099, duplicate identifier, amounts 69,000 and 55,000
disagree and one row has a blank close_date. Cannot determine which is
authoritative.
Notice: The definition of done is what forced two counts onto the page, and the stopping condition is why the run reported the conflict instead of quietly keeping the row that had a close date.
In Module 1 you ran a snapshot. In Module 3 you picked a workflow worth repeating. Now you
write the thing that makes it repeatable. Your first attempt will be a sentence, because
sentences are what people type.
A brief contains the four things a sentence leaves for the agent to decide: what counts as
finished, what not to touch, when to stop, and what to do with the rows that do not fit.
Either you decide those in advance, or something else decides them at run time.
What you will be able to do
Write a seven-part brief and get a usable first result from it.
Look at a failure and name which part of the brief it came from.
Stop polishing phrasing, the smallest lever you have.
The seven parts
Same task, two ways, on the shipped data.
The request:
Give me a summary of the Oldcastle Leisure account.
The brief:
Outcome
A one-page renewal-ready snapshot of one account that a CSM can read before a
call without opening any file.
Inputs
10-data/accounts.csv fields: account_id, account_name, segment, industry,
country, region, owner, csm, contract_start_date,
arr_usd, currency, health_score, renewal_date, tier,
lifecycle_status
10-data/opportunities.csv fields: opportunity_id, opportunity_name, type, stage,
amount, currency, close_date, forecast_category
10-data/contacts.csv fields: contact_name, title, persona, last_activity_date
10-data/activities.csv fields: timestamp, activity_type, subject, owner, outcome
Account: ACC-02848.
Output shape
30-output/account-snapshot/<YYYY-MM-DD>-ACC-02848.md
Sections in this order: Account facts. Opportunity history (oldest first).
Contacts. Last five activities. Reconciliation. What the data does not tell you.
Constraints
Read-only. Do not modify anything in 10-data/.
Do not infer a value that is not in the files. Blank means blank, say so.
Every number carries the file and column it came from.
Never add a Renewal amount to a New Business amount: renewal amounts restate
current ARR (DATA-DICTIONARY.md, opportunities.type).
Definition of done
Every opportunity on this account appears exactly once in the history section,
and the Reconciliation section states the row count, the distinct
opportunity_id count, and whether they agree.
Stopping condition
If the row count and the distinct opportunity_id count disagree, do not choose
between the conflicting rows. Report both and continue with the rest.
Exceptions rule
Anything unresolvable goes to
30-output/account-snapshot/exceptions.csv with the row identifier and the
reason. Do not guess.
Longer, yes. But the length went into named files, named columns, a named output filename,
and three rules for when something does not fit. Not adjectives.
Those last three sections earn their keep here. ACC-02848 has four opportunity rows but
three opportunities: OPP-007099 appears twice with different amounts
(answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab,
class duplicate-opportunity-rows). The request gets you a clean summary with a wrong total
and no warning. The brief cannot: the definition of done forces both counts onto the page,
and the stopping condition forbids picking a row.
Which of the seven parts was your last request missing?
Prompting, sized honestly
This is the only section in this course about phrasing, and it stays one section. Most of the
discipline is choosing the right work (Module 3), judging the output (Lesson 3), and
deploying responsibly (Modules 2 and 8). A course that is mostly prompt technique has
mis-sized the problem.
Three techniques earn their place in revenue work. There is no fourth.
1. Give the shape of the output, not adjectives about it. "Comprehensive" describes a
feeling. A section list, a column list and a filename describe an artifact. Reaching for an
adjective means you have not decided what you want yet.
2. Give one worked example of a correct row.
Example of a correct history row:
| OPP-007468 | New Business | Closed Won | 55,000 USD | 2023-10-19 |
One line settles date format, money format, currency placement, column order and level of
detail at once.
3. Say what to do when the data is ambiguous. The highest-yield sentence in any brief,
and Lesson 3's whole subject. "If two rows share an identifier and disagree, report both and
write the conflict to exceptions.csv. Do not choose."
Then stop. If the output is wrong, the fix is almost never a better adjective. It is a
missing input, an unstated constraint, or a definition of done you never wrote. Open-ended
tasks fail more often, and they fail bigger.
Both versions, run
Pass 1, the request. Back in seconds. Clean prose, three headings, a total: "total
closed-won value 234,000 USD across 4 opportunities". Reads well. Both numbers are wrong, in
different ways, and the page gives you no way to tell. I did not reply "are you sure?": that
buys a confidence statement, not a comparison. The fix goes in the brief, not the chat.
Pass 2, the brief. Same folder, same files. The reconciliation section: four rows, three
distinct opportunity_id values, they do not agree. One line, and the summary becomes
something you can defend.
The stopping condition earned its place at once. OPP-007099 appears at 69,000 USD with
close date 2025-10-19 and at 55,000 USD with a blank close date. Without the stopping
condition, a helpful run picks one. With it, both rows go in the snapshot and the conflict
goes to exceptions.csv.
The reconciliation also surfaced something unasked: accounts.arr_usd here is 55,000,
matching one conflicting row and not the other. A signal, not a resolution. The row matching
ARR is the one missing its close date, so the evidence points both ways. The right output
raises the question and names who could settle it, instead of quietly picking for you.
And one number I refused to print: "total closed-won value". Three of the four rows are
type = Renewal, and renewal amounts restate current ARR rather than adding to it
(DATA-DICTIONARY.md in your pack's data folder). Summing them gives 234,000, which describes nothing. The brief now
says what may be totalled and what may not.
Your turn
You are the CSM covering Oldcastle Leisure Inc. Renewal date 2026-10-19, call this week. You
want a snapshot you can regenerate every quarter, for any account, in under a minute. Budget
six minutes. This produces the first file of your brief pack.
Everything this lesson uses came in the practice pack from the course page; if you have not
downloaded it, do that first.
Setup: grant revenue-workspace/, the folder you built in Module 2, with the pack's dataset
copied into 10-data/ (accounts.csv,
opportunities.csv, contacts.csv, activities.csv). No connectors. Read-only. Have your
Module 2 CLAUDE.md in place.
Try this now
Open templates/M4/account-snapshot.brief.md from your practice pack and fill the four
gaps (starter below).
Save it as 20-briefs/account-snapshot.brief.md: this is a file you create, and it lives
in your workspace from here on.
In a session, grant the folder and say only: Run the brief in 20-briefs/account-snapshot.brief.md for account ACC-02848. Nothing else. If you have to
add anything, the brief was incomplete, and that is the finding.
Open the output. Read the Reconciliation section first.
Open exceptions.csv. If it is empty on this account, your Exceptions rule or Stopping
condition is not doing its job. Fix the brief, not the output.
The starter brief, with gaps marked:
# Brief: account snapshot
## Outcome
A one-page renewal-ready snapshot of one account that a CSM can read before a call
without opening any file.
## Inputs
10-data/accounts.csv, 10-data/opportunities.csv, 10-data/contacts.csv,
10-data/activities.csv.
<<GAP 1: name the exact columns you need from each file. Not "the relevant
fields". Column names.>>
Account: supplied at run time as an account_id.
## Output shape
30-output/account-snapshot/<YYYY-MM-DD>-<account_id>.md
Sections: Account facts. Opportunity history, oldest first. Contacts. Last five
activities. Reconciliation. What the data does not tell you.
<<GAP 2: one worked example row for the opportunity history section, showing
date format, money format and column order.>>
## Constraints
Read-only. Do not modify anything in 10-data/. Do not infer a value that is not in
the files; blank means blank.
<<GAP 3: the constraint that stops a wrong total on this data. Look at
DATA-DICTIONARY.md, opportunities.type, and at the currency convention.>>
## Definition of done
<<GAP 4a: the condition that makes the Reconciliation section non-optional. It
should force two counts to be printed and compared.>>
## Stopping condition
<<GAP 4b: what to do when those two counts disagree.>>
## Exceptions rule
Anything unresolvable goes to 30-output/account-snapshot/exceptions.csv with the
row identifier and the reason. Do not guess.
Stuck? Three hints, in rising order of help.
Three of the four gaps make the output nicer. One makes it possible to tell that the
output is wrong. A snapshot that reads beautifully and cannot be checked is the failure
this module exists for.
Before writing GAP 4, count by hand for ACC-02848: the rows in opportunities.csv for
that account, and the distinct values of opportunity_id. Then write the instruction
that makes the agent print those two numbers every time, for any account, unasked.
GAP 4a should name two counts and require them printed and compared, not just computed.
GAP 4b should forbid a choice: report both rows, route the conflict to exceptions, never
pick the row that looks more complete. GAP 3 needs two sentences, one about
type = Renewal amounts and one about currency, because opportunities.amount is in the
account's local billing currency while accounts.arr_usd is in USD.
What you should see on ACC-02848 (the filled brief ships as
templates/M4/account-snapshot.brief.md):
Account facts: Mid-Market, Hospitality, 302 employees, United States, NA, owner Ursula
Ulrich, CSM Helena Zabala, contract start 2023-10-19, ARR 55,000 USD, currency USD, health
score 37, renewal date 2026-10-19, Tier 2, Customer.
Opportunity history: OPP-007468 New Business Closed Won 55,000 USD 2023-10-19;
OPP-007132 Renewal Closed Won 55,000 USD 2024-10-19; OPP-007099 Renewal Closed Won
appearing twice, at 69,000 USD with close date 2025-10-19 and at 55,000 USD with a blank
close date.
Reconciliation: 4 rows, 3 distinct opportunity ids, they do not agree.
exceptions.csv: at least one row, OPP-007099, reason: duplicate identifier with
conflicting amount and one blank close date, cannot determine which row is authoritative.
What the data does not tell you: which of the two 2025 renewal amounts is correct; that no
renewal opportunity exists for the 2026-10-19 renewal date; that health_score 37 has no
stated definition in the data.
Your run may differ in section order, table versus prose, activity count, exception wording,
and whether you note the account's declining seat utilisation. It may not differ here: a
missing Reconciliation section, or a run that silently picked one OPP-007099 row, means
the brief failed, however good the page reads. And GAP 3 must contain both rules from
DATA-DICTIONARY.md: renewal amounts restate current ARR and are never added to New
Business amounts, and never total across currencies without converting.
Worth knowing: the parts of a brief that feel like bureaucracy, the definition of done
and the stopping condition, are the only two parts that can tell you the output is wrong.
What you keep
revenue-workspace/20-briefs/account-snapshot.brief.md, the first file of the brief
pack, runnable by naming the file, on any account id. Also keep the pass-1 output next to
the pass-2 output. You will need both in Lesson 2.
Check yourself
A brief produced a snapshot that stopped after twelve of forty accounts and reported
success. Which of the seven parts was missing? Name a second part that could also have
caused it, and how you would tell them apart from the output.
Your output has the right numbers and the wrong section order. Which part do you edit, and
how much time should that take?
Someone proposes fixing a recurring wrong total by adding "please be accurate and
double-check your work" to the brief. What will that change, what will it not, and which
part should they edit instead?
Answers
1. Definition of done: nothing stated the condition that makes it finished, so twelve
accounts satisfied it as well as forty. The second candidate is the stopping condition:
it may have hit something unresolvable and stopped correctly. The output tells them apart.
Stopped with a named row and reason: the stopping condition fired, the system worked.
Reported success with no reason: the definition of done was missing.
2. Output shape, in under a minute. Section order is the cheapest thing in a brief to
change and the least important to get right first.
3. It changes the tone, not the total. Nothing in that sentence says which rows may be
added to which, so the same wrong total comes back sounding more careful. Edit
Constraints, naming the rule that makes the total wrong, and Definition of done,
requiring the reconciliation that would have exposed it.
Where this goes next
Next lesson: the same brief, run with and without the folder context file.
Track labs: the exercise sheet in labs/LAB-04 of your practice pack (start with its
README). Part A runs three passes on this task and requires the account
chosen to contain one of the four documented duplicate opportunity ids, so the reconciliation
has something real to catch.
Module 4 · The operator's working method, assign, check, keep
M4 lesson / about 13 min
Failure mode
The folder changes the plan
Do this nowAsk for it the lazy way60s
Open a fresh session with revenue-workspace/ granted.
Type only: Same account snapshot as before, for ACC-02848.
Do not name the brief file, and do not correct it while it runs.
You know it worked when: You get a snapshot, and it differs from your last one in at least one way you did not choose.
What you type
Same account snapshot as before, for ACC-02848.
What comes back
Account snapshot: Oldcastle Leisure Inc (ACC-02848)
Mid-Market, Hospitality, United States. Owner Ursula Ulrich.
ARR 55,000. Renewal 2026-10-19.
Opportunities (3)
OPP-007468 New Business Closed Won 55000 2023-10-19
OPP-007132 Renewal Closed Won 55000 2024-10-19
OPP-007099 Renewal Closed Won 69000 2025-10-19
Total closed won: 179,000
Contacts: Ulrich Steenkamp, Fabian Nyberg, Tobias Lombardi.
Notice: Three opportunities is the right count reached the wrong way: the second OPP-007099 row was dropped rather than reported, and 179,000 adds two renewal amounts to a new business amount.
You now have a brief, the one you wrote yourself in Lesson 1. It sits in 20-briefs/, and it
works.
Then a colleague copies it into their own workspace, or you open a session with the folder
not quite set up. The output comes back close, not the same, and it differs in ways you would
not expect.
This lesson is about where instructions live and what happens when a layer is missing. It is
the last time this will surprise you.
What you will be able to do
Sort your rules between global instructions, the folder file, and the brief.
Recognise the failure shape of a missing context layer.
Say honestly what this course can and cannot tell you about packaging a brief for others.
The folder changes the plan, not just the wording
Per-folder context files carry the project rules. At session start Claude reads MEMORY.md
for cross-project learnings and, inside a project folder, that folder's own CLAUDE.md and
MEMORY.md.
The part people underestimate: the folder file does not decorate the output, it changes the
plan. If CLAUDE.md says renewal amounts must never be added to new business amounts, the
plan for "total this account" is a different plan, not the same plan with a nicer label.
Remove the file and you get different work, not the same work presented differently.
This is why Module 2 came before Module 4: rails are built once and inherited, and this is
where you find out what you inherited. Try naming one rule in your own CLAUDE.md whose
removal would change a number, not a format.
Three places a rule can live
Three layers now, not two. Sorting a rule into the wrong one is the most common configuration
mistake in this course, and each mistake has its own smell.
Layer
What belongs there
The smell when a rule is in the wrong layer
Global instructions
Who you are, how you write. True in every session you ever run.
A global rule that refers to a folder is wrong most of the time, so you quietly delete it, and then it is missing from the sessions where it mattered.
Folder CLAUDE.md
What this body of work is. Facts about this data, output conventions, standing safety rails, the exceptions rule.
A folder rule that only applies to one workflow makes every other workflow in the folder behave oddly.
Brief file
What this run does. Outcome, inputs, output shape, definition of done, stopping condition.
A brief that restates the folder rules is a brief nobody maintains, and the copies drift.
The test, applied in order:
Would this be true in a session about my personal calendar? Then it is global.
Would this be true for every task in this folder? Then it is folder.
Otherwise it is brief.
Worked on real rules from earlier modules:
"Write plainly, no marketing adjectives": global.
"opportunities.amount is in local billing currency; accounts.arr_usd is USD":
folder. A fact about this data, true for every task that touches it.
"Sections in this order: account facts, opportunity history, contacts, activities,
reconciliation": brief. Only this workflow has those sections.
"Anything unresolvable goes to exceptions.csv and do not guess": folder, wanted on
every task. Lesson 3 shows why it is the highest-yield line in the course.
Now sort your own: which rules in your brief should move down to the folder?
Where reusable work lives, and what we will not teach you
The brief you wrote in Lesson 1 is worth more than the output it produced. The output answers
one question about one account. The brief answers it every quarter, for any account, for
anyone who has the file. So: 20-briefs/<workflow>.brief.md, in version control if you
have any, named after the workflow, not the day.
Now the honest part. Cowork supports skills, domain expertise that activates
automatically, and plugins, which package skills, MCP connectors, subagents, slash
commands and hooks. So packaging a mature brief as an installable skill or plugin is
possible, and it is the natural next step when a team wants to use your workflow rather than
copy your file.
But this course's verified facts do not include the authoring format for a skill file. So we
teach the file route, which we can teach accurately, and name the packaging route as one to
explore with current documentation, without inventing a format.
That models the standard to hold on your own outputs: say what you know, name what you do
not, never fill the gap with something plausible. If you take one habit into work where
nobody is grading you, take that one.
The experiment: one layer removed
The same brief, run twice. Worth doing, not just reading.
Run A: everything in place.revenue-workspace/CLAUDE.md (the folder file you wrote in
Module 2) present, brief file present.
The request: Run the brief in 20-briefs/account-snapshot.brief.md for account ACC-02848.
Output: the six sections in order, money as 55,000 USD, dates as 2025-10-19, a
Reconciliation section reading four rows and three distinct opportunity ids, one row in
exceptions.csv, and a first line naming the source files.
Run B: CLAUDE.md moved out of the folder. Same brief file, same request, same data. I
changed nothing in the brief: otherwise I would not know which layer made the difference.
The brief still carried the output shape, so the sections were right. Missing was everything
the brief relied on the folder for: the source-files line, the output location convention,
the standing exceptions rule. The reconciliation survived, because it lives in the brief's
definition of done. The exceptions file was thinner, because "do not guess" lived in the
folder file.
The output did not fall over. It got quietly less defensible, and nothing announced it.
There is no error message for "your configuration was less complete than usual".
Try this now
Open a fresh session. Temporarily move revenue-workspace/CLAUDE.md out of the folder.
Rename it, do not delete it.
Ask the way a busy person asks on a Tuesday, without naming the brief file. That is what
gets typed when you are in a hurry and believe the setup remembers you.
Same account snapshot as before, for ACC-02848.
Restore CLAUDE.md, open another fresh session, and run
Run the brief in 20-briefs/account-snapshot.brief.md for account ACC-02848.
Compare the two outputs side by side.
What you should see: two failures stacked, and separating them is the exercise.
Failure one: "as before" refers to nothing. There is no memory between sessions, so "before"
is not something the session can resolve. What you get depends on what is lying around: if a
previous snapshot sits in 30-output/, the run may find it and imitate its shape. That looks
like memory and is not.
Failure two: the missing layer is invisible in the output. Different section order, a
different sort, money without currency codes, no exceptions file. Each difference is small.
Together they produce a document close enough that a busy person pastes it into a deck.
The thing to look for on this account: whether the run silently picked one of the two
OPP-007099 rows. The stopping condition that forbids picking lives in the brief, so naming
the brief keeps it. Saying "as before" never loads it.
Worth knowing: the guard is two files doing different jobs. The brief file, named in the
request, carries what this run does. The folder context file, read at session start, carries
what this folder always does. Plus one habit: never ask for "the same as before". Name the
brief. Eight words, and the difference between a reproducible run and a run that inherited
whatever was on disk.
Write down, in one sentence, which difference between your two outputs would have been
costliest unnoticed. That sentence goes in your run log in Lesson 4.
What you keep
revenue-workspace/CLAUDE.md, updated with any rule from your brief that belonged one
layer down.
20-briefs/account-snapshot.brief.md, unchanged, plus a note at the bottom recording
which folder rules it depends on: what the next person must bring when they copy the file.
Both outputs from the experiment, kept side by side until Lesson 4.
Check yourself
Sort these into global, folder or brief: (a) "Our fiscal year equals the calendar year";
(b) "Output goes to 30-output/<workflow>/"; (c) "Sort the opportunity history oldest
first"; (d) "Never write to a file in 10-data/"; (e) "I prefer short sentences".
A colleague copies your brief file into their own workspace and gets differently shaped
output. Name the two most likely causes, which one you would check first, and how.
Someone says they have packaged their brief as a skill so their team can install it. What
one question would you ask, and what would you not assert about the format?
Your run produced a snapshot that looks like last quarter's, and you did not name the
brief file. Give two explanations consistent with there being no memory between sessions,
and what you would look at to tell them apart.
Answers
1, one line per option:
(a) Fiscal year: folder. A fact about this body of work, wrong or irrelevant anywhere
else.
(b) Output path: folder. It names a path that only exists inside this workspace, which
is exactly what makes it a bad global rule.
(c) Sort oldest first: brief. Only this workflow has an opportunity history to sort.
(d) Never write to 10-data/: folder. A standing safety rail for every task in the
workspace.
(e) Short sentences: global. How you write, true in every session.
2. Either they do not have your folder CLAUDE.md, or theirs differs. Check the first,
because it is one look: if their output lacks the Source files: line and the 30-output/
path, the folder layer is absent rather than different. This is why your brief carries a note
listing the folder rules it depends on.
3. Ask which current documentation they used for the format, and assert nothing about it
yourself. This course documents that skills and plugins exist and what they package, not how
a skill file is written.
4. Either the run read last quarter's output from the granted folder and imitated it, or
it produced a generic default that resembles it. Neither is memory. Move last quarter's
output out of the granted tree and re-run: if the resemblance disappears, it was reading the
file.
Where this goes next
Next lesson: evidence that proves an output right at a cost below the work, and the one
instruction that turns unresolvable rows from invisible guesses into a list you can read.
Track labs: the exercise sheet in labs/LAB-04 of your practice pack (start with its
README). Part A runs this experiment as a required step and asks you to diff
the two outputs and name the costliest difference.
Module 4 · The operator's working method, assign, check, keep
M4 lesson / about 16 min
Concept
Evidence cheaper than the work
Do this nowReconcile one account's ARR60s
In opportunities.csv, take the Closed Won rows on ACC-02848 whose type is New Business or Expansion, and add their amounts.
Read arr_usd on ACC-02848 in accounts.csv.
Say whether they agree, and name the rule that kept the Renewal rows out of your sum.
You know it worked when: One New Business row at 55,000 USD against an arr_usd of 55,000, agreeing exactly, and you can say why the three Renewal rows were excluded.
What you type
Verify 30-output/account-snapshot/2026-08-06-ACC-02848.md against source, using
40-reference/verification-checklist.md. Report only. Fix nothing.
1. Row count in opportunities.csv for this account against the distinct
opportunity_id count in the same filter. State both.
2. Sum of Closed Won New Business and Expansion amounts against
accounts.arr_usd. Exclude Renewal amounts: they restate ARR
(DATA-DICTIONARY.md).
3. Tripwire: OPP-007099 is a known duplicate id. Confirm it is counted once and
that its conflict appears in exceptions.csv.
State pass or fail on each, and name the file and column behind every number.
What comes back
1. Counts: FAIL, by design
opportunities.csv rows where account_id = ACC-02848: 4
distinct opportunity_id in the same filter: 3
They disagree. Reported, not resolved.
2. Amounts: PASS
Closed Won New Business and Expansion, opportunities.amount: 55,000 USD
(OPP-007468 only)
accounts.arr_usd for ACC-02848: 55,000
Renewal rows excluded: OPP-007132, and both rows of OPP-007099.
3. Tripwire: PASS
OPP-007099 appears once in the history section.
exceptions.csv holds one row for it, with both amounts named.
Handed back to you: arr_usd agrees with the 55,000 row, and the 55,000 row is
the one missing its close_date. The evidence points both ways.
Notice: The last block is the check doing its job. A verification that ends in a question for Ursula Ulrich has produced more than one that ends in confidence.
On account ACC-02848 the roll-up is internally perfect at 4 rows, 234,000 USD and 58,500 average, and the source says 3 distinct opportunity ids and 55,000 USD of ARR.Reconciliation looks at the whole population and catches drops and double counts, sampling looks at a few rows and catches rows that are present and wrong, and a tripwire checks a row whose answer you already knew.A row the brief cannot resolve has three destinations, and the middle one is credited as a half catch because declaring what could not be resolved is the only route that leaves a visible trace.
Four hundred rows came back. You are not going to read four hundred rows, and if you were,
you would have done the job by hand.
So you scroll, spot two rows that look right, and decide the whole thing is fine. That is not
checking. It is reassurance with a keyboard, and it fails exactly where it matters: the
plausible-wrong output is the one that survives scrolling.
Verification is a design problem with a budget.
What you will be able to do
Design a verification with a named source of truth, a stated sample size and at least one
reconciliation.
Decide when verification cost alone should stop you handing a task over.
Read an exceptions file as a signal about your brief, not about the data.
The governing rule
If checking costs more than doing it by hand, do not hand it over.
That rule excludes real workflows. Some tasks are faster to do than to verify, and the
correct decision is to keep them. Saying so keeps every time-saved claim in this course
honest.
It also means verification cost sits inside any time-saved number: a workflow that saves 40
minutes and adds 35 of checking saves 5, and you write 5.
And the procedure is written before the run. A check designed after you have seen the output
is justification, not verification: you know what the answer says, and you will design a
check it passes.
Name one workflow you were about to hand over that fails on verification cost.
Three cheap kinds of evidence
Three instruments. Use at least two. They catch different things.
Reconciliation. Totals in the output must agree with totals in the source: row counts,
sums of amounts, distinct counts of accounts or opportunities. This catches silent drops, the
most common failure and the hardest to see, because a list missing 40 of 400 rows looks
exactly like a list. Cost: seconds. If you do only one thing, do this one.
Sampling. Pull n rows and check them by hand against the source. State n before the run.
A rule of thumb, not a standard:
at least 10 rows, always;
at least 5 percent of rows for anything draft-only;
every row in the top decile by amount for anything that touches a number a VP will read,
because the distribution of consequences is not the distribution of rows.
Sampling catches what reconciliation cannot: rows that are present and wrong. A total can be
right while individual rows are misclassified.
Tripwires: known-answer rows. Rows whose correct treatment you already know, planted in
the input, then checked in the output. On your own data you plant them yourself; that is the
habit.
On the shipped Talvera data you do not need to plant any: documented rows with a known
correct treatment already exist. The four opportunity ids that appear twice with different
amounts are OPP-006790, OPP-006886, OPP-007099 and OPP-007110
(answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab,
class duplicate-opportunity-rows). Each must be counted once, and any
roll-up that counts one twice has failed a check you could have designed in advance. That is
all a tripwire is: a row where you know the answer before you look.
The exceptions file
One instruction. It is the highest-yield habit in this course.
Anything you cannot resolve goes to exceptions.csv with the row identifier and
the reason. Do not guess.
Without it, unresolvable rows do not disappear. They get resolved anyway, by something that
had to pick, and the picking is invisible: a guessed value looks exactly like a known one in
a CSV. Then read the file, because it is diagnostic:
An empty exceptions file on messy data is a warning sign, not a success. The shipped
dataset carries 320 documented defect records across 23 classes (PLANTED-ISSUES.md). A
pass that resolves everything either did not look or did not tell you.
A long exceptions file usually means the brief was underspecified, not that the data was
bad. Read five rows. If they share a shape, that shape is a rule you did not write.
"Cannot determine currency" forty times is one missing sentence in your Constraints, not
forty problems.
Exceptions are never failures, not in this course and not in your own head. A run that
flags 30 rows it could not resolve did better work than one that silently resolved all 30,
and the second looks better on the page. Rewarding the second is how organisations end up
with confident wrong numbers.
The checklist, written first, and what it caught
The checklist for the account snapshot, written before the run:
Source of truth: 10-data/accounts.csv and 10-data/opportunities.csv, as shipped.
Reconciliation 1: row count of opportunities for this account_id, against the
distinct count of opportunity_id. State both.
Reconciliation 2: sum of Closed Won New Business and Expansion amounts, in the
account's currency, against accounts.arr_usd converted at the
documented fixed rate. Renewal amounts excluded: they restate
ARR (DATA-DICTIONARY.md).
Sample: all rows. This account has 4, so the sample is the population.
On a 400-row output the sample would be 10 rows plus the top
decile by amount.
Tripwire: OPP-007099 is a known duplicate id. It must appear once in any
count and its conflict must appear in exceptions.csv.
Cost: about 3 minutes.
Why reconciliation 2 has a clause in it: without the exclusion, the check itself is wrong.
Three of the four rows on ACC-02848 are type = Renewal, and renewal amounts restate
current ARR. A check that sums everything against ARR fails on a correct output, which is
worse than no check: you will then "fix" a correct output.
Why a sample size appears on a four-row account: the checklist is a template for accounts
with forty rows, and the sample rule must exist before it is needed.
What it caught, on the run from Lesson 1. Reconciliation 1: 4 rows, 3 distinct ids,
disagreement flagged. Tripwire: OPP-007099 appeared once and its conflict appeared in
exceptions.csv with both amounts. Reconciliation 2: the single New Business amount, 55,000
USD, matched accounts.arr_usd of 55,000 exactly.
What it did not catch: which OPP-007099 amount is right. Nothing in the data settles it.
ARR agrees with the 55,000 row, and that row has the blank close date. The checklist's job
was to make sure I found out there was a question. Answering it means asking Ursula Ulrich,
who owns the opportunity. A good verification produces questions for humans. A bad one
produces confidence.
Consistency check or verification: tell them apart
The most expensive confusion in the course, because a consistency check feels rigorous.
Numbers were compared. Something agreed with something. The something was itself.
A consistency check compares the output against itself. A verification compares the
output against the source. An output that is internally consistent can be entirely wrong,
and it will look better than a correct output that raises a question.
Example A: a consistency check
The snapshot for ACC-02848 prints:
Opportunities: 4
Total value: 234,000 USD
Average: 58,500 USD
Check it: 55,000 plus 55,000 plus 69,000 plus 55,000 is 234,000, and 234,000 divided by 4 is
58,500. Every number agrees with every other number on the page.
Two things are wrong, and neither is visible from the page.
It counts one opportunity twice.OPP-007099 appears on two rows (PLANTED-ISSUES.md,
class duplicate-opportunity-rows). There are three opportunities, not four.
It adds renewals to new business. Three of the four rows are type = Renewal, and
renewal amounts restate current ARR rather than adding to it (DATA-DICTIONARY.md in your
pack's data folder). So
234,000 is not the total of anything: not lifetime value, not ARR, not bookings.
And the account's recorded ARR is 55,000 USD. The page is out by a factor of more than four,
and internally perfect.
Example B: a verification
Same output, three comparisons against source:
Row count against distinct identifier count, from opportunities.csv: 4 against 3.
They disagree. A source comparison, because the distinct count is a property of the file.
Sum of Closed Won New Business and Expansion amounts against accounts.arr_usd:
55,000 against 55,000. Agrees, and only because the check names which types go into the
sum.
Tripwire: OPP-007099 is a known duplicate. Does it appear once, and is the conflict
in exceptions.csv?
Check 1 fails, and that is the check doing its job. The corrected output says: three
opportunities; current ARR 55,000 USD as recorded on the account; one unresolved conflict on
OPP-007099, two amounts, one blank close date, routed to exceptions. That output is uglier
than the wrong one, and it is the only one you can put your name on.
What good feedback on this looks like
A learner submits the pass-1 roll-up with the note: "verified: the amounts sum to the stated
total and the count matches the number of rows returned." The response:
You ran a check and it passed. The problem is what it was a check of. You compared
three numbers in your output against each other. That is a consistency check. It proves the
arithmetic on the page is sound, and it cannot detect any error that entered before the
arithmetic started.
You never compared anything against opportunities.csv or accounts.csv. Two comparisons
would have taken ninety seconds. Row count against distinct opportunity_id count: four
rows, three distinct ids, because OPP-007099 appears twice with different amounts and one
blank close_date, both documented, in classes duplicate-opportunity-rows and
missing-values. Your "4 opportunities" is a row count wearing an opportunity count's
label. And your total against accounts.arr_usd: the account records 55,000 USD, you
printed 234,000. Not a rounding disagreement, a category error: three of your four rows are
renewals, and renewal amounts restate ARR.
Why this matters more than the specific error: a consistency check almost always passes,
because one process produced the output, so its parts agree by construction. "Verified"
after a check that cannot fail transfers your confidence without any evidence, and the next
person reads it and stops looking.
The repair is one line in your checklist, written before the run: name the file and column
each output number will be compared against. A number with no named source is either
derived from numbers that have one, and you say from which, or it does not go in the
output.
One thing you got right: you stated what you checked. Most people say "verified" and stop.
Because you wrote it down, this conversation is possible. Keep the habit, change what goes
in the sentence.
What you keep
revenue-workspace/40-reference/verification-checklist.md, a file you write yourself
(template: templates/M4/verification-checklist.md in your practice pack),
written before the next run, containing:
the named source of truth, by file and column;
at least one reconciliation, with the clause that makes it meaningful on your data;
a stated sample size and how it was chosen;
at least one tripwire row and its known correct treatment;
the estimated cost in minutes, which goes inside your time-saved number.
Plus exceptions.csv from your runs, kept rather than deleted: the evidence of what your
brief does not yet handle.
Check yourself
Three verification procedures are proposed for a 400-row hygiene pass: (a) re-run the pass
with a second brief and compare the outputs; (b) reconcile row counts and sums against
source, then sample 20 rows including the top decile by amount; (c) read every row. Pick
the one that is sufficient and cheaper than the work, then say why (c) is wrong and why
(a) is worse than it looks.
Your exceptions file has zero rows after a pass over the shipped accounts data. Name the
two most likely explanations and the first thing you would check.
A run reports "totals reconciled". Write the one question that tells you whether that was
a verification or a consistency check.
Your workflow saves 40 minutes and its verification takes 35. What number goes in your
delegation policy, and what would have to change for the workflow to be worth keeping?
Answers
1. (b). One line per option:
(a) Re-run and compare: wrong, and worse than it looks. Two outputs agreeing is a
consistency check with extra steps. Both runs read the same source the same way, so they
can agree and both be wrong, at the cost of a second full run.
(b) Reconcile against source, then sample 20 rows including the top decile: correct.
Two instruments, both comparing against source, minutes against a 400-row job. The top
decile is there because the distribution of consequences is not the distribution of rows.
(c) Read every row: sufficient and wrong. More expensive than doing the job by hand,
which is the rule that overrides thoroughness. A check nobody performs twice is not a
check.
2. Either the exceptions rule was not in force, or the scope was narrow enough that
nothing ambiguous fell inside it. Check whether an in-scope ambiguous case was resolved
silently.
3. "Which file and column was each number compared against?" No answer means nothing was
compared against source.
4. You write 5. To make it worth keeping, either cut the verification cost (a
reconciliation plus a sample rather than a full read) or raise the baseline by running it
over more rows per pass.
Where this goes next
Next lesson: what to do while a run is going wrong in front of you, and the record that
lets you answer "where did this number come from" six weeks later.
Track labs: the exercise sheet in labs/LAB-04 of your practice pack (start with its
README). Part A pass 3 requires you to write the checklist, run the
reconciliation and a 10-row sample against source, find the discrepancy, and record it. Every
findings-list lab from Module 5 onward is scored on detection.
Module 4 · The operator's working method, assign, check, keep
M4 lesson / about 11 min
Do it
Steering, and what survives the session
Do this nowCount the owners60s
Filter opportunities.csv to the six open stages.
List the distinct owner strings in that filter and count them.
Compare that count against the eight account executives on the team.
You know it worked when: Nine distinct owner strings for eight people, because OPP-006934 carries "Dagmar Wexford" with two spaces.
What you type
Read opportunities.csv in this folder.
List every open opportunity (stage in Qualification, Discovery, Solution Review,
Technical Validation, Proposal, Negotiation), grouped by owner, with the row
count per owner.
Before grouping, normalise owner: trim leading and trailing whitespace, collapse
repeated internal spaces, and match case-insensitively. Report every raw variant
you merged and how many rows it covered.
Do not modify opportunities.csv.
Notice: The normalisation rule belongs in CLAUDE.md and not in this session, because the second time you correct the same misunderstanding you are describing your brief rather than the run.
There is no memory between sessions, so the five things written to files cross the session boundary and everything said inside the session does not.
A run is going wrong in front of you. It has read the wrong file, or taken "the pipeline" to
mean something you did not mean, three minutes into a job that will take eight.
Most people do one thing: keep typing corrections until it looks right. That produces an
output you cannot reproduce, because the corrections lived in a session that will not exist
tomorrow. Then, six weeks later, someone senior asks about a number you no longer remember
producing.
What you will be able to do
Steer a running task with three moves, and spot the signal that says stop steering.
Keep a run record that lets a colleague reproduce a run and lets you defend a number a
month later.
Say plainly what survives a session and what does not.
Three steering moves, and one rule
Interrupt and narrow. "Stop. Only the EMEA accounts." For when the work is right and the
scope is wrong. Cheap, and it usually works.
Interrupt and correct. "Stop. The amount field is in local currency, use the account
currency and convert." For when a specific fact is wrong. Also cheap, with a catch: it fixes
this run only. The fact belongs in the folder file or the brief, or you will say it again
next Tuesday.
Abandon. Stop, do not correct, rewrite the brief and start again.
The rule that tells you which one you are in:
If you have corrected the same misunderstanding twice, stop and rewrite the brief.
Twice. Not "when it gets annoying". The second identical correction is evidence about your
brief, not the run: something in it is genuinely ambiguous, and you have confirmed it by
observation. A third in-place correction produces work you cannot reproduce, because
corrections live in a session that will not exist tomorrow. Abandoning feels like losing
three minutes. It is buying back every future run.
Before you abandon, write down the correction you were about to make. It is usually the
missing line in your Constraints, easier to capture now than later. Which of the three moves
did your last frustrating session need?
The run record
One line per run in 50-runs/RUN-LOG.md, a file you keep yourself: start it from
templates/M4/RUN-LOG.md in your practice pack. Nine fields, and each one exists because a specific
question gets asked later.
Field
The question it answers
Date
"When was this produced?"
Brief file used
"What exactly was asked for?"
Input files and their dates
"Was this before or after the July export?"
Output files
"Which file are we even looking at?"
Verification performed, and the result
"Did anyone check this?"
Exceptions count
"What did it not handle?"
Blast-radius class
"Did this need a reviewer?"
Approver, if draft-only or writes-to-system
"Who said yes?"
One line of note
"Anything odd?"
The input file dates are the field people skip and the field asked about most. "The
pipeline export" is not an answer when there have been six of them, and an entry without
dates cannot say whether two runs used the same data, the first question anyone asks when two
reports disagree.
Worth knowing: the Compliance API and Audit Logs do not cover Cowork yet. At team level
that is a rollout question for Module 8. At your level it means the audit trail is a file you
maintain, about twenty seconds per run. Do not make the log clever: a markdown table you
actually fill in beats a system you abandon in three weeks.
Your memory is a folder
Here is the sentence the whole module is built to earn: there is no memory between sessions.
Continuity is a file convention, not a product feature. What survives is exactly what you
wrote to disk, nothing else.
What survives
Where it lives
What it does next time
The brief
20-briefs/<workflow>.brief.md
The next run costs a sentence instead of a rewrite
The folder rules
revenue-workspace/CLAUDE.md
Every task in this folder inherits your conventions
The verification procedure
40-reference/verification-checklist.md
You check the same way twice, which is what makes two runs comparable
The run record
50-runs/RUN-LOG.md
You can answer "where did this number come from"
The exceptions file
next to each output
You know what your brief does not handle yet
What does not survive: the correction at minute four, the clarification at minute six, the
reason you chose that folder. If those mattered, they were supposed to become one of the five
rows above.
The test that separates Assign from Keep: deleted, and the next run is still correct but
slower to set up? Assign. Deleted, and you can no longer prove what happened or reproduce the
judgment? Keep.
Which of the five files are you missing, and what question can you not answer without it?
A real correction, and the moment it became a rewrite
The task. This week's list of open opportunities whose close date has already passed,
grouped by owner, from the shipped data. Fourteen such rows, documented as class
overdue-open-pipeline in answer-keys/PLANTED-ISSUES.md in your pack, which you open only
after attempting the lab.
Minute two, correction one. The output had more owner groups than there are account
executives. I interrupted: "Trim whitespace on owner and treat Dagmar Wexford with two
spaces as the same person as Dagmar Wexford." The issue log documents six
inconsistent-values rows in opportunities.owner, including OPP-006696, OPP-006746
and OPP-006934 as Dagmar Wexford, and OPP-006795 as bettina costa.
Minute five, correction two. The same thing on a different spelling, Aisling Calderon
with a trailing space (OPP-006805), which I had not named.
The decision. Second correction of the same misunderstanding: owner names are dirty and
my brief never said what to do about it. I stopped. What I wrote instead, before restarting,
went into the folder file rather than the brief, because it is true of every task that
groups by owner in this data:
The owner field in opportunities.csv is not clean. Before grouping by owner,
normalise: trim leading and trailing whitespace, collapse repeated internal
spaces, and match case-insensitively. Report the raw variants you merged and
how many rows each covered.
The cost, honestly: the restart took four minutes, the two corrections had already cost
three. A third correction would have left a correct output with no route back to it, and next
Tuesday the seven minutes again. The rule paid for itself on the second run. The run log
entry, written immediately after:
| 2026-08-02 | overdue-pipeline.brief.md | opportunities.csv (2026-08-02 copy),
accounts.csv (2026-08-02 copy) | 30-output/overdue-pipeline/2026-08-02.md |
Reconciliation: 14 rows out, 14 open rows with close_date < today in source.
Sample 10 checked against source, all correct | exceptions: 2 |
read-only | n/a | Rewrote brief after second owner-normalisation correction.
Rule moved to CLAUDE.md. |
Twenty seconds, and six weeks from now that entry answers which export, which brief, whether
anyone checked, and why the brief changed.
What you keep
The complete brief pack, four files you created across this module, the module artifact:
20-briefs/account-snapshot.brief.md
40-reference/verification-checklist.md
50-runs/RUN-LOG.md, with at least three real runs recorded
revenue-workspace/CLAUDE.md, updated with the output conventions and the standing
exceptions rule
Try this now
The reusability test is literal. Before you call this done, hand the brief file to another
person, or open it yourself in a clean session with nothing else, and see whether it works
without additions. A brief only its author can run is not reusable, however good the
output.
Check yourself
You have corrected the same misunderstanding twice in a running session. State the correct
next move and give the reason in terms of reproducibility, not frustration.
A run log entry records the brief, the outputs and the verification, but not the input
file dates. Name the question it can no longer answer, and give a situation where that
question is the only one that matters.
Your workflow now runs from a saved brief. Which of these breaks it: deleting the brief
file, the run log, CLAUDE.md, or last month's output? For each, say whether what you
lost was Assign or Keep.
You measured three numbers on this workflow: time to first output, time to verified
output, time to second run using the saved brief. Which one justifies the method, and
which one is usually quoted instead?
Answers
1. Abandon and rewrite the brief. Not because it is annoying, but because a correction
lives in a session that will not exist tomorrow. A third correction gives you an output
nobody can reproduce, including you. Write down the correction you were about to make: it is
the missing line in your Constraints.
2. "Did these two runs use the same data?" That becomes the only question that matters
the moment two reports disagree and somebody senior is deciding which to believe.
3, one line per item:
Brief file: breaks it. Assign. Not runnable as it stands, though rewritable from the
output.
Run log: does not break the next run. Keep. Nothing stops working and you can no
longer prove what happened.
CLAUDE.md: degrades it quietly. Assign. The run still produces something,
conventions gone, nothing announcing their absence. The one people mis-sort, because the
exceptions rule it carries is what produces a Keep artifact.
Last month's output: breaks nothing, and it is neither. A product of a run,
reproducible from the brief.
4. Time to second run justifies the method, and it is usually the smallest of the three.
Time to first output is the number people quote, and it says least.
Where this goes next
Next module: Module 5 runs T, R, A, C, K in order against CRM data hygiene. Every
findings list from here on is measured on detection: how much you found, how much was real,
whether you caught the defects that do not announce themselves.
Track labs: the exercise sheet in labs/LAB-04 of your practice pack (start with its
README). Part B builds this same pack on your own top shortlist workflow from
Module 3, behind the redaction gate, and measures time to first output, verified output and
second run.
Module 4 · The operator's working method, assign, check, keep
Lab / ? min / read-only
Lab
Request, brief, verified brief
LAB-04: Request, brief, verified brief
The situation
You handed something over and got something back. It looks right. Three questions decide whether it was
worth doing:
Would you sign your name under it, and on what evidence?
Can you get the same thing next Tuesday without retyping the request?
If a VP asks where a number came from, can you show them?
Most people who try an agent for revenue work answer no to all three, conclude it is unreliable and go
back to doing it by hand. The gap is not the model. They wrote a request instead of a brief, checked by
reading rather than by evidence, and kept nothing.
This lab runs the same task three times and makes the difference visible. The account you run it on has
a defect planted in it that a reconciliation catches and a careful read does not, which is the point.
What you will produce
File
What it is
30-output/pass-1-request/
The one-line request output. Kept as the before picture
20-briefs/account-snapshot.brief.md
The seven-part brief
40-reference/verification-checklist.md
The reconciliation, the sample size and the known-answer procedure, written before the first run
30-output/pass-3-verified/
The verified output plus exceptions.csv
50-runs/RUN-LOG.md
At least three runs recorded
revenue-workspace/CLAUDE.md
Updated with the output conventions and the standing exceptions rule
Outcome accounting
Three numbers, not one. MODULES/04 requires all three, and the third is the one that justifies the
method.
1. Time to first output (pass 1, the one-line request)
2. Time to verified output (pass 3, including the reconciliation and the 10-row sample)
3. Time to second run using the saved brief <- the number that matters, and it is the smallest
Observed while building this lab, stated as observed and not as a promise:
pass 1 (request) 3 min, and the output was wrong in a way I did not see
pass 2 (write the seven-part brief, run it) 14 min, of which 11 was writing the brief
pass 3 (write the checklist, run, verify) 16 min, of which 9 was the manual sample
second run on a different account 4 min including the reconciliation
Manual baseline for the same page by hand: 19 min (measured in LAB-01, one account, timed)
The arithmetic that matters:
First account: 33 min against a 19 min baseline. The method LOSES 14 minutes on run one.
Second account: 4 min against 19. Saves 15.
Break-even: run two.
Five accounts: 33 + (4 x 4) = 49 min against 5 x 19 = 95 min. Saves 46 minutes.
Your actual: 1 _____ 2 _____ 3 _____
Verification cost: inside every number above. Pass 3's 16 minutes includes 9 minutes of manual
checking, and the 4-minute second run includes the reconciliation. There is no
version of this lab's numbers with the checking removed, because a number with
the checking removed is not a number about this method.
Blast radius
Read-only. Everything writes into 30-output/. 10-data/ is not modified: your LAB-02 CLAUDE.md
says so and this lab is where that rule earns its keep, because a reconciliation against a source you
have already cleaned is not a reconciliation.
Declared scorecard scope
None. LAB-04 produces no findings list, so no recall or precision is computed. It does use two
planted rows as known-answer tripwires, which is a different instrument: rows whose correct treatment
you decided in advance, used to confirm the check works. See part-a.md step 4 and SCORING.md §1.
How you verify your own output
The verification checklist you write in step 4 is the answer to this section, which is the point of
the lab. It has three parts, and MODULES/04 beat 5 names all three:
Reconciliation. Totals in the output agree with totals in the source: row counts, sums of
amounts, distinct account counts. This catches silent drops, which are the most common failure and
the hardest to see.
Sampling. n rows checked by hand against source, with n stated. Rule of thumb, labelled a rule of
thumb: at least 10 rows, at least 5% for anything draft-only, and every row in the top decile by
amount for anything that touches a number a VP will read.
Known-answer rows. Rows whose correct treatment you already know, confirmed to have come out
right. On your own data you plant these yourself. On the shipped Talvera data you do not need to:
rows with a known correct treatment already exist and are documented.
The four duplicate-opportunity-rows IDs, key 20260801, are OPP-006790, OPP-006886,
OPP-007099 and OPP-007110. Each appears on two rows with different amounts. Each must be
counted once, not twice, and any roll-up that counts one twice has failed a check you could have
designed in advance.
Part A puts you on an account that contains one of them.
Governing rule: if checking costs more than doing it by hand, do not hand it over. Write the
verification procedure before you run the task, in the brief, so you cannot be talked into believing
an output you find pleasant.
Files you need
course/labs/data/ in 10-data/talvera/, plus the workspace from LAB-02 and the shortlist from LAB-03.
What comes next
Modules 5 to 8 reproduce this pack for a different body of work. The brief pack is also three of the
seven capstone rubric rows (CAP-A, CAP-C, CAP-K).
Module 4 · The operator's working method, assign, check, keep
6 questions / pass at 70%
Checkpoint
Module 4 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
The brief:
"Outcome: a pipeline summary for the EMEA region I can send to the VP.
Inputs: 10-data/opportunities.csv and 10-data/accounts.csv, joined on account_id.
Output shape: 30-output/pipeline/emea-summary-2026-08-03.md, one section per stage, each with deal
count and total amount.
Constraints: open stages only. Do not include closed opportunities.
Definition of done: every open EMEA opportunity appears in exactly one stage section.
Stopping condition: stop and report if more than 5% of rows cannot be assigned a region.
Exceptions rule: anything unresolvable goes to exceptions.csv with the row identifier and the
reason."
Which of the seven parts failed, and how do you know?
Hint
Ignore the shape of the output and interrogate one column. Ask what unit the numbers in it are in, then
go and check what the data dictionary says that column holds.
A. Incorrect. A stopping condition threshold is a judgment call and 5% is a defensible one. Zero would be worse, not better: it stops the run on the first ambiguous row, which is what the exceptions rule exists to handle without stopping. Nothing in the output suggests the stopping condition fired at all.
B. Incorrect. You cannot conclude anything about dropped rows from a count you have not reconciled, and that is the actual lesson hiding behind this option. If you suspected a silent drop, the move is to reconcile the section counts against the source row count for the same filter, not to judge the total against an intuition about how big EMEA ought to be.
C. Incorrect. Renewals are a real modelling decision and worth deciding on purpose, because renewal amounts restate current ARR and should not be added to new business when computing ARR. But it is a refinement, not the failure. The report is wrong before you get to the question of which deal types belong in it.
D. Correct. "Total amount" is an output shape that names a column and not a unit. The data dictionary is explicit that opportunity amounts are in the account's local billing currency, so a sum over EMEA adds euros to pounds and reports the result as a number a VP will read as money. Note what makes this the module's signature failure: the output is internally consistent, the stage counts add up, and nothing on the page announces itself as wrong. The fix is in the brief, not in the output: name the currency, name the conversion source, and require the summary to state which it used.
Question 2 of 6
The task: produce a monthly roll-up of closed-won amount by segment from opportunities.csv joined to
accounts.csv. Doing it by hand takes about 50 minutes. Three verification procedures are proposed.
Procedure 1. Read the output and confirm the segment totals sum to the grand total.
Procedure 2. Reconcile the output's distinct opportunity count and summed amount against the same
filter applied to the source; then hand-check 10 rows chosen across segments in proportion to segment
size; then confirm the four known duplicated opportunity IDs are counted once rather than twice. About
12 minutes.
Procedure 3. Hand-check every closed-won row against source. About 3 hours, and it would find
everything.
Which do you pick, and why is the thorough one the wrong answer?
Hint
Two of the three are defensible on correctness. Put the manual baseline (50 minutes) next to the cost of
each check and ask which one you would still be running in month four.
A. Correct. You picked the procedure with all three cheap evidences in it and you gave the right reason for rejecting the thorough one. The rule is not that thoroughness is bad, it is that if checking costs more than doing it by hand then the workflow failed triage on verification cost and should not have been handed over. There is a second reason, and it is the one that bites: a three-hour check is run once, out of good intentions, and never again, so the workflow ends up with no verification at all from month two onward.
B. Incorrect. This is the answer that sounds most responsible and is the most likely to leave you unprotected. Three hours against a 50-minute manual baseline means you spent more to check the output than to produce it yourself, which is the definition of a workflow that should not have been delegated. It is also the check that quietly stops happening, and an unverified monthly number is worse than a manual one because everybody now believes it.
C. Incorrect. Procedure 1 is a consistency check, not a verification, and telling those two apart is the discrimination this module is built on. An output whose totals agree with themselves is internally consistent and can still be wrong against the source in every row. It would not catch a silent drop, and it would not catch a duplicated primary key inflating a segment, because the inflated number is included in both the part and the whole.
D. Incorrect. Right procedure, wrong justification, and the justification is what you carry to the next task. Ten is a rule of thumb and the course labels it as one. What makes procedure 2 sufficient is that the sample has a stated selection rule, that a reconciliation sits in front of it, and that it includes rows whose correct treatment was known before the run. A sample size with no selection rule is decoration.
Question 3 of 6
You ran a contact hygiene pass over contacts.csv, which you know is messy: it contains addresses with
two at-signs, addresses with no at-sign at all, a value that is just the word none, role-based
mailboxes, and rows in upper case.
exceptions.csv came back with a header row and nothing else.
What does this most likely mean?
Hint
The exceptions file is not a log of failures, it is a measurement of how much your brief did not cover.
Ask what an honest run on this data should have been unable to decide.
A. Incorrect. Possible, and it is the third thing to check rather than the first. A file with a header and no rows is what the instruction produces when nothing was written to it, which is exactly the situation the correct answer describes. Assuming a mechanical fault lets you skip the question of what the agent did with `none` in an email column.
B. Incorrect. This is the reading that makes the instrument useless. Zero exceptions is only good news on data where you had independent reason to expect zero ambiguity, and the stem tells you the opposite. The empty file is the cheapest early warning you will get on this run, and treating it as a pass is how the wrong classification reaches the proposed-changes file.
C. Correct. The exceptions file is a signal about the brief, and reading it is the debugging move the module is built around. On data you know contains a value of `none` in an email column, a run that resolved everything did not find clarity, it made decisions you did not ask for and did not record. Note the symmetric reading, because you will need it: a very long exceptions file usually means the brief was underspecified too, just in the opposite direction, and rarely means the data was bad.
D. Incorrect. The stem states what is in the file, so this option asks you to disbelieve the premise rather than diagnose the output. It is worth naming because it is a real habit: when an output disagrees with what you knew about the data, the reflex to trust the output over your own knowledge is the reflex the Check letter exists to interrupt.
Question 4 of 6
You are twenty minutes into a run. You have now told it twice that renewal amounts restate current ARR
and must not be added to new business. It acknowledged both times and the third intermediate output has
done it again.
What is the correct next move, and what is the reason?
Hint
Ask where the correction you just gave will be tomorrow morning. Then ask where it would be if you had
put the same sentence somewhere else.
A. Incorrect. Patching output is the habit this module exists to break. You will fix the renewal rows you can see, the same defect will be in next month's run, and the run record will show a verification that passed on a file you hand-edited, which is not a verification. Fix the rule, not the row.
B. Incorrect. The twenty minutes are already spent and cannot be recovered by continuing, which is the trap. Even if the third correction lands, what you own at the end is an output you cannot reproduce and a brief that still does not contain the rule. Next Monday you pay the twenty minutes again, plus three corrections, and you have no record of why the number is what it is.
C. Correct. The abandon rule is that if you have corrected the same misunderstanding twice, stop and rewrite the brief. The reason is not that the model cannot be corrected, it is that the correction is not durable: there is no cross-session memory, so a fix that exists only in a running conversation is thrown away with the conversation (FACTS §7.1). Adding a worked example of a correct row is the specific technique that turns a repeated correction into an instruction that holds.
D. Incorrect. This over-reads the signal. Two corrections tell you the brief is wrong, not that the task is unsuited. The renewal treatment is a written rule about a named field with a source of truth behind it, which is the easiest kind of constraint to specify. Abandoning here means abandoning most workflows on their first difficult afternoon.
The inputs column names the entities and not the files or their dates. Which question can this entry no
longer answer?
Hint
Imagine re-running this brief today and getting a different total. Work out which column you would need
in order to tell whether anything actually went wrong.
A. Correct. A run record exists to answer "where did this number come from" a month later, and an input without a version is not a source. `opportunities` names an entity; `10-data/opportunities- 2026-07-13.csv` names an artifact you can go and open. The distinction becomes load-bearing the first time somebody challenges a number: without it you cannot separate a data refresh from a rule change from an error, and this is your own audit trail because the enterprise Compliance API and Audit Logs do not cover Cowork yet (FACTS §5.10, §7.7).
B. Incorrect. "Not applicable, read-only" is a correct entry, not a gap. Read-only work needs no approver, and recording the class alongside the absence is exactly what makes the entry auditable: a reader can see that the question was asked and answered rather than skipped.
C. Incorrect. "Reconciled, 10-row sample passed" is thin and it is not the defect. It names the two evidences that were run and their result, which is what the column is for. If you want it stronger, the improvement is to name what was reconciled against what, and that is a refinement rather than a question the entry cannot answer at all.
D. Incorrect. Duration is useful for outcome accounting and it belongs in the lab's measurement block rather than in the audit line. Nobody defending a number six weeks later needs to know it took eleven minutes. They need to know which file it read.
Question 6 of 6
Four rules. Each belongs in global instructions, in the folder's CLAUDE.md, or in the brief file for
one workflow.
Never use a table where a list would do. I read these on a phone.
In this workspace, output files go to 30-output/<workflow>/<name>-YYYY-MM-DD.md.
For this weekly pipeline pack, exclude opportunities below 5,000 in account currency and say so in
the exclusions section.
Anything you cannot resolve goes to exceptions.csv with the row identifier and the reason. Do not
guess.
Hint
For each rule ask the smallest scope at which it is still true. One of the four is true of you, one is
true of the workspace, one is true of a single job, and the fourth is more local than it sounds.
A. Incorrect. Rule 3 is the error. "Anything about a repeating workflow belongs with the folder" collapses two levels that need to stay separate: the folder holds what is true of all work here, and the brief holds what is true of this job. Push a threshold up to the folder and it silently applies to the next workflow you write, which is the kind of inherited constraint nobody remembers setting.
B. Incorrect. Rule 2 names a path that exists in this workspace and nowhere else, so as a global rule it would be wrong the moment you open a different folder. Rule 4 in the brief file is not harmful, it is just work: you restate it in every brief you ever write, and the day you forget is the day the agent guesses.
C. Incorrect. You have inverted rules 1 and 4. How you prefer to read output is a property of you, not of a project, and it is the textbook case for global instructions along with tone, role background and output format. The exceptions rule names a file in this workspace, which is what ties it to the folder.
D. Correct. The split holds at three levels. Rule 1 is a fact about you and it is true in every session, which is the global level (FACTS §5.7). Rule 2 is a convention of this workspace, so it lives in the folder file, read at session start. Rule 3 has a threshold and a scope that belong to one workflow and would be wrong applied to another, so it lives in that workflow's brief. Rule 4 is the interesting one: it reads like a universal habit and it is a workspace convention, because the file it names is a path in this workspace. Putting it in the folder file means every brief inherits it without restating it.
Last verified: 2026-08-06.
Product claims in this module trace to course/reference/COWORK-FACTS.md (cited as FACTS §x.y).
Dataset claims trace to course/data/DATA-DICTIONARY.md and the pack's answer-keys/PLANTED-ISSUES.md
for the Talvera dataset generated with seed 20260801.
The module at a glance
The forecast is wrong and everybody knows it. Nobody can say by how much, because the export it is
built from contains records that are counted twice, amounts in three currencies added together as if
they were one, and open deals whose close date passed months ago.
This module builds the hygiene pass that turns "the data is bad" into a reviewed, importable set of
proposed changes plus an honest list of everything the pass refused to guess about. It runs the full
TRACK loop once, over CRM data hygiene.
Lesson
Title
TRACK
Minutes
Module-level piece
L5.1
What a hygiene pass may touch, and what it must never decide
T, R
10
L5.2
Name the defect classes before you count anything
A
13
Worked exercise
L5.3
Match rules with bands, and the merge that costs you a customer
A
11
Discrimination pair
L5.4
Reconcile, then sample, then earn the write
C, K
14
Failure lane
Total lesson time: 48 minutes. Module budget is 45 (PEDAGOGY.md §3.1), tolerance plus or minus
15%.
Then: checkpoint CP-05 (about 10 minutes) and LAB-05 (Part A 55 minutes, Part B Full 40 or Part B
Lite 15).
Outcomes
By the end of this module you can:
O5.1 Produce a data quality audit over account, contact and opportunity exports that quantifies
each defect class with counts, percentages and named examples.
O5.2 Build a duplicate candidate list from a match rule you wrote down, split into an
auto-mergeable band and a human-review band, with the rule stated for each band.
O5.3 Produce a field completeness report against a required-field policy, broken down by owner
and by segment, that names who has to fix what.
O5.4 Produce a draft-only proposed-changes file suitable for review and import, paired with an
exceptions file covering everything unresolved.
O5.5 Verify the audit against source with a stated sample and at least one reconciliation, and
correct a defect the audit got wrong.
O5.6 State the conditions under which a hygiene workflow may be promoted from draft-only to
writes-to-system, and write the reversal plan that promotion requires.
Outcome IDs are defined in curriculum/OUTCOMES.md. Every one is assessed.
The artifact you leave with
The hygiene pack, in your own revenue-workspace/:
File
What it is
Built in
40-reference/field-policy.md
Required fields, valid values, staleness thresholds, defect definitions, match rules with bands
L5.2, L5.3
20-briefs/crm-hygiene-audit.brief.md
The seven-part brief, rerunnable monthly
L5.2, completed L5.4
30-output/hygiene/audit-YYYY-MM-DD.md
The defect report
L5.2
30-output/hygiene/duplicates.csv
Banded duplicate candidates with the rule that matched each
L5.3
30-output/hygiene/proposed-changes.csv
The draft-only import file
L5.4
30-output/hygiene/exceptions.csv
Everything the pass refused to guess about
L5.4
50-runs/RUN-LOG.md entry
Reconciliation result, sample result, who ran it, when
L5.4
Starting templates for the two durable files are in templates/ beside this README. Copy them into
your workspace, do not work inside the course folder.
The field policy is the one that outlives the course. It is a RevOps document, it is portable to any
CRM, and most teams have never written one.
Blast radius
Everything in this module is read-only or draft-only. Nothing here writes to a CRM.
Promotion to writes-to-system is taught in L5.4 as a set of conditions to be met, and every learner
on every track writes the promotion checklist and the reversal plan as a paper exercise in LAB-05
Part A. Executing a tested reversal on a live batch is TL-OPS-2, for RevOps learners only.
Prerequisites
Modules 1 to 4: the workspace layout and fenced folders (M2), the delegation policy and blast-radius
vocabulary (M3), the seven-part brief, the three cheap evidences and the exceptions rule (M4).
What this unlocks
Module 6 assumes the opportunity data is trustworthy enough to inspect. Module 8 assumes an account
record you can reason about. On the OPS track this is the deepest module you take.
A note on the numbers in these lessons
Every count in these lessons was computed from the shipped dataset at the seed named above. They are
synthetic. They are also exact, and you can reproduce every one of them, which is the point: if a
number in a lesson does not match what you get, one of us is wrong and it is worth finding out which.
The generator is seed parameterised. A different seed produces the same shape of problem with
different identifiers, so memorising OPP-006790 teaches you nothing. Learning to reconcile a sum
against its source does.
Module 5 · The truth layer, CRM hygiene at volume
M5 lesson / about 10 min
Concept
What a hygiene pass may touch, and what it must never decide
Do this nowTry to pick a merge survivor60s
Open accounts.csv and put ACC-02947 and ACC-03047 side by side.
Decide, from the export alone, which of the two records should survive a merge.
Write down the field you used, or write down that no field answers it.
You know it worked when: Every identifying field is identical (Halloway Logistics AG, halloway-logistics.example.com, Germany, 712 employees, owner Bettina Farrow) and the only difference is created_date: 2025-03-15 against 2024-09-19.
What you type
Read every CSV in 10-data/. For each file, report the filename, the row count,
and the column names in order. Then state whether any file failed to parse.
Do not clean, interpret or summarise anything. Do not write any file.
Notice: If any of those three row counts differs from what your CRM export tool reported, stop here, because every number you produce afterwards inherits the difference.
You are the person three teams ask for a number, and you know the number is not clean. Last quarter the board pack and the CRM disagreed, and the reconciliation cost you two evenings. Now a well-meaning leader suggests the obvious fix: point an agent at the CRM and let it clean up. This lesson is the answer you give.
Everything this module uses came in the practice pack from the course page; if you have not downloaded it and unzipped it into your course folder, do that first.
What you will be able to do
Score three hygiene tasks on the five-factor fit test and say which two you would hand over today.
Say why the third stays human, even though it is the most tedious.
Choose between a file export and a live CRM connector, and defend the choice in one sentence.
Three hygiene tasks, three different answers
"Clean up the CRM" is not one task. It is at least three, and they score differently. Run them through the Module 3 fit test: Volume, Checkability, Judgment density, Cost of being wrong, Specification cost, each 1 to 5. A low Checkability score is a veto, whatever the rest say.
Count the defects
Find duplicate candidates
Choose which duplicate survives
Volume
5. Three files, 1,747 rows in the shipped set
5. Every account against every other
3. Fifteen or twenty decisions a month
Checkability
5. Every count is reproducible from the export
4. Every candidate pair can be read by a human
1. No file says which record is right
Judgment density
1. None, once the definitions exist
2. Low, once the rule exists
5. High, and most of it is outside the CRM
Cost of being wrong
2. A wrong count is embarrassing and reversible
2. A wrong candidate is rejected at review
5. A wrong merge destroys history for two teams
Specification cost
2. Twenty minutes if you know your fields
3. An hour to write the rule properly
5. You cannot write down what "right" means
Verdict
Hand it over
Hand it over, with a review band
Keep it human. The agent proposes, a person decides
Row counts: accounts.csv 308, contacts.csv 911, opportunities.csv 528 in the shipped Talvera dataset (DATA-DICTIONARY.md in your pack's data folder). Together: 1,747.
The third column is the lesson. Which duplicate survives depends on things the export cannot see: which record billing points at, where support's tickets sit, whether the rep on one is mid-negotiation. Checkability scores 1, and that is a veto. The agent brings you the pair and the evidence. You make the call.
Work from an export first, connect later
Two ways to give an agent CRM data: a file export in a folder you granted, or a connector wired to the live system. Connectors are MCP servers wired to your tools, and installing a plugin can add one.
For a first hygiene pass, use the export. Three reasons your CRM owner will accept:
A snapshot you can reconcile against. Re-run the same brief on the same file next week: any difference is your brief, not the data moving under you. That is what makes checking possible.
No write path. Cowork reads, creates and modifies files only inside the folders you grant it. A CSV in 10-data/ cannot become a CRM record by accident. No such guarantee exists once a connector with write scope is authorised.
Read-only blast radius. You get to be wrong for free until the rules are right.
Default: export to 10-data/, output to 30-output/hygiene/, connector later or never.
Treat any CRM connector as writes-to-system until you have read its permission screen yourself. Our fact base names Gmail, Google Drive, Slack and a Zoom connector, and does not enumerate CRM connectors, so this course cannot tell you what yours can do. Write down what it can read and write, and put that line in your data-boundary.md from Module 2.
Worth knowing: a large export consumes allowance quickly, because complex multi-step work uses materially more than chat, out of a pool shared with chat and Claude Code. One batched monthly pass is cheaper than a live connection you poke at all week. And no trigger fires when a record changes, so a hygiene pass is a cadence whatever you connect. Module 8 designs the cadence.
Try this now
Point Cowork at your revenue-workspace/ from Module 2, not at Downloads. Downloads also holds a compensation spreadsheet, and the directory boundary is the real boundary.
Copy accounts.csv, contacts.csv, opportunities.csv from your pack's data folder into 10-data/. Nothing else goes there.
Before any analysis, paste this and read the answer.
Read every CSV in 10-data/. For each file, report: filename, row count,
column names in order, and the first two data rows verbatim.
Do not clean, interpret or summarise anything. Do not write any file.
What you should see:accounts.csv 308 rows, contacts.csv 911, opportunities.csv 528. You asked for an inventory first on purpose: proof the agent sees what you think it sees. If a count differs from what your CRM export tool reported, stop. You have an encoding problem, a quoting problem or the wrong file, and every later number inherits it. Thirty seconds, and the cheapest tripwire in this course.
What you keep
Nothing yet. Two decisions, written at the top of 40-reference/field-policy.md, which you start next lesson:
Source: file export, refreshed monthly, in 10-data/. Connector status: not authorised.
Never delegated: the survivor decision on a duplicate merge. Reason: no source of truth exists in the export, so the decision cannot be checked.
That second line belongs in your Module 3 delegation-policy.md too.
Check yourself
Your VP asks why the agent cannot just fix the duplicates it can obviously see. Answer in two sentences, using the fit test.
A colleague says the export route is slower because you re-export every month. Name what the export buys you that the connector does not, and say whether you would trade it.
You are handed a connector already authorised with write scope. What do you do first, and what do you write down?
Where this goes next
L5.2 turns "count the defects" into a brief that gives the same answer every month, which means naming the defect classes first.
Part A of the exercise sheet in labs/LAB-05 of your practice pack (start with its README) runs the full pass on this dataset. TL-OPS-2, in your pack under labs/track-labs, is where a RevOps learner takes the workflow to a controlled write with a tested reversal.
Module 5 · The truth layer, CRM hygiene at volume
M5 lesson / about 13 min
Do it
Name the defect classes before you count anything
Do this nowCount one defect class by hand60s
Filter opportunities.csv to the six open stages.
Within that, count the rows whose close_date is earlier than 2026-08-02.
Write the count, the two fields you used, and the population you counted over.
You know it worked when: You get 14, over a population of 524 distinct opportunity_id values in a file of 528 rows.
What you type
OUTCOME
Produce a defect audit of the opportunity export against the definitions in
40-reference/field-policy.md, sections D1 to D10.
INPUTS
10-data/opportunities.csv (528 rows) and 10-data/accounts.csv (308 rows).
accounts.csv is used only to resolve account_id and to read accounts.currency.
OUTPUT SHAPE
30-output/hygiene/audit-2026-08-06.md
One markdown table, one row per defect class, columns exactly:
defect_class | definition_ref | count | pct_of_population | example_ids
Sort descending by count. Percentages to one decimal place.
Below the table, one line: "Distinct opportunity records failing at least one
check: N of 524 (X%)."
CONSTRAINTS
Population is the 524 distinct opportunity_id values unless a definition says
otherwise. Run date is 2026-08-02, the dataset's own today. Change no input file
and propose no fixes here.
DEFINITION OF DONE
Every class D1 to D10 appears, including classes with a count of zero, and every
example_ids cell holds exactly three real ids.
EXCEPTIONS RULE
Any record you cannot classify goes to 30-output/hygiene/exceptions.csv with
record_id, reason and definition_ref. Do not guess.
What comes back
30-output/hygiene/audit-2026-08-06.md written.
| defect_class | ref | count | pct | example_ids |
| open opportunity, close date in the past | D4 | 14 | 2.7% | OPP-006738, OPP-006894, OPP-006905 |
| owner name variant | D7 | 6 | 1.1% | OPP-006696, OPP-006746, OPP-006795 |
| closed, close date blank | D5 | 5 | 1.0% | OPP-006728, OPP-007099, OPP-007174 |
| closed lost, loss reason blank | D6 | 5 | 1.0% | OPP-006670, OPP-006750, OPP-006884 |
| duplicated primary key | D1 | 4 (8 rows) | 0.8% | OPP-006790, OPP-006886, OPP-007099 |
| owner blank | D2 | 4 | 0.8% | OPP-006774, OPP-006841, OPP-007062 |
| currency contradicts account | D8 | 4 | 0.8% | OPP-006697, OPP-006729, OPP-007089 |
| fiscal quarter disagrees with close date | D9 | 4 | 0.8% | OPP-006764, OPP-006773, OPP-006805 |
| amount at or below zero | D10 | 3 | 0.6% | OPP-006774, OPP-006833, OPP-007117 |
| account id does not resolve | D3 | 2 | 0.4% | OPP-006774, OPP-006894 |
Distinct opportunity records failing at least one check: 46 of 524 (8.8%).
Notice: The class counts add to 51 and only 46 records are involved, because OPP-006774 is in three classes at once, so the line under the table is what stops you overstating the problem by eleven percent.
The shipped Talvera data carries 320 defect records across 23 classes, unevenly: accounts and opportunities are near 13 percent defective while pipeline snapshots are 0.07 percent, and the six classes that do not announce themselves sit in only four sources.
You ask for "a data quality report on the CRM export". You get four paragraphs saying the data has some issues. You cannot act on it, compare it to next month, or tell whether it read every row. It is not wrong. It is not a measurement either.
The fix is not a better model or a longer prompt. You never said what a defect is, so the agent picked. This lesson does that job, once, in a file.
What you will be able to do
Write a defect definition that is falsifiable, and spot one that is not.
Fix the output shape so this month's audit is comparable to last month's.
Produce an audit table with counts, percentages and named example IDs, from a brief you can re-run.
A defect definition is a test, not an adjective
Two definitions of the same class:
Stale records. Records that have not been touched in a long time and look abandoned.
Stale opportunity. An opportunity whose stage is one of the six open stages and whose close_date is earlier than the run date. Counted per opportunity record. Population: all rows in opportunities.csv.
The first cannot be wrong, so it cannot be right either. Two people running it get two answers. The second can be checked by anybody with the file, and argued with. That makes it useful.
A usable definition has four parts:
The fields it reads, by their real column names.
The comparison, with any threshold stated as a value you chose, not a vibe.
The population: "12 records missing an owner" means something different over 528 rows than over 60.
The unit of count: records, or rows. They differ the moment your file has a duplicated primary key. This dataset does.
Write them in 40-reference/field-policy.md, the policy file you create in this lesson, not in the chat window. Cowork carries no memory between sessions. A definition that lives only in a conversation is gone next month, and the next audit silently uses a different one. Rules in a file in the working folder are the memory.
The output shape is half the brief
Say exactly what the output looks like, or you get a different shape every month and can never draw a trend. The shape for this audit:
Column
Why it is there
defect_class
The name from the field policy. Not free text.
definition_ref
The heading in field-policy.md that defines it. Makes the audit auditable.
count
Records failing the test.
pct_of_population
Count over the stated population, one decimal place.
example_ids
Exactly three record IDs, so a reviewer can spot-check in a minute.
population
The row count the percentage is over. Printed, not implied.
Three examples, not all, is deliberate. The audit is a summary a manager reads; the full list belongs in proposed-changes.csv and exceptions.csv, two files your pass creates in 30-output/hygiene/. An audit nobody reads has changed nothing.
Ask for defects by name
Do not ask for "data quality issues". Name the classes. This taxonomy covers most CRM exports, each class definable against a real column:
missing required fields
invalid values against a picklist
stale records by last-modified age
orphaned records: no owner, or a parent key that does not resolve
impossible dates: an open record with a close date in the past, a created date after a modified date
inconsistent formatting: country, industry, currency, owner name
duplicate candidates (its own lesson, L5.3)
You know which matter in your CRM. That part you cannot delegate, and it takes about twenty minutes. This is why specification cost is a fit-test factor.
See it run
The real brief, run against opportunities.csv in the shipped dataset, in the seven-part shape from Module 4:
OUTCOME
Produce a defect audit of the opportunity export against the definitions in
40-reference/field-policy.md.
INPUTS
10-data/opportunities.csv (528 rows) and 10-data/accounts.csv (308 rows).
accounts.csv is used only to resolve account_id and to read accounts.currency.
Definitions: 40-reference/field-policy.md, sections D1 to D10.
OUTPUT SHAPE
30-output/hygiene/audit-2026-08-06.md
One markdown table, one row per defect class, columns exactly:
defect_class | definition_ref | count | pct_of_population | example_ids | population
Sort descending by count. Percentages to one decimal place.
Below the table, one line: "Distinct opportunity records failing at least one
check: N of 524 (X%)."
CONSTRAINTS
Population for every class is the 524 distinct opportunity_id values unless the
definition says otherwise. Run date is 2026-08-02, which is the dataset's own
"today" (DATA-DICTIONARY.md). Do not clean, modify or rewrite any input file.
Do not propose fixes in this output.
DEFINITION OF DONE
Every class in field-policy.md D1 to D10 appears in the table, including classes
with a count of zero. Every example_ids cell holds exactly three real IDs.
STOPPING CONDITION
Stop after writing the audit file. Do not continue into duplicate detection.
EXCEPTIONS RULE
Any record you cannot classify against a definition goes to
30-output/hygiene/exceptions.csv with columns record_id, reason, definition_ref.
Do not guess.
The result, from the shipped data. Population: the 524 distinct opportunity_id values in a file of 528 rows.
defect_class
ref
count
pct
example_ids
open opportunity, close date in the past
D4
14
2.7%
OPP-006738, OPP-006894, OPP-006905
owner name variant
D7
6
1.1%
OPP-006696, OPP-006746, OPP-006795
closed, close date blank
D5
5
1.0%
OPP-006728, OPP-007099, OPP-007174
closed lost, loss reason blank
D6
5
1.0%
OPP-006670, OPP-006750, OPP-006884
duplicated primary key
D1
4 (on 8 rows)
0.8%
OPP-006790, OPP-006886, OPP-007099
owner blank
D2
4
0.8%
OPP-006774, OPP-006841, OPP-007062
currency contradicts account
D8
4
0.8%
OPP-006697, OPP-006729, OPP-007089
fiscal quarter disagrees with close date
D9
4
0.8%
OPP-006764, OPP-006773, OPP-006805
amount at or below zero
D10
3
0.6%
OPP-006774, OPP-006833, OPP-007117
account id does not resolve
D3
2
0.4%
OPP-006774, OPP-006894
Distinct opportunity records failing at least one check: 46 of 524 (8.8%).
Two lines matter most. A careless run leaves them out.
The class counts add to 51, but only 46 records are involved. Four records fail two checks, one fails three: OPP-006774 has a blank owner, an unresolvable account, and a negative amount. Report "51 defective opportunities" and you overstate by 11%. The first checker will find it. Hence the distinct-record line, exactly what the module quiz in the interactive app asks you to diagnose.
The duplicated primary key class counts 4 and touches 8 rows. That distinction is L5.4's failure section, and where your pipeline total goes wrong.
Every ID above is documented in answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab.
Try this now: three definitions, one number back
About eight minutes, one real thing end to end. You are the RevOps analyst at Talvera. Sales leadership wants "how much of the open pipeline is not real" by Monday. You have the export and no definitions, because nobody ever wrote any.
Try this now
Open opportunities.csv in your pack's data folder and read the columns, or the schema in DATA-DICTIONARY.md next to it. The six open stages are Qualification, Discovery, Solution Review, Technical Validation, Proposal, Negotiation.
Create 40-reference/field-policy.md from templates/M5/field-policy.md in your practice pack.
Write three definitions, each with all four parts: an open opportunity whose close date has passed; an opportunity with no owner; an opportunity whose currency contradicts its account's billing currency.
Fill the gaps in the starter brief and run it.
Compare your counts to the reference below. If one differs, look at the definition first, not the run.
Starter brief (fill the gaps marked <<...>>):
OUTCOME
Count the records failing each of the three defect definitions in
40-reference/field-policy.md.
INPUTS
10-data/opportunities.csv
<<which second file do you need, and what exactly do you need from it?>>
OUTPUT SHAPE
A markdown table printed in the chat, not written to a file. Columns:
defect_class | count | pct_of_population | example_ids (exactly 3)
CONSTRAINTS
Population is <<state the population and say whether it is rows or distinct
opportunity_id values>>.
Today's date for any date comparison is <<which date, and why that one?>>.
Do not modify any input file.
DEFINITION OF DONE
Three rows. Every count reproducible from the file.
EXCEPTIONS RULE
<<what should happen to a record you cannot classify?>>
Stuck? Open the hints in order.
Hint 1 (Concept)
Two of your three definitions can be answered from one file. The third cannot: one class is only visible when two files disagree. Single-row defects are the easy half of data quality. Defects in the relationship between two rows are the half that survives a careless run.
Hint 2 (Move)
Join opportunities.account_id to accounts.account_id and compare opportunities.currency against accounts.currency row by row. Then decide what happens to the two opportunity rows whose account_id resolves to no account: they cannot pass this test and cannot fail it. Your exceptions rule answers that.
Hint 3 (Shape)
A correct answer has three counts in single figures and low double figures, all under 3% of the population, plus one thing not in the table: at least one record routed to exceptions because its account did not resolve. If your exceptions output is empty, re-read your currency definition. An empty exceptions file on messy data is a warning sign, not a clean bill of health.
What you should see:
defect_class
count
pct_of_population
example_ids
open opportunity, close date in the past
14
2.7%
OPP-006738, OPP-006894, OPP-006905
owner blank
4
0.8%
OPP-006774, OPP-006841, OPP-007062
currency contradicts account
4
0.8%
OPP-006697, OPP-006729, OPP-007089
Population: 524 distinct opportunity_id values. Run date 2026-08-02.
Exceptions: OPP-006774 and OPP-006894, reason "account_id does not resolve in accounts.csv, so the currency test cannot be applied".
Your answer can legitimately differ. A population of 528 rows instead of 524 records gives slightly lower percentages: fine if the audit says which you used. Excluding closed opportunities from "owner blank" gives a count below 4: also fine, since a blank owner on a 2023 closed deal is a different problem from one on live pipeline. Not fine: a count you cannot reproduce, or a population you did not state.
If "open, close date in the past" came back far above 14, check whether you included closed stages. If it came back 0, check whether it used your machine's today instead of the dataset's 2026-08-02.
Worth knowing: the three counts took the agent seconds. The twenty minutes of definition writing is the part still worth something next month.
What you keep
40-reference/field-policy.md, your first defect definitions. The artifact of the module. It survives your CRM, your job and this course.
20-briefs/crm-hygiene-audit.brief.md, started from the brief above (a seed is in templates/M5/crm-hygiene-audit.brief.md of your practice pack).
30-output/hygiene/audit-YYYY-MM-DD.md, the first dated audit.
Check yourself
Which of these is not usable as a defect definition, and rewrite it so it is: (a) accounts with a blank owner; (b) opportunities where amount is at or below zero; (c) records that look wrong; (d) contacts whose email contains no @.
Your audit says 51 defective opportunities. A colleague counts distinct IDs in your proposed-changes file and gets 46. Who is wrong, and what does the audit need to add?
You inherit an audit with counts and percentages but no population column and no definition_ref. Name the two questions you now cannot answer.
Where this goes next
L5.3 takes on duplicates, the class where an agent will produce confident nonsense, and where the control is a written rule rather than a better model.
Module 5 · The truth layer, CRM hygiene at volume
M5 lesson / about 11 min
Failure mode
Match rules with bands, and the merge that costs you a customer
Do this nowCompare the Follifoot pair60s
Open accounts.csv and put ACC-02839 and ACC-03086 side by side.
Read country, employee_count, lifecycle_status and arr_usd on each.
Decide whether they are the same company.
You know it worked when: Germany against Sweden, 1452 employees against 132, Churned against Customer, arr_usd 0.0 against 27000.0, on two names that differ by one letter.
What you type
OUTCOME
Produce a banded duplicate candidate list for accounts.
INPUTS
10-data/accounts.csv (308 rows). Rules: 40-reference/field-policy.md sections
R1, R2 and R3, including the normalisation list (trim, collapse double spaces,
lowercase, strip legal suffixes Inc Ltd Limited PLC GmbH AG AB SA SAS SL NV BV
LLC Corp, strip punctuation and ampersands).
OUTPUT SHAPE
30-output/hygiene/duplicates.csv, columns exactly:
account_a, account_b, rule, band, name_similarity, domain_match, country_match,
lifecycle_a, lifecycle_b, arr_usd_a, arr_usd_b, evidence
CONSTRAINTS
Every row's evidence cell names the fields compared and their values.
Do not merge anything. Do not modify accounts.csv.
DEFINITION OF DONE
Every candidate pair appears exactly once, A before B by account_id, and no row
has an empty rule or evidence cell.
What comes back
30-output/hygiene/duplicates.csv written.
account_a,account_b,rule,band,name_similarity,domain_match,country_match
ACC-02947,ACC-03047,R1 exact-domain-plus-name,auto,1.00,TRUE,TRUE
ACC-02839,ACC-03086,R1 name-similarity,auto,0.94,FALSE,FALSE
ACC-02890,ACC-02988,R2 name-only,review,0.97,FALSE,TRUE
evidence, ACC-02839 vs ACC-03086:
names identical after normalisation (suffixes AG and AB both stripped)
domains follifoot-hospitality.example.com and follifoot-hospitality.example.org
countries Germany and Sweden
employee_count 1452 and 132
lifecycle Churned and Customer, arr_usd 0.0 and 27000.0
Notice: The second row is the failure: normalisation stripped AG and AB, the only characters separating two companies, and a merge accepted in a bulk review removes a live customer carrying 27,000 USD.
The R3 veto runs before any similarity score, so a pair that differs on country, lifecycle or open pipeline goes to review however similar the names are.
Somebody in your company is going to run a dedupe. It might be you. When it goes wrong, it is not quiet: two real customers become one record, support cannot find the second one, the renewal sits on an account nobody owns, and the fix is a restore from a backup you have to ask for.
Duplicate detection is the one place in a hygiene pass where an agent hands you confident, well-formatted, wrong answers. Not because it is careless. Because "same company" is a judgment, and you did not write down what you meant.
This lesson runs on the accounts.csv you copied from your pack's data folder into 10-data/ in L5.1.
What you will be able to do
Write a match rule with bands: some pairs safe to auto-merge, the rest to a human with evidence attached.
Tell a duplicate apart from two records that merely resemble each other, and name the field that decides it.
Reject any duplicate output whose rows do not say why they matched.
Ban the unexplained score
You will be offered output like this:
account_a
account_b
match_score
ACC-02947
ACC-03047
0.96
ACC-02839
ACC-03086
0.94
Unusable, and not because 0.96 is wrong. A reviewer cannot tell which pair is safe, what the score is made of, or how to stop a rejected pair coming back next month. A number with no components is an opinion wearing a lab coat.
The rule for this module, and the exercise sheet in labs/LAB-05 of your practice pack checks for it (start with its README): if the output cannot state why two records matched, it is not reviewable, and therefore not usable.
Same output, made reviewable:
account_a
account_b
rule
evidence
band
ACC-02947
ACC-03047
R1 exact-domain-plus-name
domain halloway-logistics.example.com identical; names identical after casing normalisation; country Germany both
auto
ACC-02839
ACC-03086
R3 name-only-cross-country
names differ only in legal suffix (AG, AB); domains follifoot-hospitality.example.com and .example.org differ; countries Germany and Sweden differ
review, never auto
Same amount of agent work. The difference is entirely in your brief.
Bands, and the one that is not a band at all
Write the rule in 40-reference/field-policy.md, the file you started in L5.2, under a heading the brief can reference. The thresholds are yours. A starting shape that works on this dataset:
R1 auto band
Normalised domain equal
AND normalised name similarity >= 0.90
AND country equal
-> propose merge, still draft-only, still reviewed in bulk
R2 review band
Domain equal, names not similar
OR names similar >= 0.90, domains not equal
-> one human decision per pair, with evidence
R3 never auto, regardless of any score
Country differs
OR one record is a Customer and the other is Churned
OR either record has open pipeline
-> human decision, and the survivor question goes with it
Normalisation, stated so it is reproducible: trim whitespace, collapse internal double spaces, lowercase, strip the legal suffix set you name (Inc, Ltd, Limited, PLC, GmbH, AG, AB, SA, SAS, SL, NV, BV, LLC, Corp), strip punctuation and ampersands.
Three things worth saying out loud:
R3 is not a third band, it is a veto. It runs first and ignores the similarity score. Every veto in it exists because something happened to somebody: merging across countries destroys a legal entity distinction, merging a customer into a churned record destroys revenue history, merging a record with open pipeline moves a live deal to an account the rep does not own.
The auto band is still draft-only. "Auto" means "accept in bulk after a scan", not "apply without a human". Nothing in this module writes to a system of record. Promotion is L5.4.
You chose 0.90, so be able to say why. Write the number and one line of reasoning in the file. Next month, when a pair lands just under it, you will want to know what you were thinking.
Try this now
Paste the brief section that produces the duplicate list:
OUTCOME
Produce a banded duplicate candidate list for accounts.
INPUTS
10-data/accounts.csv (308 rows). Rules: 40-reference/field-policy.md sections
R1, R2, R3 including the normalisation list.
OUTPUT SHAPE
30-output/hygiene/duplicates.csv, columns exactly:
account_a, account_b, rule, evidence, band, name_similarity,
domain_match, country_match, lifecycle_a, lifecycle_b, arr_usd_a, arr_usd_b
CONSTRAINTS
Apply R3 first. A pair that trips any R3 condition is band=review regardless of
similarity, and the evidence column states which condition fired.
Every row's evidence column names the fields compared and their values.
Do not merge anything. Do not modify accounts.csv.
DEFINITION OF DONE
Every candidate pair appears exactly once (A before B by account_id).
No row has an empty rule or evidence cell.
EXCEPTIONS RULE
A record with a blank domain and a blank owner cannot be scored. Write it to
30-output/hygiene/exceptions.csv with reason "insufficient fields to match".
What you should see: a candidate list you can check against the key afterwards. answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab, documents 8 account duplicates under a name variant, each with its partner named, for example ACC-02947 against ACC-03047 (Halloway Logistics AG, upper-case variant) and ACC-02890 against ACC-02988 (Highmoor Retail Inc, punctuation variant).
Note what the second pair does to a lazy rule: ACC-02890 is highmoor-retail.example.net and ACC-02988 is highmoor-retail.example.com. Same company, different domain string. A rule that requires domain equality misses it entirely. That pair belongs in R2, gets a human, and the human takes eight seconds: the evidence column already says employee count 7956 on both and owner Bettina Costa on both.
Tell these apart: a duplicate, and two records that resemble each other
This is the distinction that costs money when it is got wrong.
A duplicate
ACC-02947
ACC-03047
account_name
HALLOWAY LOGISTICS AG
Halloway Logistics AG
domain
halloway-logistics.example.com
halloway-logistics.example.com
country
Germany
Germany
employee_count
712
712
segment
Mid-Market
Mid-Market
owner
Bettina Farrow
Bettina Farrow
lifecycle_status
Prospect
Prospect
created_date
2025-03-15
2024-09-19
Everything that identifies the company is identical. The only differences are casing and a creation date six months apart: the fingerprint of one company entered twice, most likely by a form fill that did not check. PLANTED-ISSUES.md lists ACC-02947 in the duplicate-accounts class, partner ACC-03047. Band: auto. Action: propose merge, survivor is the older record unless the reviewer says otherwise.
Two records that resemble each other
ACC-02839
ACC-03086
account_name
Follifoot Hospitality Group AG
Follifoot Hospitality Group AB
domain
follifoot-hospitality.example.com
follifoot-hospitality.example.org
country
Germany
Sweden
employee_count
1452
132
segment
Mid-Market
SMB
owner
Aisling Calderon
Franka Mertens
csm
Tomasz Jonker
Tomasz Jonker
lifecycle_status
Churned
Customer
arr_usd
0.0
27000.0
Name similarity here is higher than in the first example on most string measures: the names differ by one character. Same domain stem. Same CSM. A name-similarity rule with a 0.90 threshold merges these two without hesitating. And merging them deletes a live customer worth 27,000 into a churned record. The renewal date disappears, the CSM's book is short one account, and the first person to notice is the customer.
What separates them is not similarity, it is identity evidence: different country, different top-level domain, an eleven-fold difference in employee count, one record live and one churned. Any one of those is enough to veto. That is what R3 is for.
Worth knowing: neither row appears in the duplicate-accounts class in PLANTED-ISSUES.md; that class documents 8 pairs, all genuine duplicates, and this is not one of them. The Follifoot pair is real data that behaves exactly like this trap, so practice on it, but it is not a documented planted pair and the key will not confirm it.
What the feedback looks like
What a learner hears after submitting a duplicate list with ACC-02839, ACC-03086 in the auto band:
Your list put ACC-02839 and ACC-03086 in the auto band on a name similarity of 0.94, above your stated 0.90 threshold. The rule fired correctly. The rule is the problem.
These are two different companies. ACC-02839 is in Germany with 1,452 employees and is churned. ACC-03086 is in Sweden with 132 employees, is a live customer, and carries 27,000 USD of ARR. The names differ only in the legal suffix, AG against AB, precisely the part your normalisation strips before comparing. It removed the only piece of the name that distinguished the two entities.
Accepted in a bulk review, this merge removes a paying customer from your account list. That is the failure the band structure exists to prevent: a safety failure, not just a wrong count.
What would have caught it. Not a higher threshold: 0.94 still clears 0.95 on some measures, and raising the bar loses ACC-02890 against ACC-02988, a real duplicate whose domains differ. What catches it is a veto that runs before the score: country_match = false sends the pair to review no matter the similarity. The same veto catches the lifecycle mismatch, the second independent reason this pair must not merge.
What to change. Add R3 to field-policy.md and re-run. Then check for the opposite error: a real duplicate R3 pushed into review unnecessarily. If R3 sends half your candidates to review, it is too broad and you have moved the cost, not removed it. Report the trade in your run log.
What you keep
40-reference/field-policy.md, now with R1, R2 and R3, your thresholds and your normalisation list.
30-output/hygiene/duplicates.csv, banded, with rule and evidence on every row.
Check yourself
Two accounts share a domain exactly, have completely different names, and one has open pipeline. Which band, and which condition decided it?
A colleague proposes raising the similarity threshold from 0.90 to 0.97 "to be safe". Name the real duplicate class in this dataset that change would start missing, and say what you would do instead.
Your duplicate list has 40 rows and 38 are in the review band. Is the rule too tight or too loose, and what does that cost the reviewer?
Where this goes next
L5.4 is where the pass gets checked, where the naive roll-up gets caught, and where you write down the conditions under which any of this would ever write to the CRM.
Module 5 · The truth layer, CRM hygiene at volume
M5 lesson / about 14 min
Failure mode
Reconcile, then sample, then earn the write
Do this nowSum Closed Won the wrong way60s
Filter opportunities.csv to stage Closed Won.
Add up the amount column without deduplicating and without converting currency.
Write the number down and keep it.
You know it worked when: You get 16,830,000, a clean, well-formatted number that no system in the company holds.
What you type
Read opportunities.csv and accounts.csv in this folder.
Produce the total Closed Won value in USD.
Before any aggregation, deduplicate on opportunity_id. Where two rows share an
id and their amounts disagree, do not pick one: write both to
30-output/hygiene/exceptions.csv with the reason "duplicate primary key, amounts
disagree", and use the first row for the headline figure.
Treat accounts.currency as the authority over opportunities.currency. Convert at
the fixed rates in DATA-DICTIONARY.md: EUR 1.08, GBP 1.27, USD 1.00.
Report four figures: the raw sum, the deduplicated sum, the converted total on
opportunities.currency, and the converted total on accounts.currency. Say which
one you would defend and why.
What comes back
Raw sum of amount, 240 Closed Won rows, currencies mixed: 16,830,000
Deduplicated on opportunity_id, 236 records, currencies mixed: 16,679,000
Converted, opportunities.currency as authority: 17,980,560 USD
Converted, accounts.currency as authority: 17,995,680 USD
exceptions.csv, 4 rows:
OPP-006790 9,000 and 11,000 EUR
OPP-006886 57,000 and 71,000 USD
OPP-007099 69,000 and 55,000 USD
OPP-007110 15,000 and 14,000 USD
Four rows carry a currency label their account contradicts (OPP-006697,
OPP-006729, OPP-007089, OPP-007326). They are the 15,120 USD between the last
two figures.
Notice: Four plausible totals out of one file, and the only defensible one names which currency column it trusted and lists the four collisions it refused to resolve.
Reconciliation looks at the whole population and catches drops and double counts, sampling looks at a few rows and catches rows that are present and wrong, and a tripwire checks a row whose answer you already knew.A row the brief cannot resolve has three destinations, and the middle one is credited as a half catch because declaring what could not be resolved is the only route that leaves a visible trace.Total Closed Won on the shipped data is 16,830,000 raw, 16,679,000 after deduplicating on the primary key, 17,980,560 after currency conversion, and 17,995,680 once the account is treated as the authority for currency.
Your audit is done, your duplicate list is banded, and the output looks good. That is the dangerous moment. A wrong hygiene pass looks exactly like a right one: same columns, same confident tone, same tidy counts. Then somebody puts your number next to the board pack, the two disagree, and you find out in the meeting.
What you will be able to do
Reconcile an audit against its source so every row is accounted for.
Choose a sample size you will actually run every month, and defend it.
Read the exceptions file as a statement about your brief, not the data.
State the conditions under which this workflow could write to the CRM, with the reversal plan to match.
Reconcile first: everything gets accounted for
Sampling is the check people reach for. It is the second check. The first is free and catches what sampling cannot, because sampling only looks at rows you already know about. Reconciliation has three parts:
Row accounting. Every row in the export is clean, defective in one or more named classes, or written to exceptions. Total the three against the source. In the shipped opportunity file: 528 rows, 524 distinct records, 65 records failing at least one check, 459 clean. The lines must sum to the distinct count. If not, the accounting is wrong before the analysis starts.
Referential accounting. Every ID in your duplicate list and proposed-changes file exists in the source. An agent that invents an ID is rare. One that carries a truncated ID through a join is not.
Cross-source accounting. One number in your output is rebuilt from a different source and shown to match. This catches the silent failures, and it is the subject of the breakage below.
Sample ten rows, on purpose
The temptation is ten rows per class. With ten classes that is a hundred manual checks inside a pass meant to save you time, and the course rule forbids it: if checking costs more than doing it by hand, do not hand it over. The sample that survives a real month:
10 rows in total, allocated across defect classes in proportion to class size, minimum one row for any class with at least one member. A class of 14 gets three rows, a class of 4 gets one.
Plus every auto-band duplicate pair above a stated ARR threshold. You pick the threshold and write it in the field-policy.md you started in L5.2. A merge on a Tier 1 account gets looked at every single time.
When somebody asks why the sample is not bigger: the correct sample size is the one you will actually run every month. A twenty-row sample you skip in December is worth less than a ten-row sample you never skip. Either way, record the result in the run log: ten rows checked, N errors found, which classes.
The exceptions file is feedback on your brief
Anything the pass cannot resolve goes to exceptions.csv with the record ID and the reason. It does not guess. Read it like this:
What you see
What it usually means
Empty, on messy data
The brief gave the agent no way to say "I do not know". Suspect silent guessing.
A handful of rows with distinct reasons
Healthy. This is the file working.
40% of the file, one repeated reason
Your field policy is wrong, not the data. Fix the definition.
In this dataset the honest exceptions on the opportunity pass are OPP-006774 and OPP-006894, whose account_id values resolve to no account, so no test needing the account applies. Two rows out of 528 is what a healthy exceptions file looks like.
The shape that makes review possible
proposed-changes.csv, which your pass writes to 30-output/hygiene/, is the deliverable. It is draft-only: a human reviews it and a human applies it, through your CRM's normal change process.
Columns, exactly:
Column
Why a reviewer needs it
record_id
What is changing
object
accounts, contacts or opportunities
field
Which field
current_value
So they do not have to open the CRM
proposed_value
The change
reason
Plain English, one line
rule
The rule ID from field-policy.md that produced it
confidence_band
auto or review, from L5.3
The rule column makes bulk review possible: a reviewer sorts by rule, reads one example, and accepts or rejects 60 rows at once. Without it they read 60 unrelated decisions, stop after eleven, and your pass never gets applied. That is the difference between a pass a manager approves and one they do not, and the module quiz in the interactive app tests exactly this.
Earning the write, and the reversal plan
You write this as a paper exercise in Part A of the exercise sheet in labs/LAB-05 of your practice pack (start with its README), whatever your role, and the module quiz in the interactive app tests it.
The workflow may be proposed for promotion from draft-only to writes-to-system only when all of these are true:
It has run clean on three consecutive passes, with the run log to show it.
Sampling found no errors in the auto band across those passes.
The reversal path has been tested on a small batch: export the before state, apply, confirm you can restore, write down how long the restore took.
A named approver signs each run. Not a role. A person, per run.
The run log captures what changed, per record, before and after.
Then the part that is not a checklist item. The audit trail for a CRM write initiated through Cowork is the one you build. The Compliance API and Audit Logs do not cover Cowork. Enterprise has an admin dashboard, an Analytics API with per-day usage and per-user activity, RBAC, spend limits, and OpenTelemetry events covering tool and connector calls, files read or modified, and whether an action was approved manually or automatically, all SIEM compatible. That is real telemetry, genuinely useful, and not a compliance audit log. If your organisation needs the second one, this workflow stays draft-only and the import goes through your CRM's normal change process. Say both halves to your security reviewer. They will find out anyway, and better from you.
Cadence, because there is nothing to trigger on
There is no event trigger, no "when a record changes, run this". A hygiene pass is therefore weekly or monthly, chosen by you, and the lag before your pass notices a bad record is a number you accept, not a problem you solve. Large exports also consume allowance quickly, out of a pool shared with chat and Claude Code: a second argument for one batched monthly pass over constant small ones. Module 8 designs the cadence, including what happens the month it does not run.
Try this now
Run the reconciliation and the sample:
OUTCOME
Reconcile the audit against source and produce the verification sample.
INPUTS
10-data/opportunities.csv, 10-data/accounts.csv,
30-output/hygiene/audit-2026-08-06.md, 30-output/hygiene/exceptions.csv
OUTPUT SHAPE
Append to 50-runs/RUN-LOG.md under a dated heading:
1. Row accounting: source rows, distinct records, records with >=1 defect,
clean records, exception records. State whether they reconcile.
2. Referential check: count of IDs in duplicates.csv and proposed-changes.csv
that do not appear in the source. Expected 0.
3. Cross-source check: for account ACC-02858, compare accounts.arr_usd against
the amount on the most recent Closed Won opportunity converted at the fixed
rate for that account's currency. Report both numbers.
4. Sample: 10 records, allocated across defect classes in proportion to class
size, minimum 1 per non-empty class. Print record_id, class, and the field
values a human needs to verify it.
CONSTRAINTS
Do not fix anything found here. Report only.
What you should see: the cross-source check on ACC-02858 (Vireo Retail PLC) resolves. arr_usd is 121,920, the account bills in GBP, and the most recent Closed Won renewal OPP-006984 is 96,000 GBP. 96,000 times 1.27 is 121,920. They tie exactly. That is your tripwire: one account, about forty seconds, and it fails loudly the moment somebody sums an amount column without converting it.
When this goes wrong
Break it yourself
Ask for the simplest thing a leader ever asks for. No deduplication instruction, no currency instruction. Run it.
From 10-data/opportunities.csv, give me total Closed Won by fiscal year and the
total open pipeline.
What comes back
You get 16,830,000 for total Closed Won. A clean number, nicely formatted, and wrong in two directions at once.
Direction one: it double counts.opportunity_id is meant to be the primary key of opportunities.csv and is not unique in the file (DATA-DICTIONARY.md in your pack's data folder says so explicitly). Four IDs appear on two rows each, with different amounts: OPP-006790 (9,000 and 11,000 EUR), OPP-006886 (57,000 and 71,000 USD), OPP-007099 (69,000 and 55,000 USD), OPP-007110 (15,000 and 14,000 USD). All four are Closed Won, so all four are counted twice. Deduplicating on the key brings the raw total to 16,679,000.
Worse than the size of the error: you cannot tell which row is right. Keep the lower amount of each pair and the converted total is 17,965,560 USD. Keep the higher and it is 17,996,720 USD. The 31,160 USD gap depends entirely on a choice nobody made deliberately. This is why the guard sends collisions to exceptions rather than picking a row.
Direction two: it adds three currencies together.opportunities.amount is in the account's local billing currency and accounts.arr_usd is in USD (DATA-DICTIONARY.md). Summing the amount column treats a EUR deal and a USD deal as the same unit. Converting properly, at the fixed rates in the data dictionary (EUR 1.08, GBP 1.27), takes the deduplicated Closed Won total to 17,980,560 USD.
And a third thing, hiding inside the second. Four opportunities carry a currency label that contradicts their account: OPP-006697, OPP-006729, OPP-007089 and OPP-007326, all labelled USD on accounts that bill in EUR (answer-keys/PLANTED-ISSUES.md in your pack, class SUBTLE currency-mismatch, which you open only after attempting the lab). Even a careful analyst converts these four the wrong way. Using the account as the authority moves the Closed Won total again, to 17,995,680 USD.
Nothing in any of those outputs looks broken. Every one is a plausible number somebody would put in a board pack.
The check that catches it
Take OPP-007089, on account ACC-03089 (Quarrybank Hospitality Group SL, Spain, billing EUR).
The opportunity row says 25,000 USD.
The account's arr_usd says 27,000.
25,000 times the EUR rate of 1.08 is 27,000.
And the renewal on the same account, OPP-006747, records the same 25,000 labelled EUR.
The account's own USD figure only reconciles if the deal is read as EUR. The currency label on that row is wrong, and the reconciliation says so without anybody noticing a suspicious-looking row. That is the difference between a check and a hunch.
The guard
Three lines in the brief, permanently:
Deduplicate on the primary key before any aggregation, and write every collision to exceptions.csv with both amounts and the reason duplicate primary key, amounts disagree. Do not pick a row silently.
Treat accounts.currency as the authority over opportunities.currency, convert with the rates named in the brief, and flag every row where the two disagree as its own defect class.
Put both checks inside the reconciliation section, so they run every month rather than the month you remembered.
Then re-run and confirm the three numbers move the way you predicted. A guard you did not watch fire is not a guard.
What you keep
30-output/hygiene/proposed-changes.csv and 30-output/hygiene/exceptions.csv.
A 50-runs/RUN-LOG.md entry with the reconciliation result, the sample result and the ten record IDs you checked.
In 40-reference/field-policy.md: the promotion checklist, the named approver, the reversal procedure and the ARR threshold above which every auto-band merge is reviewed individually.
20-briefs/crm-hygiene-audit.brief.md, complete, with the three guards in it.
Check yourself
Your audit reconciles perfectly: every class count checks out, the sample of ten found no errors, the totals agree with each other. Name one thing that could still be badly wrong, and the check that would find it.
A colleague proposes promoting the workflow to direct CRM writes after one clean run, because the sample found nothing. List the promotion conditions not yet met.
Your exceptions file has 214 rows out of 528, all with the reason "cannot determine required fields for this record type". What do you change, and where?
Your security reviewer asks: "if this ever writes to the CRM, what is the audit trail?" Answer in three sentences, including what Enterprise telemetry does and does not give them.
Where this goes next
The exercise sheet in labs/LAB-05 of your practice pack (start with its README) runs the whole pass on the shipped dataset, then on your own export behind the redaction gate, and scores how much of the planted mess you caught.
M6 takes the data you can now trust and asks what changed this week.
Module 5 · The truth layer, CRM hygiene at volume
Lab / ? min / draft-only
Lab
The hygiene pass that a manager will approve
LAB-05: The hygiene pass that a manager will approve
The situation
The forecast is wrong and everybody knows it. Use the shipped Talvera export, because it is the file in
front of you and its defects are documented with exact row identifiers in
course/labs/data/PLANTED-ISSUES.md, key 20260801:
14 open opportunities carry a close date in the past.
8 account records duplicate an existing account under a name variant, out of 308 accounts.
4 opportunity IDs appear on two rows each with different amounts, so any naive sum double counts.
4 opportunities carry a currency label that contradicts the account's billing currency: all four
are labelled USD on accounts that bill in EUR.
Owners are blank on 3 accounts and 4 opportunities, and the same person is spelled several ways on
others: Aisling Calderon also appears as Aisling Calderon and Aisling Calderon , and
Bettina Costa also appears as bettina costa. Group by owner and one rep becomes three rows.
Every number in that paragraph is a count from the shipped synthetic dataset, not a claim about your
CRM. Yours will be worse in different ways, which is what Part B is for.
Fixing this by hand is a multi-day job you are not being given. And the tempting shortcut, letting an
agent write corrections straight into the CRM, is the version that ends with two real customers merged
into one record and a support team that cannot find either.
What you will produce
File
What it is
40-reference/field-policy.md
Required fields, valid values, staleness thresholds, and the duplicate match rule with its bands. A RevOps document that outlives the course
20-briefs/crm-hygiene-audit.brief.md
The seven-part brief, rerunnable monthly
30-output/hygiene/audit-2026-08-02.md
The defect report: one row per class with count, percentage and three example IDs
30-output/hygiene/duplicates.csv
Banded candidate pairs with the rule that matched each one
30-output/hygiene/completeness-by-owner.md
Completeness by owner and by segment, turned into named actions
30-output/hygiene/proposed-changes.csv
Draft-only. record ID, field, current value, proposed value, reason, rule, confidence band
30-output/hygiene/exceptions.csv
Everything the run refused to guess about
40-reference/promotion-checklist.md
The conditions for promoting this workflow to writes-to-system, and the reversal plan that promotion requires
Outcome accounting
Manual baseline: 6 hours (360 minutes), conservative end of a measured range.
How measured: the lab author hand-audited a 200-row slice (100 accounts, 50 contacts,
50 opportunities) against the eleven declared defect classes in a spreadsheet. Timed: 68 minutes.
Full declared scope is 308 + 911 + 528 = 1,747 rows.
Linear extrapolation: 68 x (1747 / 200) = 594 minutes, about 10 hours.
We do not claim 10 hours, for two reasons that pull in opposite directions and are both stated:
- Duplicate detection is pairwise, so it gets HARDER than linear as the file grows.
- The per-row cost of the field checks FALLS once the spreadsheet formulas exist, which is most
of the first 68 minutes.
The honest answer is a range of roughly 6 to 10 hours. We claim the bottom of it.
Target with Cowork: 55 minutes on run one, observed while building this lab.
41 minutes of running, of which about 18 was writing field-policy.md.
14 minutes of verification (see below).
Run two, monthly, with the policy and brief already written: 22 minutes
observed, of which 14 is the same verification.
Verification cost: 14 minutes, counted separately and inside every number above.
Reconciliation about 4 minutes; the 10-row sample about 8; the auto-band
duplicate check about 2.
Net, run one: 360 - 55 = 305 minutes saved.
Net, run two onward: 360 - 22 = 338 minutes saved, monthly.
Your actual: _____ run one _____ verification _____ run two
The number we are NOT claiming: that the output is as good as six hours of your attention. It is
not. It is a proposal file and an exceptions file, and the six hours it replaces produced a cleaned
CRM. The comparison is honest only because the deliverable here is reviewed by a human before
anything is applied, which is what draft-only means.
Blast radius
Draft-only. The deliverable is proposed-changes.csv, a file a human reads and a human imports.
Nothing in this lab writes to a CRM and nothing modifies 10-data/.
Obligation
This lab
Review gate
Full human read of proposed-changes.csv before it leaves the folder, sorted by rule so a reviewer can accept or reject in bulk
Approver
The RevOps lead, as a role
Reversal plan
Do not import the file. Nothing has been written, so there is nothing to unwind
Promotion to writes-to-system is a paper exercise in this lab and is required of every learner on
every track, including those taking M5 at core pass. See part-a.md step 8. Executing a tested
reversal on a live batch is TL-OPS-2.
Work from an export, not a connector, for the first pass. An export is a snapshot you can reconcile
against, it has no write path, and it lets you build the whole workflow at read-only blast radius. A CRM
connector is an MCP connector wired to your tool (FACTS §5.8) and is treated as writes-to-system by
default. No claim is made here about any specific CRM connector's capabilities, because our fact base
does not enumerate them.
Per-class in-scope counts are in SCORING.md §1.1. Two things to know before you start:
date-ordering has 0 in-scope rows in this key version. Every date-ordering row lives in
activities.csv, product_usage.csv or support_tickets.csv, which are not in this lab's file
scope. The class is declared anyway, and if you find date-ordering rows in support_tickets.csv and
report them, they route to adjudication as correct but out of scope: excluded from numerator and
denominator, no penalty, no credit (SCORING.md §4).
Recall, coverage and precision are computed only over the declared scope. You are not marked down
for missing a defect in a file this lab never asked you to open.
You do not hold PLANTED-ISSUES.md while attempting this lab. It is released, sliced to these
eleven classes, after you submit. Requesting it early is permitted and marks this lab's practice
scorecard key-seen, which excludes its numbers from every record (ASSESSMENT-MODEL.md §2.1).
How you verify your own output
MODULES/05 beat 5 sets the sample size and the reason for it, and both matter.
Reconcile first. Every row in the export is accounted for as clean, defective in one or more
named classes, or written to exceptions. The three counts must sum to the source row count, per file.
And the duplicate list must not reference an ID that is absent from the source.
Then sample: 10 rows in total, allocated across defect classes in proportion to class size, plus
every duplicate pair in the auto band above a stated ARR threshold.
Ten rows per class across eleven declared classes would be 110 manual checks inside a 55-minute
Part A, which fails the module's own rule that the check must cost less than the job. Say it out
loud: the correct sample size is the one you will actually run every month.
Two reconciliations that catch what the page cannot show you. A duplicated primary key that makes
a sum double count, and an amount column summed across mixed currencies. Reconcile the opportunity
amount total against accounts.arr_usd, which is stated in USD in the data dictionary while
opportunities.amount is in the account's local currency. The two do not agree and the size of the
disagreement points at the cause.
Files you need
accounts.csv, contacts.csv, opportunities.csv and DATA-DICTIONARY.md in 10-data/talvera/, plus
the workspace, the brief method from LAB-04, and field-policy.md which you write in step 1.
A requirement this lab states rather than claims
MODULES/05 and ASSESSMENT-MODEL.md §2.2.3 both require a near-name control pair: two genuinely
distinct companies with near-identical names in different countries, planted as its own class so a
name-similarity-only rule visibly merges them.
That planted class does not exist in key 20260801. All eight duplicate-accounts pairs are true
duplicates and all eight share a country. Part 3 does not own course/labs/data/, so this remains a
requirement on the dataset owner and this lab does not assert that such a pair ships.
What the lab teaches instead is real and checkable today: the accounts file contains naturally
occurring near-name pairs that share a normalised domain root and sit in different countries and are not
in the key. A rule without the country condition finds them. They are the reason the country condition
exists, and they route through the out-of-key adjudication path rather than through the false-positive
count. Details in solution/notes.md.
What comes next
LAB-06 assumes the opportunity data is trustworthy enough to inspect. A movement analysis over
duplicated accounts produces a reconciliation that cannot be made to balance, which is why M5 comes
first.
Module 5 · The truth layer, CRM hygiene at volume
6 questions / pass at 70%
Checkpoint
Module 5 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
An agent produced this from a learner's own account export.
Illustrative excerpt. The records below are invented for this item and are not rows from the Talvera practice dataset.
These are two different companies: a Dutch operator and an unrelated Swedish one with a similar trading
name. The rule that produced the group was: normalised name similarity above 0.90 goes to the auto-merge
band.
Which component is missing from the match rule?
Hint
Ask whether any value of the single number in this rule would have produced the right answer for both
this pair and the ordinary casing variants you also need to catch. If not, the rule needs another field
rather than another threshold.
A. Incorrect. Raising the threshold trades one error for the other. At 0.97 you separate this pair and you also stop catching the whitespace, punctuation, casing and legal-suffix variants that are the reason you built the list, so the recall collapses while the underlying weakness stays. A rule that can only be made safer by finding less is a rule with a missing dimension.
B. Correct. A one-dimensional similarity rule can only be tuned, not corrected, because there is no threshold at which "Verhoeven Logistics BV" and "Verhoeven Logistics AB" stop resembling each other. What separates them is a second field that carries entity identity: the country differs, the domains differ, and either fact is enough to route the pair to human review. Banding is the mechanism: exact domain match plus high name similarity to the auto band, everything else to review, different country to review regardless.
C. Incorrect. The output already carries a number. Adding a second number produced by the same process does not make the pair reviewable, because a reviewer cannot check a score, only a reason. What makes a duplicate list reviewable is the stated rule that fired and the fields it matched on, which is why the unexplained score is banned in this module.
D. Incorrect. This is true and it is the wrong lesson to take from it. If every auto-merge row has to be reviewed by hand, the auto band does not exist and the workflow has no benefit over reading the file yourself. The band is a promise about which rows are safe without a human, and a merge of two real companies is exactly the failure that promise has to exclude by construction.
Question 2 of 6
A colleague has built the hygiene workflow, run it once, and produced a proposed-changes.csv of 214
rows. They now want to authorise a CRM connector and have the next run apply the changes directly, "to
skip the import step".
Which promotion conditions are not yet met?
Hint
Write out the promotion checklist from memory, then tick off which items the described situation
satisfies. Count how many runs have happened and what evidence exists from them.
A. Correct. Promotion is earned rather than granted, and the checklist has five parts: clean runs on three consecutive passes, sampling that found no errors in the auto band, a reversal path tested on a small batch, a named approver per run, and a run log that captures what changed. One run meets none of them. The second half of your answer is the part that convinces a manager: because the enterprise Compliance API and Audit Logs do not cover Cowork yet (FACTS §5.10, §7.7), the run log and the before-state export are the audit trail rather than a supplement to one.
B. Incorrect. "It looked clean" is the evidence that the promotion conditions exist to replace. You have no sample result, so you do not know the band is clean, you know the file is well formatted. A named approver signing off on a workflow with no tested restore is a person carrying accountability for something they cannot undo.
C. Incorrect. This is a blanket refusal where a staged path exists. A CRM connector is a write path into the system of record and is treated as writes-to-system by default until proven otherwise (FACTS §5.8), and the word doing the work there is "default". The whole point of the promotion checklist is that there is a route from draft-only to writes-to-system, and refusing it entirely means the workflow never improves and people route around you.
D. Incorrect. You picked the most important condition and then made the wrong inference. A reversal plan is not something you add while running live: it has to be written, tested and in the run record *before* the first write, because the first write is when you might need it. The other conditions are also not optional extras, since without three clean passes and a sample you have no basis for believing the changes are right.
Question 3 of 6
A proposed-changes.csv arrives for review, 214 rows, of which the first four are:
Illustrative excerpt. The records below are invented for this item and are not rows from the Talvera practice dataset.
record_id,field,current_value,proposed_value
ACC-1180,country,netherlands,Netherlands
ACC-2244,industry, Retail,Retail
ACC-1902,owner,,D. Marchetti
ACC-2801,account_name,Northgate Facilities AG ,Northgate Facilities AG
Three of the columns the module specifies for this file are missing. Which omission is the one that
makes the file unusable for review at all, rather than merely weaker?
Hint
Put yourself in the reviewer's seat with 214 rows and twenty minutes. Work out what you would need in
order to handle most of them without opening the CRM.
A. Incorrect. The whitespace in `current_value` is the defect being reported, not a defect in the file. Preserving the current value exactly as it appears in the source is correct behaviour, because a reviewer has to be able to see what is actually stored.
B. Incorrect. The confidence band is genuinely one of the three missing columns, and the stem asks which omission makes the file unusable rather than weaker. A missing band costs a reviewer the ability to triage marginal rows, and they can still work: they read the reason, check the rule, accept or reject. Without the reason and the rule there is nothing to read, so the review has no cheaper path than redoing the analysis. A band with no rule behind it is also just another unexplained score.
C. Correct. The shape that makes review possible is record ID, field, current value, proposed value, reason, rule, and confidence band, and the two missing here are the two that turn 214 individual decisions into a handful of decisions about rules. A reviewer who can sort by rule accepts every country-canonicalisation row in one action and spends their attention on the rows that need it. Row 3 is the sharpest illustration: filling a blank owner is a different kind of act from correcting a casing error, and nothing in the file says where "D. Marchetti" came from.
D. Incorrect. The four columns present are the four that describe the change. None of them describe the decision. The test for a draft-only artifact is whether a reviewer can act on it in less time than doing the work themselves, and 214 unexplained single-row judgments fails that test.
Question 4 of 6
An audit over a learner's own account export.
Data quality audit, accounts export
Source rows: 4,180
Defect class rows % of records
Missing required field 1,412 33.8%
Non-canonical picklist value 806 19.3%
Stale record (no update 18m+) 1,004 24.0%
Orphaned (no owner or no parent) 587 14.0%
Duplicate candidate 441 10.6%
Impossible date 362 8.7%
-----
Total defective rows 4,612 110.3%
What is the likely cause, and which reconciliation catches it?
Hint
Take one plausible row from a real CRM and ask how many of the six classes it could belong to at once.
Then look again at the line labelled Total.
A. Incorrect. A join fan-out is a real failure and it produces a different signature: the per-class counts inflate together and the source row count in the header is the one you cannot trust. Here the header states 4,180 and the individual class counts are all below it. The impossible number is the total, which is the line that was computed rather than counted.
B. Correct. A record with a blank owner that has also not been updated in two years is orphaned and stale, and it is one row. Per-class counts are the right thing to report, and summing them into a "total defective rows" line is the error, because the classes overlap by design. The reconciliation that catches it is a row-level accounting rather than a column-level one: every source row lands in exactly one bucket, and the distinct defective count is reported alongside the per-class counts rather than derived from them.
C. Incorrect. The percentages are consistent with the counts (1,412 of 4,180 is 33.8%), so the percentage column is a faithful rendering of a wrong total rather than an independent error. Checking that would have taken one division, and doing it is what tells you the arithmetic is fine and the model of the data is not.
D. Incorrect. Nothing here is invented. Every class count is plausible and the arithmetic is correct. Reaching for hallucination when the output has an ordinary logical flaw is expensive, because the rerun produces the same structure and you will have learned nothing about your own report design.
Question 5 of 6
Four defect definitions proposed for 40-reference/field-policy.md. One of them cannot be used. Which,
and what is the rewrite?
Overdue open pipeline: the opportunity's stage is one of the six open stages and close_date is
earlier than today.
Suspect record: the record looks wrong or has fields that do not seem right for the account.
Orphaned contact:account_id is blank, or account_id does not resolve to a row in
accounts.csv.
Stale account:lifecycle_status is Customer and there is no row in activities.csv for this
account_id in the last 180 days.
Hint
For each definition, imagine handing it to a colleague with the same file and comparing your two outputs.
Three of them would match. One of them could not.
A. Incorrect. You picked the right definition for the wrong reason, which is scored as a miss because the reason is what you use on the next definition you write. The problem is not that definition 2 flags too many rows. A definition that flagged too many rows would at least be checkable. The problem is that no two people, and no two runs, produce the same list from it. "Narrow it to the worst records" reproduces the same defect in a smaller box.
B. Incorrect. A threshold does need a reason, and it is not what makes a definition usable. 180 days is arbitrary in the sense that 150 or 200 would also be defensible, and it is precise in the sense that everybody applying it gets the same rows, which is the property that matters. Write the reason next to the number, and keep the number.
C. Incorrect. A definition relative to the run date is normal for an operational audit, and the standard fix is to record the run date in the run log rather than to reject the definition. Comparing two runs from different dates is a reconciliation question, not a definitional one.
D. Correct. The test for a defect definition is whether a second person applying it to the same file produces the same rows. Definitions 1, 3 and 4 all pass because they name fields and conditions. Definition 2 fails on falsifiability rather than on breadth: there is no output it could produce that you could call wrong, which means there is nothing to sample against and no reconciliation to run. The rewrite in this option is the important half, because it converts an impression into a condition and routes the genuinely unclassifiable rows to exceptions rather than into a class.
Question 6 of 6
A field completeness report.
Required-field completeness, accounts export
Field populated missing % complete
owner 4,061 119 97.2%
industry 3,798 382 90.9%
country 4,177 3 99.9%
renewal_date 2,904 1,276 69.5%
health_score 3,110 1,070 74.4%
It has no breakdown by owner and no breakdown by segment, and each omission costs you something.
The report was commissioned to get the gaps closed. Which action does it fail to enable, the one
without which nothing gets fixed at all?
Hint
Read the report and then write the next email you would have to send. Notice what you would have to go
and find out before you could address it to anybody.
A. Incorrect. Progress measurement works fine from the totals in this report: run it monthly and the percentages move. It is a real loss that you cannot see *whose* completeness improved, and that is a consequence of the missing owner breakdown rather than a separate failure.
B. Correct. The deliverable of a completeness pass is not a report, it is an assignment list. Every blank field was created by somebody, usually a small number of somebodies, and the breakdown by owner is what converts a statistic into a set of tasks with names on them. Without it the report gets forwarded, agreed with, and changes nothing, which is the most common fate of a data quality document.
C. Incorrect. This is a genuine loss and it is not the one the stem asks about. Segment gives you prioritisation, owner gives you accountability, and the stem asks which action fails without the report producing any fix at all. A report with no segment breakdown still gets the gaps closed, in a worse order. A report with no owner breakdown gets nothing closed, because no task has a name on it. The module names both breakdowns for this reason, and this item is about the one that is load-bearing.
D. Incorrect. "It answered the question it was asked" is exactly the reasoning that produces reports nobody uses. The question behind a completeness pass is always "so what do we do", and a report that stops one step short of that is a report that gets read once. Deciding what a report has to enable before writing it is the design step this item is testing.
Last verified: 2026-08-06.
Product claims trace to course/reference/COWORK-FACTS.md (FACTS §x.y). Dataset claims trace to
course/data/DATA-DICTIONARY.md and the pack's answer-keys/PLANTED-ISSUES.md for the Talvera
dataset generated with seed 20260801.
The module at a glance
Every Monday your manager asks the same two questions: what changed, and what is at risk. You
rebuild the same view by hand, it eats most of a morning, and by Thursday it is stale. The risk half
is worse: "feels solid" is not a forecast input, and when a deal slips nobody can point at the
signal that was there and ignored.
This module builds a weekly pipeline review pack that says what moved, flags risk with the evidence
line attached to every flag, and states plainly what it does not judge. It runs the full TRACK loop
once, over deal inspection and forecast inputs.
Lesson
Title
TRACK
Minutes
Module-level piece
L6.1
Split the review into retrieval and judgment
T, R
11
L6.2
Movement that reconciles, and the week that did not happen
A
12
Failure lane 1
L6.3
Risk rules that emit evidence, and the deal card
A
14
Worked exercise
L6.4
Hunt the false positive, then keep the pack
C, K
13
Discrimination pair, Failure lane 2
Total lesson time: 50 minutes. Module budget is 45 (PEDAGOGY.md §3.1), tolerance plus or minus
15%.
Then: checkpoint CP-06 (about 10 minutes) and LAB-06 (Part A 50 minutes, Part B Full 40 or Part
B Lite 15).
Outcomes
O6.1 Produce a week-over-week pipeline movement analysis from snapshot data covering created,
advanced, slipped, shrunk, grown and closed, with counts and amounts that reconcile.
O6.2 Write risk rules that produce a flag plus the evidence line that justifies it, and reject
any risk output that cannot show its evidence.
O6.3 Produce a per-deal inspection card containing stage, amount, age in stage, next step,
owner, risk flags with evidence, and what the data does not tell you.
O6.4 Produce a review pack a manager can read in five minutes, and state explicitly what it
excludes and why.
O6.5 Verify a risk flag against source activity data, find a false positive, and change the
rule rather than the output.
O6.6 State which forecast judgments must remain human and defend the boundary.
The artifact you leave with
The pipeline review pack, in your own revenue-workspace/:
File
What it is
Built in
20-briefs/pipeline-review.brief.md
The weekly brief, rerunnable
L6.2, completed L6.4
40-reference/risk-rules.md
Named rules, thresholds, and the evidence line each one emits. Versioned.
L6.3, revised L6.4
30-output/pipeline/review-YYYY-WW.md
The five-minute pack
L6.4
30-output/pipeline/deal-cards/<opportunity-id>.md
One card per flagged deal
L6.3
50-runs/RUN-LOG.md entry
Reconciliation result plus the one number you rebuilt by hand
L6.4
Templates are in templates/ beside this README.
risk-rules.md is the durable asset. It is a written statement of what your team believes predicts
a slipped deal, and most teams have never written one down. Its version history is what shows
judgment: every entry is a false positive you found and a rule you changed because of it.
We do not print a time saving here. LAB-06 measures your own before and after, including the
verification cost, and the number that reaches your capstone is the one you measured.
Blast radius
Read-only, all of it. The pack reads exports and writes markdown into your own output folder.
Nothing in this module writes to a CRM. If you want next steps written back into the system, that is
a Module 5 promotion decision with a reversal plan and a named approver, and it is not this
workflow.
Prerequisites
Modules 1 to 4 for the method. Module 5 at least at the core pass, because a movement analysis over
duplicated records produces a reconciliation that cannot be made to balance, and you will spend an
hour looking for an error that is in the export rather than in your work.
What this unlocks
Module 8's renewal pipeline work is this pattern applied to a different object. On the SELL and CS
tracks this is one of your deepest modules.
A note on the numbers
Every count in these lessons was computed from the shipped dataset at the seed above. They are
synthetic and they are exact. The snapshot file holds 1,383 rows covering 26 Mondays from
2026-02-02 to 2026-07-27, and the dataset's "today" is 2026-08-02
(DATA-DICTIONARY.md). Older weeks age out of the warehouse, which is why the history does not
reach back to January, and which is itself the kind of thing you have to know before you write a
year-to-date anything.
Module 6 · Pipeline inspection and forecast inputs
M6 lesson / about 11 min
Concept
Split the review into retrieval and judgment
Do this nowPhotograph the snapshot file first60s
Point a session at 10-data/ and ask for the total row count of pipeline_snapshots.csv.
Ask for the distinct snapshot_date values in order, with the row count for each.
Read the first date, then look for two consecutive weeks with the same count.
You know it worked when: 1,383 rows across 26 Mondays from 2026-02-02 to 2026-07-27, and the weeks of 2026-04-13 and 2026-04-20 both come back at exactly 50.
What you type
Read opportunities.csv, contacts.csv and 10-data/notes/call-notes/ in this
folder.
Report the current position on OPP-007372: amount, close date, stage, and who
the economic buyer is.
Where a call note and a CRM row disagree on a field, do not reconcile them.
Print both values, name the file each came from, and flag the conflict.
Call notes are evidence, never the source of a number.
What comes back
OPP-007372, Dovecote Marketplace SA (ACC-02826)
opportunities.csv notes/call-notes/dovecote-marketplace-update.md
amount 78,000 EUR 54,000 EUR "once the volume discount is applied"
close_date 2026-09-22 "signature around 12 August 2026"
stage Proposal "with their legal team for contract review"
Stakeholder conflict:
contacts.csv CON-009444 Lorcan Ilkka, IT Manager, persona Technical Evaluator
the note calls him the person who signs off the budget and says it is already
approved on their side
The note closes with "Record updated accordingly." The record reflects none of
it. Three fields in conflict. Not reconciled.
Notice: There is no correct single answer here. The useful output names both numbers, says which system owns which field, and hands the gap to a person.
Monday, 8:40am. Your manager wants to know what changed and what is at risk. Behind that sits a
third question they will not ask directly: are we going to make the quarter. Before you hand any
of this review over, you need to know which parts stay yours, because the third question is the
one that gets you into trouble if you let a tool answer it.
By the end of this lesson you can: split the weekly review into three tasks and score them on
the fit test, name the one that stays human and defend it with the deciding factor, and rank your
four data sources by trust.
Three tasks wearing one name
What moved
Which deals are at risk
Are we calling the quarter
Volume
5. 60 open deals, 26 snapshot weeks
5. Every open deal, every week
1. One decision
Checkability
5. Two snapshots, arithmetic
4. Every flag is checkable against the row that fired it
1. Nothing to check it against until the quarter ends
Judgment density
1
2, once the rules exist
5
Cost of being wrong
2. Corrected next week
3. A wrong flag wastes an inspection
5. Your credibility, and somebody's hiring plan
Specification cost
2
3. An hour to write the rules
5. You cannot write down what "right" looks like in advance
Verdict
Hand it over
Hand it over, with evidence on every flag
Human. Permanently.
The third column fails on checkability, and low checkability is a veto no matter what the other
rows say. There is no source of truth to compare a forecast call against until the period closes,
and by then the call is not needed. A better model does not fix that.
This module builds columns one and two. L6.4 gives you the words to defend column three to your
VP.
Four sources, four levels of trust
Confident wrong flags come from treating everything in the folder as equally true. Rank your
sources, out loud, in the brief. All four sources below came in your practice pack's data folder;
if you have not downloaded the pack from the course page, do that first.
1. Pipeline snapshots are facts about the record.pipeline_snapshots.csv is a weekly
photograph of open pipeline: one row per open opportunity per Monday, 1,383 rows over 26 weeks
from 2026-02-02 to 2026-07-27 (DATA-DICTIONARY.md). If a snapshot says the close date was
2026-06-30 on 28 April, that is what the record said that day. It is the only source where "what
changed" has an unambiguous answer, which is why movement is built on it. But facts about the
record are not facts about the deal: the snapshot says the close date moved, not that the deal
moved.
2. The opportunity file is current state, not history.opportunities.csv tells you where a
deal is now. It has 528 rows and 524 distinct opportunity_id values, because the intended
primary key is not unique in the file (DATA-DICTIONARY.md). If you have not done Module 5 on
this data, do it before you build a roll-up on it.
3. Activity records are facts about logging behaviour. This one catches people.
activities.csv tells you what somebody wrote down. A rep who runs three good calls and logs
none looks dormant. A rep who logs every unanswered voicemail looks busy. duration_minutes is 0
for emails and notes, and contact_id is blank for internal notes (DATA-DICTIONARY.md), so a
deal can look active in a week where nobody spoke to the customer at all. L6.4 shows a real deal
in this dataset where that happens.
4. Call notes are written for a human reader. Sixteen files, mixed .txt and .md,
inconsistent headers, fragments, action items buried in prose, one filename with spaces
(DATA-DICTIONARY.md). Rich, unstructured, and an agent will read them confidently.
And one planted warning: one call note contradicts the CRM.answer-keys/PLANTED-ISSUES.md
in your pack, which you open only after attempting the lab, documents it
under the class note-crm-contradiction: for opportunity OPP-007372 on account ACC-02826,
the note states an amount, a close date and a stakeholder role the CRM rows flatly disagree with,
and claims the record has been updated when it has not. Every other statement in every other note
agrees with the CSVs.
There is no single correct answer to "what is this deal worth". The right output names both
numbers, says which system is authoritative for which field, and raises the gap. "Summarise the
deal" gets you one source, stated with confidence. This instead gets you the most useful line in
the pack:
Movement comes only from pipeline_snapshots.csv.
Current state comes only from opportunities.csv.
Engagement comes from activities.csv and is described as "logged activity",
never as "engagement".
Call notes are evidence, never a source of a number. Where a note and a CRM row
disagree, report both values and flag the conflict. Do not reconcile silently.
Read-only, and why that is good news
The entire pack is read-only: it reads exports from 10-data/ and writes markdown into
30-output/pipeline/. Cowork reads, creates and modifies files only inside folders you have
granted, so the blast radius of a wrong flag is a wrong sentence in a file you have not sent yet.
That is why you can iterate fast: wrong ten times on a Monday morning, at no cost. The moment
this workflow writes next steps back into the CRM, it becomes a Module 5 promotion decision with
a named approver and a tested reversal, not a thing you tune casually.
Worth knowing: two product facts shape the design.
Nothing fires when a stage changes. The pack is a cadence you run or schedule. A deal that
goes wrong on Tuesday is caught the following Monday. That latency is a number you accept in
writing; Module 8 makes you write it down.
There is no memory between sessions. Next Monday's session knows nothing about this one.
The rules file is the memory, which is why risk-rules.md, the rules file you create in L6.3,
is a file and not a habit.
Parallel tasks run genuinely in parallel, each with its own context and file access: useful when
you want movement and risk computed at the same time. Any hard cap on how many is unverified, so
plan for "several", not a number.
Try this now
Point Cowork at revenue-workspace/. In 10-data/, place pipeline_snapshots.csv,
opportunities.csv, activities.csv, and the notes/ folder, all copied from your pack's data
folder. Do not put the compensation plan
or the board deck in there: the directory boundary is the boundary.
Then run the inventory, before any analysis:
Read 10-data/pipeline_snapshots.csv. Report: total row count, the list of
distinct snapshot_date values in order, and the row count for each date.
Report nothing else. Write no file.
What you should see: 1,383 rows across 26 Mondays, 2026-02-02 through 2026-07-27, between 45
and 61 rows a week. Two things to notice before you analyse anything.
First, the history starts on 2026-02-02, not in January. Older weeks age out of the warehouse
(DATA-DICTIONARY.md). Any "year to date" statement from this file is wrong.
Second, the weeks of 2026-04-13 and 2026-04-20 both hold exactly 50 rows. That proves nothing by
itself, but you want it in front of you before you start, and L6.2 is about what it turns out to
mean.
What you keep
Two lines for the top of 40-reference/risk-rules.md, once you create it in L6.3:
Trust ordering: snapshots for movement, opportunity file for current state, activities as
logged activity only, notes as evidence and never as a number.
Human-only, permanently: the forecast call, the decision to inspect a rep, the decision to
discount, and the judgment of whether a champion is real. Reason: no source of truth exists to
check the output against, which is a veto under the fit test.
Check yourself
A colleague builds "deals with no engagement in 30 days" from activities.csv. Name two
different ways that flag can be wrong, one in each direction.
Your manager asks for pipeline created year to date from the snapshot file. What do you say?
A call note says a deal is at 54,000 EUR and closing on 12 August. The CRM says 78,000 EUR and
22 September. What does your pack print?
Where this goes next
L6.2 builds the movement table and the reconciliation line that proves it, then breaks it on a
week that never happened.
Module 6 · Pipeline inspection and forecast inputs
M6 lesson / about 12 min
Failure mode
Movement that reconciles, and the week that did not happen
Do this nowDiff two snapshot weeks60s
Filter pipeline_snapshots.csv to snapshot_date 2026-04-13, then to 2026-04-20, and count each.
Pick one opportunity_id present in both weeks.
Compare stage, amount, close_date and age_days across the two rows.
You know it worked when: Both weeks hold 50 rows, and every field on the row you picked is identical including age_days, which cannot be true seven days apart.
What you type
OUTCOME
Produce the week-over-week pipeline movement table for the week ending
2026-04-20.
INPUTS
10-data/pipeline_snapshots.csv, weeks 2026-04-13 and 2026-04-20 only.
10-data/opportunities.csv, used only to resolve the closed stage of deals that
left the file. 10-data/accounts.csv, used only to read accounts.currency.
OUTPUT SHAPE
30-output/pipeline/review-2026-W17.md, section "Movement".
Columns: change_type | count | amount_usd | additive.
Amounts converted at EUR 1.08, GBP 1.27, USD 1.00.
Below the table, two reconciliation lines, each stating pass or fail:
rows: opening + created - exits = closing
amount: opening + created - exits + increases - decreases = closing
DEFINITION OF DONE
Both reconciliation lines print, and both say pass, or the run reports why not.
Notice: Both reconciliation lines pass and the week is fiction: the 2026-04-20 snapshot is a copy of 2026-04-13, and correct arithmetic on a duplicated file stays correct and stays useless.
Total Closed Won on the shipped data is 16,830,000 raw, 16,679,000 after deduplicating on the primary key, 17,980,560 after currency conversion, and 17,995,680 once the account is treated as the authority for currency.The week ending 2026-04-20 shows zero movement on every change type and both reconciliation lines pass, because the snapshot is a byte copy of 2026-04-13 rather than a quiet week.
"What changed this week" sounds like the easiest question in the pack, and it goes wrong the
most quietly. Nobody reads "12 deals advanced" and thinks "that cannot be right". They think
"fine" and move on. So the table needs a proof attached: arithmetic that either closes or does
not.
By the end of this lesson you can: specify a movement table with fixed change types and a
reconciliation line, tell additive rows from non-additive ones, and refuse a zero-movement week
until you have run the check. Everything this lesson uses came in the practice pack from the
course page and was copied into 10-data/ in L6.1; if you have not downloaded the pack, do that
first.
Fix the change types, or you will get different ones every week
The change types, closed set, in this order:
Change type
Definition, between snapshot week W-1 and W
created
opportunity_id present in W, absent in W-1
advanced
stage index in W greater than in W-1
slipped a stage
stage index in W lower than in W-1
close date pushed
close_date in W later than in W-1
close date pulled in
close_date in W earlier than in W-1
amount increased
amount in W greater than in W-1, converted
amount decreased
amount in W lower than in W-1, converted
closed won
present in W-1, absent in W, and the opportunity file says Closed Won
closed lost
present in W-1, absent in W, and the opportunity file says Closed Lost
left the file, other
present in W-1, absent in W, and neither of the above
That last row exists because "an opportunity leaves the snapshot file the week after it closes"
(DATA-DICTIONARY.md in your pack's data folder). A deal that leaves for any other reason is something you want to see, not
something quietly absorbed into "closed lost".
Two kinds of row, and only one kind adds up
Most hand-built movement tables get this wrong:
Some rows change the total. created, closed won, closed lost, left-other, amount increased,
amount decreased. These carry a delta.
Some rows do not. advanced, slipped, close date pushed, close date pulled in. A deal that
advances a stage is worth the same money it was worth last week.
Print one amount column and add it up, and you double count every deal that both advanced and
grew. So the table prints the deal value for non-additive rows, the delta for additive rows,
labels which is which, and reconciles on the additive ones only:
opening amount
+ created
- closed won
- closed lost
- left the file, other
+ amount increases
- amount decreases
= closing amount
And the same for counts:
opening rows + created - all exits = closing rows
The reconciliation is not decoration. It catches a snapshot join done wrong, the most common
silent failure in this workflow.
One currency, and one authority for it
pipeline_snapshots.amount is in the deal's billing currency (DATA-DICTIONARY.md). Convert
before you compare anything, using the fixed rates in the data dictionary (EUR 1.08, GBP 1.27),
and say in the brief that the account is the authority for the currency, exactly as in Module 5.
A live example: OPP-006697 is labelled USD in opportunities.csv while its account
ACC-03045 bills in EUR, a documented defect (answer-keys/PLANTED-ISSUES.md in your pack,
class SUBTLE currency-mismatch, which you open only after attempting the lab), and the same
deal is labelled EUR in all 26 of its snapshot rows.
So snapshot totals and opportunity totals disagree on this one deal, and the account file settles
it. Not a trick question. A Tuesday.
Try this now
Run the movement section of the brief for real:
OUTCOME
Produce the week-over-week pipeline movement table for the week ending
2026-07-27.
INPUTS
10-data/pipeline_snapshots.csv, weeks 2026-07-20 and 2026-07-27 only.
10-data/opportunities.csv, used only to resolve the closed stage of deals that
left the file.
10-data/accounts.csv, used only to read accounts.currency.
OUTPUT SHAPE
30-output/pipeline/review-2026-W30.md, section "Movement".
One table. Columns: change_type | count | amount_usd | additive.
Amounts converted to USD at EUR 1.08, GBP 1.27, USD 1.00.
For additive rows amount_usd is the delta. For non-additive rows it is the
current value of the deals concerned, and the additive column reads "no".
Below the table, two reconciliation lines, each stating pass or fail:
rows: opening + created - exits = closing
amount: opening + created - exits + increases - decreases = closing
CONSTRAINTS
Stage order is Qualification, Discovery, Solution Review, Technical Validation,
Proposal, Negotiation.
accounts.currency is the authority where it disagrees with the snapshot.
Before computing anything, compare the two snapshot weeks on row count and on a
hash of the ordered rows. If they are identical, stop and report it.
DEFINITION OF DONE
Both reconciliation lines print and both say pass, or the run reports why not.
What you should see, from the shipped data:
Week ending 2026-07-27. Opening 59 open deals, 6,126,525 USD. Closing 60 open deals,
6,200,285 USD.
This gives your Monday meeting what "pipeline is up 73,760" does not: ten deals advanced worth
537,640, two close dates out on 322,000 of pipeline, one deal closed won at 27,000. The total
went up and the quarter got harder, both at once.
When this goes wrong
The break. Run the same movement analysis across the whole snapshot history and trust the
file. Specifically, run the week ending 2026-04-20.
Stop on that, because it is the whole lesson. The reconciliation passes. Both lines say pass.
The week is fiction.
The 2026-04-20 snapshot is a byte-for-byte copy of 2026-04-13, all 50 rows, identical apart from
the date: the weekly load failed and the previous file was reused (PLANTED-ISSUES.md, class
SUBTLE broken-etl). In a business that advanced between 4 and 17 deals a stage in every other
week of the file (a median of 11 or 12 across the other 24 week pairs), nobody touched anything
across 50 open deals. Not a quiet week. A missing week.
The damage spreads, too. The following week absorbs two weeks of movement into one:
week
advanced
closed won
closed lost
created
2026-04-06 to 04-13
10
1
1
5
2026-04-13 to 04-20
0
0
0
0
2026-04-20 to 04-27
13
5
4
4
A manager reading that sequence sees a stall and then a surge, and will explain both. Neither
happened.
What catches it. A consistency check is not a verification. Arithmetic on a duplicated file
is still correct arithmetic, so what catches it sits outside the arithmetic: compare each
snapshot week against the previous one on row count and on a hash of the ordered rows, before
any movement runs. Two consecutive weeks that hash identically did not happen. If the hash is
unavailable, fall back to checking that at least one of stage, amount or close_date differs
somewhere between the two weeks.
The guard. Three things, in this order, permanently in the brief:
Snapshot integrity test runs first. Any week identical to its predecessor is reported and
the run stops.
A zero-movement week is never accepted silently. If the integrity test passes and movement
is still zero, that is a finding and it goes in the pack in words.
The affected week is a judgment, not a patch. Exclude it and say so in the exclusions
section: "the week of 2026-04-20 is excluded because the snapshot is a duplicate of 2026-04-13;
the week of 2026-04-27 therefore contains two weeks of movement." Do not smooth it,
interpolate it, or quietly drop it. The reader is about to draw a trend line through it, and
they need to know.
What you keep
20-briefs/pipeline-review.brief.md, a brief you start this lesson from
templates/M6/pipeline-review.brief.md in your practice pack, with the movement section, the
change type list, the reconciliation lines and the snapshot integrity test.
The first 30-output/pipeline/review-YYYY-WW.md with a movement table that reconciles.
Check yourself
Your movement table says opening plus changes does not equal closing, and the gap is exactly
one large deal. Name the two most likely causes and the check that separates them.
Why is "advanced" excluded from the amount reconciliation? What breaks if you include it?
Your integrity test flags two identical weeks. Your manager wants the pack in ten minutes. What
goes in the pack?
A colleague fixes the duplicated week by averaging the movement across the two affected weeks
so the trend looks smooth. Name what that costs the reader.
Where this goes next
L6.3 turns "which deals are at risk" into named rules that emit evidence, and produces the
first deal card.
Module 6 · Pipeline inspection and forecast inputs
M6 lesson / about 14 min
Do it
Risk rules that emit evidence, and the deal card
Do this nowRead three fields on one deal60s
Open opportunities.csv and find OPP-006969.
Read days_in_stage, close_date and next_step.
Open accounts.csv, find ACC-02995, and read segment.
You know it worked when: 214 days in stage against a Mid-Market norm of 32, a close date of 2026-06-23 that has passed, and a next step reading "Waiting on their IT security review".
What you type
OUTCOME
Produce one deal inspection card for OPP-006969.
INPUTS
10-data/opportunities.csv, accounts.csv, activities.csv, notes/call-notes/
Rules: 40-reference/risk-rules.md, rules R1 to R3.
OUTPUT SHAPE
30-output/pipeline/deal-cards/OPP-006969.md
Sections: header, position, flags, corroboration, what the data does not tell you.
Every flag prints on its own line as:
<rule name> | <the field or record that fired it> | <the value>
CONSTRAINTS
Run date is 2026-08-02. The segment norm comes from DATA-DICTIONARY.md via the
account's segment (SMB 18, Mid-Market 32, Enterprise 48).
For the engagement gap, count a row only where contact_id is populated and
activity_type is not Note.
Convert the amount using the account's billing currency as the authority.
EXCEPTIONS RULE
If a rule cannot be evaluated, say so on its line rather than omitting it.
What comes back
30-output/pipeline/deal-cards/OPP-006969.md written.
# OPP-006969, Selby Supply Chain BV
Account ACC-02995, Netherlands, Mid-Market, Logistics. Owner Bettina Farrow.
Stage Technical Validation
Amount 54,000 EUR (58,320 USD at 1.08)
Forecast Commit
Close date 2026-06-23
Next step Waiting on their IT security review
## Flags
R1 stage-age | opportunities.days_in_stage vs Mid-Market norm 32 | 214 (6.69x)
R2 overdue | opportunities.close_date vs run date 2026-08-02 | 2026-06-23, 40 days past
R3 gap | last qualifying activity, ACT-0043361 | 2026-06-16, 47 days
## Corroboration
notes/call-notes/call-selby-supply.txt: "Nobody in the building owns the third
gate."
Notice: Each flag names a file, a field and a value, so a manager can argue with your threshold of 32 rather than arguing with the list.
You send a manager a list of at-risk deals. They ask why one is on the list. If your answer is
"the model scored it 0.72", the conversation is over, and so is the credibility of every future
list. The real unit of work is the sentence under each flag that lets somebody argue with it.
By the end of this lesson you can: write a risk rule that names a field, a comparison and a
threshold you chose; make every flag emit an evidence line; and produce a deal card a manager
reads in ninety seconds without mistaking it for a verdict.
Rules first, in a file, before any run
Write 40-reference/risk-rules.md, versioned, before you ask for anything. Cowork carries no
memory between sessions. The rules file is the memory.
Shapes that work, with the thresholds left to you:
Rule
Shape
The threshold you must choose
stage age
days_in_stage above N times the segment norm
the multiple
overdue
open stage and close_date before the run date
none, this one is binary
engagement gap
no two-way customer activity in N days on a deal above X
N and X
single thread
one contact with activity on a deal above X
X
repeat slip
close_date moved more than N times in the snapshot history
N
no next step
next_step blank, or a date inside it that has passed
none
stated blocker
a call note contains an explicit blocker the stage does not reflect
which words count
Segment norms for stage age are in the data dictionary (DATA-DICTIONARY.md in your pack's data
folder): SMB about 18 days, Mid-Market about 32, Enterprise about 48. One flat number across segments is the most common way this rule produces
noise: a 60-day Enterprise deal is normal and a 60-day SMB deal is not.
State your norm basis in the rules file, in one line. "Segment norm as published in
DATA-DICTIONARY.md" is a different basis from "the observed mean days_in_stage for that
segment" and from "the observed median for that segment and stage", and the three give different
multiples on the same deal. None of them is wrong. An unstated basis is, because two people then
reproduce different numbers from the same file and neither can find the disagreement. Everything
in this lesson uses the published segment norms.
Every flag emits an evidence line
The evidence line has exactly three parts:
<rule name> | <the field or record that fired it> | <the value>
For example:
R1 stage-age | opportunities.days_in_stage vs Mid-Market norm 32 | 214
R2 overdue | opportunities.close_date vs run date 2026-08-02 | 2026-06-23
R3 gap | last two-way activity on OPP-006969 | 2026-06-16, 47 days
Three tests a flag has to pass before it goes in the pack:
Can a person check it in under a minute without asking you? The evidence names the file,
the field and the value, so yes.
Can a person disagree with it? "32 days is the wrong norm for this segment" is an
argument about the rule, which is productive, not about the output.
If it is wrong, does the fix land in the rule? If the only fix is deleting the row, the
flag was an assertion, not evidence.
A risk output without evidence lines is rejected in this module. Not a style preference: it
is the difference between a pack that changes somebody's Tuesday and a pack that gets skimmed.
The deal card, and its last line
One card per flagged deal. You write each card yourself; a seed layout ships in
templates/M6/deal-card.md of your practice pack. Fixed sections:
Deal, account, owner
Stage, amount in local currency and USD, forecast category
Age in stage, against the segment norm
Days since last two-way customer activity
Next step, as recorded
Flags, each with its evidence line
Corroboration from notes, quoted, with the filename
What the data does not tell you
The last section is the honesty gate, and it is required. Without it the card is a verdict, and
a manager acts on a verdict. With it the card is an inspection, and a manager asks a question.
It is also the easiest section to write badly: "further information may be required" is not a
sentence. The good version names a specific unknown and who could answer it.
Try this now: three rules, one deal, one card
About nine minutes. This is the exercise scaled up by the exercise sheet in labs/LAB-06 of your
practice pack (start with its README).
You are the AE covering EMEA logistics at Talvera. Monday 2026-08-02. Your manager has stopped
asking "how is Selby going" in the meeting and started asking it in writing, which is worse. You
have opportunities.csv, accounts.csv, activities.csv and the call notes folder, all from
your pack's data folder.
Create 40-reference/risk-rules.md from templates/M6/risk-rules.md in your practice pack.
Write three rules with your own thresholds: stage age, overdue close date, and
engagement gap. Each needs a name, a field, a comparison, a threshold, and the evidence
line it emits.
Fill the gaps in the starter brief below and run it against opportunity OPP-006969.
Read notes/call-notes/call-selby-supply.txt yourself. Add one quoted line of corroboration
to the card, with the filename.
Write the "what the data does not tell you" section by hand. Do not delegate it. It is the one
place in the card where you are the source.
Compare against the reference card below.
Starter brief (fill the gaps marked <<...>>)
OUTCOME
Produce one deal inspection card for OPP-006969.
INPUTS
10-data/opportunities.csv, 10-data/accounts.csv, 10-data/activities.csv
10-data/notes/call-notes/
Rules: 40-reference/risk-rules.md, rules R1 to R3.
OUTPUT SHAPE
30-output/pipeline/deal-cards/OPP-006969.md
Sections in this order: header, position, flags, corroboration, what the data
does not tell you.
Every flag prints on its own line as:
<<state the three parts of the evidence line format>>
CONSTRAINTS
Run date is 2026-08-02.
The segment norm for stage age comes from <<where, and how do you get the
segment for this deal?>>.
For the engagement gap, count only <<which activity rows count as two-way
customer contact, and which do not?>>.
Convert the amount to USD using the account's billing currency as the authority.
Do not write anything to any file in 10-data/.
DEFINITION OF DONE
Every flag has an evidence line naming a field and a value.
The corroboration section quotes the source file by name or says "none found".
The final section is present and is not empty.
EXCEPTIONS RULE
<<what happens if a rule cannot be evaluated, for example because the deal has
no activity rows at all?>>
Stuck? Open these in order
Hint 1 (Concept)
The card is a by-product you throw away next week. The real output is the rules file, and the
test of a rule is whether somebody who disagrees with the flag can point at the part they
disagree with. Write the rule so it has parts.
Hint 2 (Move)
For stage age you need the account's segment: join opportunities.account_id to
accounts.account_id, read accounts.segment, then compare days_in_stage against the norm for
that segment (SMB about 18, Mid-Market about 32, Enterprise about 48, from DATA-DICTIONARY.md).
For the engagement gap, activities.csv has a related_type column: filter to
related_type = Opportunity and related_id = OPP-006969. Then look at what is left before you
take the maximum timestamp, because not every row in it is a customer conversation.
Hint 3 (Shape)
A correct answer fires more than one of your three rules, and independently: one about how long
it has sat, one about a date that has passed, one about silence. If only one fires, check whether
you wrote one compound rule instead of three.
The stage-age multiple here is high but not the highest in the file, so a threshold set to catch
only the worst deal misses this one. Set it at the level you would actually inspect at.
Your "what the data does not tell you" section should hold at least one thing the call note tells
you that no CSV column holds. If it only says "we do not know if it will close", write it again.
Reference card
# OPP-006969, Selby Supply Chain BV
Account ACC-02995, Netherlands, Mid-Market, Logistics. Owner Bettina Farrow.
## Position
Stage Technical Validation
Amount 54,000 EUR (58,320 USD at 1.08)
Forecast Commit # self-reported, see risk-rules.md section 0
Age in stage 214 days
Close date 2026-06-23
Next step Waiting on their IT security review
Last two-way 2026-06-16
## Flags
R1 stage-age | opportunities.days_in_stage vs Mid-Market norm 32 | 214 (6.69x)
R2 overdue | opportunities.close_date vs run date 2026-08-02 | 2026-06-23, 40 days past
R3 gap | last two-way activity on OPP-006969 | 2026-06-16, 47 days
## Corroboration
notes/call-notes/call-selby-supply.txt, written by the deal owner:
"Three gates. Legal (done, months ago). Data protection (done). Infrastructure
security, which is one contractor two days a week and a backlog. Nobody in the
building owns the third gate."
"close date on the record reads 2026-06-23 and that date has been and gone."
The note independently confirms R2, and gives R1 a cause: the deal is not stalled
on our side of the table and it is not stalled on price.
Worth reading next to the header: the record carries forecast_category = Commit
on a deal whose close date passed 40 days ago and which has sat in one stage for
214 days. This card does not treat that as a flag, because forecast_category is
a field somebody sets rather than something that happened. L6.4 is about the
difference.
## What the data does not tell you
- Whether the third approval gate has an owner now. No field in the CRM holds
this. Only the deal owner can answer it, and the answer is the deal.
- Whether the budget is real. The note records the champion saying it is funded
and declining to say from where. That is not in any column.
- Whether 214 days in Technical Validation is a dying deal or a slow procurement
cycle at an account that will eventually buy. The record cannot separate those
two, and this card does not try.
Your card may differ, and still be right. A stage-age threshold of 3x rather than 2x still
flags this deal, and a 30-day engagement gap threshold rather than 45 still flags it, so your
flags will match even if your numbers differ. Your multiple will differ from 6.69x if you used
observed means or medians rather than the published segment norms, a legitimate basis as long as
your rules file names it.
On the engagement date, this deal is unusually forgiving: the most recent activity row of any
kind and the most recent two-way customer contact are the same row, ACT-0043361 on 2026-06-16,
so a careless rule and a careful one agree here. A date later than 2026-06-16 means you read
another deal's rows. Counting only Meetings and Demos gives 2026-05-22 and a 72-day gap,
defensible if your rule says so. The two rules diverge badly on a different deal, and L6.4 is
that deal.
Worth knowing: the three flags took seconds and were already visible in three columns. The
line that changes your manager's Tuesday came out of a text file no rule can read, which is why
the card has a section only you can write.
What you keep
40-reference/risk-rules.md, version 1, with three rules and your thresholds.
30-output/pipeline/deal-cards/OPP-006969.md, your first card.
Check yourself
A risk output arrives with a column risk_score and no evidence. Name the minimum change that
makes it usable, and say why raising the score threshold is not it.
Your stage-age rule uses one flat threshold of 60 days across all segments. Which segment does
it under-flag and which does it over-flag, and what does that do to the credibility of the
pack?
A card's final section reads "further analysis may be required". Rewrite it for a deal where
the champion has gone quiet and the economic buyer has never been on a call.
Where this goes next
L6.4 hunts a false positive in these rules, changes the rule rather than the output, and
assembles the pack a manager reads in five minutes.
Module 6 · Pipeline inspection and forecast inputs
M6 lesson / about 14 min
Failure mode
Hunt the false positive, then keep the pack
Do this nowCheck the biggest deal's amount60s
Open opportunities.csv and read the amount on OPP-007017.
Filter pipeline_snapshots.csv to the same opportunity_id and read the amount on 2026-07-27.
Write both numbers next to each other.
You know it worked when: 1,000,000.0 on the opportunity row against 100,000.0 in the snapshot six days earlier, and no week in the 26-week history where it grew.
What you type
Read activities.csv and opportunities.csv in this folder.
For every open opportunity, find the most recent row in activities.csv where
related_id equals the opportunity_id, and report the days since that timestamp
against a run date of 2026-08-02.
Flag any deal above 50,000 USD whose gap is more than 45 days.
Output: opportunity_id, account_id, amount, last activity date, days, flagged.
For OPP-007409, also print the three most recent activity rows in full.
What comes back
OPP-007409 ACC-02908 212,000 EUR 2026-07-16 17 days not flagged
three most recent rows on this opportunity:
ACT-0043721 2026-07-16 Note contact (blank) No answer Handover note
ACT-0045630 2026-07-10 Note contact (blank) Completed Call notes logged
ACT-0041766 2026-06-17 Email contact CON-008733 Connected Chasing signature
Fairholm Facilities Ltd, Enterprise, Ireland. In Technical Validation 101 days.
Reported as engaged.
Notice: The two most recent rows are internal notes with a blank contact_id, so nobody at Fairholm was on the other end, and the last real customer contact was 2026-06-17, 46 days before the run date.
Your rules ran. You have eleven flagged deals and a pack that looks convincing. Before it goes
anywhere, assume it is wrong somewhere and find out where, in less time than redoing the work.
Then do the right thing with what you find: change the rule, not the output. A corrected pack is
worth one week. A corrected rule is worth every week after it.
By the end of this lesson you can: verify a flag against source and find the wrong one, fix
the rule rather than the row, rebuild one number by hand every week, assemble a pack a manager
reads in five minutes, and defend the boundary around the forecast call out loud.
The hunt
The procedure, before the pack goes anywhere:
Rank by amount and verify the top one by hand. Not a random one: the one a manager will
act on first.
For each flag on it, open the source and read the value. That is what the evidence line
is for.
Cross-check the number against a second source. A single source cannot contradict itself.
When one is wrong, ask where the fix belongs. Deleting the flag is a patch. Changing the
rule is a fix. Only the fix survives to next week.
Try this now
Rank this week's flags by amount. The top is OPP-007017, at 1,000,000 USD, in Solution
Review, no two-way customer activity since 2026-04-30, which is 94 days. By a wide margin the
largest at-risk deal in the book, and the first thing anybody will read.
Verify it. The flag makes two claims, and they are not equally true.
"No contact in 94 days" is true.activities.csv has nothing on this opportunity after
2026-04-30.
"1,000,000 USD" is false. The snapshot from 2026-07-27, six days before the run date,
records the same deal at 100,000 USD. The current opportunity row carries an extra zero.
Is it a real increase or an error? You can tell without asking anybody: a jump from 100,000 to
1,000,000 would show in your movement table as an amount increased row of +900,000 USD. It does
not. No week in the 26-week snapshot history shows this deal growing. The number changed in the
opportunity record without ever moving in the pipeline.
answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab,
confirms it under the class implausible-amount: "Amount 1000000.0 looks
like a fat finger extra zero (plausible at 100000.0 USD)". A second row in the same class,
OPP-007511, is recorded at 570,000 EUR against a last snapshot of 57,000 EUR.
Where the fix goes. Not in the pack. Add a rule:
R7 amount-plausibility
Compare opportunities.amount against the most recent pipeline_snapshots.amount
for the same opportunity_id.
If they differ by more than 2x and there is no corresponding amount-change row
in the movement table for any week, do not rank the deal.
Route it to exceptions.csv with reason "amount disagrees with last snapshot,
no movement row", and include it in the pack under Data Quality, not under
Risk.
Evidence line: R7 | opportunities.amount vs snapshot <date> | <a> vs <b>
Re-run. OPP-007017 leaves the risk list and appears in a two-line data quality section, where
a manager can act on it: send it to the deal owner to fix. Note what you did not do: delete it,
correct the number yourself, or quietly rank it at 100,000.
Reconcile the roll-up, and rebuild one number by hand
Two checks, every week, both cheap.
The roll-up reconciliation. The sum of deal amounts in your pack must equal the sum in the
source for the same filter. "For the same filter" is the load-bearing phrase:
Open pipeline from opportunities.csv on 2026-08-02: 62 deals.
Open pipeline from the last snapshot, 2026-07-27: 60 deals.
Both are right. Two deals, OPP-007239 and OPP-007268, were created on 2026-07-28, after the
last snapshot. The pack has to say which basis it used, or a manager comparing your number to
the CRM dashboard concludes one of you is wrong.
The two files also disagree on amount for 14 of the 60 shared deals, plus one where only the
currency label differs (OPP-006697, EUR in the snapshot and USD on the opportunity row): 15
deals differ on what they are worth. Most of that is legitimate rediscounting mid-cycle, with the
snapshot preserving the value as it stood. Two of the fourteen are the fat-finger defect from the
hunt above. A reconciliation that reports "15 deals differ" and lists them beats a single total.
The hand-rebuilt number. Pick one small number every week, vary which one, and reconstruct it
yourself before the pack goes out. Dagmar Wexford's book on the 2026-07-27 snapshot is 7 deals at
149,000 USD: a minute of adding up that either matches or does not. Record it in the run log. It
is the cheapest tripwire in the pack, and why you can say "I checked it, here is how".
Tell these apart: a signal, and an artifact of how the field is populated
This is the difference between a forecast input and a fact about a colleague's data entry habits.
An artifact
Group pipeline_snapshots.csv by owner and forecast category:
owner
rows in Commit
total rows
Commit share
Bettina Costa
43
87
49.4%
Sylvie Dekker
14
82
17.1%
Dagmar Wexford
23
171
13.5%
Bettina Farrow
29
217
13.4%
Aisling Calderon
26
219
11.9%
Renata Vasquez
19
180
10.6%
Franka Mertens
20
293
6.8%
Ursula Ulrich
8
134
6.0%
Across the whole file, about 13% of snapshot rows sit in Commit. One owner is at nearly half. It
reads like a signal about deal quality. It is a signal about a person's definition of a word.
Break the same rows down by the stage they were in when marked Commit:
Solution Review
Technical Validation
Proposal
Negotiation
Bettina Costa
11
10
6
16
Everyone else combined
0
0
0
139
Every Commit row belonging to any of the other seven owners is in Negotiation. All 139 of them.
One owner commits from Solution Review onward. "Commit" is not a measurement in this dataset: it
is a self-reported category, filled in by one person against a different definition from
everybody else, and any roll-up of committed pipeline is measuring that difference.
A signal
Same file, different question: how often does a deal's close date move? Count close-date changes
across the snapshot history and divide by the owner's open deals, because owners carry very
different numbers of deals and a raw count would rank Franka Mertens worst for having 34 of them.
owner
pushes
open deals
pushes per deal
Bettina Costa
7
8
0.88
Bettina Farrow
6
14
0.43
Aisling Calderon
6
18
0.33
Ursula Ulrich
4
12
0.33
Dagmar Wexford
3
19
0.16
Renata Vasquez
1
13
0.08
Franka Mertens
2
34
0.06
Sylvie Dekker
0
9
0.00
This is a signal. A close-date push is an observed change in a dated field between two weekly
photographs. Nobody chose a category, and it happened twice as often per deal in one book as in
the next one down. Still not a conclusion (a book of eight large enterprise deals will push more
than a book of thirty-four SMB deals), but it is evidence about deals, not about form filling.
The general test: ask what a row would have to do for the field to change. For a close date,
something has to happen to the deal. For a forecast category, somebody has to have an opinion.
What the wrong headline earns you
Suppose your pack's headline reads: "Commit coverage is strongest in Bettina Costa's book at 49%
against a team average of 13%, so her forecast is our most reliable." Here is the feedback that
sentence gets:
The number is correct and the sentence built on it is not. forecast_category is a field a
person sets, and your finding is that one person sets it differently from the other seven. The
stage breakdown says so unambiguously: all 139 Commit rows in the rest of the team sit in
Negotiation, while this owner's 43 are spread across Solution Review, Technical Validation,
Proposal and Negotiation. Half of that book is in Commit because of where the line is drawn,
not because those deals are closer to closing.
The direction is also backwards. On the one measure in this file that observes deals rather
than opinions, this book pushes close dates 0.88 times per open deal, the highest rate on the
team and roughly fourteen times the lowest. The pack recommended confidence in the forecast
that moves most.
What would have caught it: asking what would have to happen for this field to change. A close
date changes when something happens to the deal; a forecast category changes when somebody
changes their mind. A metric built on the second kind measures the people filling it in, and
must be reported per owner with the definition attached, or not at all.
What to change: in risk-rules.md, add "forecast_category is self-reported and is never used
as a risk signal without an accompanying stage distribution by owner". In the pack, replace the
headline with the push-rate table and one line naming the denominator, because the rate is only
meaningful next to the deal count.
If your team wants Commit to mean one thing, that is a management conversation, and a good one.
It is not something this pack can conclude on their behalf.
When this goes wrong
The break. Build the engagement gap rule the obvious way: contacts.csv has a
last_activity_date column sitting right there, so for each open deal take the most recent
last_activity_date among the account's contacts, and flag the deal if it is older than 45 days.
The tell. Take OPP-007409, a 212,000 EUR new business deal at Fairholm Facilities Ltd
(ACC-02908, Ireland, Enterprise), in Technical Validation for 101 days, owned by Renata
Vasquez. Three ways of asking the same question, three different answers:
Method
Answer
Verdict
Most recent contacts.last_activity_date at the account
2026-06-17, 46 days
flagged
Most recent row in activities.csv joined on related_id
2026-07-16, 17 days
not flagged
Most recent two-way customer contact
2026-06-17, 46 days
flagged
The second method is the one most people build, and on a 212,000 EUR deal it says everything is
fine. Open the rows it is reading:
The two most recent rows are internal notes. contact_id is blank for internal notes and
duration_minutes is 0 for emails and notes (DATA-DICTIONARY.md in your pack's data folder).
Nobody at Fairholm was on
the other end of either. The last time anybody spoke to the customer was 17 June, chasing a
signature.
The first method happens to give the right answer here, and it is still the wrong rule.
contacts.last_activity_date is a denormalised roll-up at contact level, not deal level: it
reports an account as engaged when the engagement was on a different opportunity entirely, and
as dormant when the deal is worked through a contact whose record has not been touched. Right
answer, wrong instrument, and you will not know which you have next week.
A second, quieter failure in the same join.PLANTED-ISSUES.md documents six rows in
activities.csv whose related_id resolves to no opportunity at all, class
broken-foreign-key. An inner join drops them without a word. One, ACT-0044746, is labelled
related_type = Account while carrying an opportunity-shaped related_id of OPP-906685, so a
filter on related_type misses it too. Neither the join nor the filter says anything is missing.
The guard.
R3 engagement-gap, corrected
Source: activities.csv joined on related_id where related_type = 'Opportunity'.
Count a row as customer contact only if contact_id is populated
AND activity_type is not 'Note'
AND outcome is not in ('No answer', 'Bounced').
Flag when the most recent qualifying row is older than 45 days on a deal above
50,000 USD.
Evidence line: R3 | last qualifying activity on <opportunity_id> | <date>, <n> days
Join integrity, reported every run, not optional:
count of activities.csv rows whose related_id does not resolve, and the count
of rows excluded by each of the three conditions above. If the excluded count
is more than 30% of the rows on a deal, name that deal in the pack.
The join integrity line is the part people skip. Reporting how many rows your filter threw away
is how you find out that the filter is the problem.
The five-minute pack, and the section that buys you trust
You assemble the pack yourself; a seed layout ships in templates/M6/review-pack.md of your
practice pack. Structure, in this order:
1. Three lines: what changed. Written as sentences, not numbers.
2. The movement table with its two reconciliation lines.
3. Top flagged deals as cards, most exposure first.
4. Data quality: anything R7 or the join integrity check caught.
5. Exclusions. What this pack does not cover, and why.
The exclusions section is a credibility instrument. A pack that says what it does not cover gets
caught out less, and stays believed. A real one for this dataset:
This pack does not judge whether any deal will close.
Movement is computed from weekly snapshots only, which start 2026-02-02. Older
weeks have aged out of the warehouse, so there is no year-to-date figure here.
The week of 2026-04-20 is excluded: the snapshot is a duplicate of 2026-04-13,
so 2026-04-27 contains two weeks of movement.
Open pipeline is stated on the 2026-07-27 snapshot basis, 60 deals. The CRM shows
62 today because two deals were created on 2026-07-28.
Engagement is measured from logged activity only. Conversations nobody logged do
not appear, in either direction.
forecast_category is self-reported and is not used as a risk signal here.
Two deals are excluded from the risk ranking under R7 pending an amount
correction: OPP-007017 and OPP-007511.
Ninety seconds to write, and the part a sceptical manager reads first.
What stays human, in words you can say out loud
Hold the line on four things, and here is the sentence for each:
The forecast call. "Nothing in this pack can be checked against an outcome until the
quarter closes, which means there is no way to know whether it was right in time for it to
matter. That is a veto under our own fit test, so the call stays mine."
The decision to inspect a rep. "The pack flags deals. Deciding that a pattern is a coaching
problem needs context that is not in any file the agent can read."
The decision to discount. "It touches money and it is not reversible by deleting a file."
Whether a champion is real. "The record shows who replied to emails. It does not show who
will spend political capital for us."
Two product facts sit under all four, worth naming to your manager once: nothing fires when a
stage changes, so the pack is a weekly cadence and a deal that goes wrong on Tuesday is caught
the following Monday; and there is no memory between sessions, so anything next week's run needs
has to be in the rules file.
What you keep
20-briefs/pipeline-review.brief.md, the brief you started in L6.2, complete.
40-reference/risk-rules.md, the file you created in L6.3, now at version 2, with R7 added
and R3 corrected. Keep the version
history, one line per change naming the false positive that caused it. That history is the part
of your portfolio that shows judgment.
30-output/pipeline/review-YYYY-WW.md with an exclusions section.
A 50-runs/RUN-LOG.md entry with the reconciliation result and the number you rebuilt by hand.
Check yourself
You find a flag that is wrong. Deleting it takes ten seconds. Changing the rule takes fifteen
minutes and will re-flag three deals that were previously clean. Which do you do, and what do
you tell your manager about the three new flags?
Your manager asks for a predicted close probability per deal. Answer using the fit test, and
offer what you can give instead.
Your pack says 62 open deals. The CRM dashboard says 60. Both are correct. Write the line that
goes in the exclusions section.
A colleague's risk rule flags deals with "no activity in 30 days" and their pack has almost no
flags. Name the two questions you would ask about the rule before you believe the pack.
Where this goes next
The exercise sheet in labs/LAB-06 of your practice pack (start with its README) runs this on the
shipped dataset, then on two exports of your own book a week apart, behind the redaction gate,
with your own measured before and after.
M7 moves from the deals you have to the accounts you do not, where the failure mode is not a
wrong number but a claim that was never true.
Module 6 · Pipeline inspection and forecast inputs
Lab / ? min / read-only
Lab
The week that did not happen
LAB-06: The week that did not happen
The situation
It is Monday morning and your manager asks the same two questions every week: what changed, and what is
at risk. You rebuild the same view by hand every week, it eats most of a morning, and by Thursday it is
out of date.
Worse, the risk answer is currently vibes. "Feels solid" is not a forecast input, and when a deal slips
nobody can say which signal was there and ignored.
The shipped Talvera export gives you the same problem at a size you can check: 528 opportunity rows
across 524 distinct opportunity IDs, and 1,383 weekly pipeline snapshots taken every Monday from
2026-02-02 to 2026-07-27, which is 26 weeks.
One of those 26 weeks shows no movement at all. A quiet week is plausible, which is exactly why nobody
questions it, and it is the reason this lab is called what it is called.
What you will produce
File
What it is
40-reference/risk-rules.md
Named rules, thresholds, and the evidence line each one emits. Versioned, because the rules change when you find a false positive
20-briefs/pipeline-review.brief.md
The weekly brief, rerunnable
30-output/pipeline/review-2026-W30.md
The five-minute pack: three lines of what changed, the movement table, the top flagged deals, and a stated exclusions section
30-output/pipeline/deal-cards/<opportunity-id>.md
One card per flagged deal, each ending in "what the data does not tell you"
50-runs/RUN-LOG.md
Updated, including the reconciliation result and the number you reconstructed by hand
Outcome accounting
Manual baseline: 47 minutes
Measured. The lab author built the movement table for one week pair and five
deal cards by hand in a spreadsheet: load both snapshot weeks, join on
opportunity_id, classify each change type, total the amounts, then pull
activities and notes for five deals. Timed once.
This is ONE week pair on 60 open deals. On a real book of 200 it is a morning,
which is what the module scenario describes.
Target with Cowork: Run one 50 minutes, observed. Of that, about 20 minutes is writing
risk-rules.md, which you do once.
Run two 16 minutes, observed, including the 9-minute verification below.
Run two is the honest number, because run one includes writing the rules.
MODULES/06 says to measure the target on run two and this lab does.
Verification cost: 9 minutes, counted separately and inside every number above.
Reconciliation (opening + changes = closing) 3 min
One number reconstructed by hand from source 2 min
Sampling 5 flagged deals against activities and notes 4 min
Net, run one: 47 - 50 = the method LOSES 3 minutes on the first week.
Net, run two onward: 47 - 16 = 31 minutes saved, every week.
Break-even: week two.
Over a quarter: 13 weeks. 47 x 13 = 611 minutes by hand against 50 + (12 x 16) = 242.
Saves 369 minutes, about 6 hours, with the checking included.
Your actual: _____ your own current preparation time, measured ONCE before you start
_____ run one _____ run two _____ verification
Measure your own baseline before you start. The only number worth putting in your capstone is the one
you measured on your own book.
Blast radius
Read-only. The entire pack reads files and writes into 30-output/. Nothing in this module writes to
a CRM.
Obligation
This lab
Review gate
Spot check on a sample: five flagged deals against source
Approver
None needed
Reversal plan
Delete the output folder
Any temptation to write next steps back into the system is a Module 5 promotion decision with a reversal
plan, and it is not this lab.
broken-etl, period-attribution, derived-field-drift, note-crm-contradiction (4 of the 6)
In-scope planted rows
62
Key version
20260801
Per-class counts are in SCORING.md §1.1. Two notes on this scope:
activities.csv and notes/call-notes/ are in the file scope, because two of the declared
classes live only there: six of the eight in-scope broken-foreign-key rows are activity rows whose
related_id resolves to no opportunity, and note-crm-contradiction is a note read against a CRM row.
Without those two files both classes would be declared with zero rows.
This lab carries four of the six hard classes, which is more than any other. broken-etl and
note-crm-contradiction are the discriminating catches: the first needs a snapshot integrity check
that nothing prompts you to run, and the second needs you to read the notes against the CRM rows
rather than trusting either alone.
Hard-class coverage is a count of classes, accumulated across the four scored core labs, and the bar
is at least 3 of the 6 (ASSESSMENT-MODEL.md §7). Coverage requires an assertive identification: an
item routed to exceptions.csv earns half recall credit and does not earn class coverage
(SCORING.md §7).
You do not hold PLANTED-ISSUES.md while attempting this lab.
How you verify your own output
The movement reconciliation. Opening pipeline plus changes equals closing pipeline, in count and
in amount. This is not decoration. It is the check that catches a snapshot join done wrong, which is
the most common silent failure in this workflow.
One number reconstructed by hand, chosen at random, before the pack goes to anybody. Every week,
a different one. This is the habit, not the check.
Sample five flagged deals against activities.csv and the call notes. You are looking for a flag
that is wrong, and there are wrong flags to find.
Every flag has an evidence line. Rule name, the field or record that triggered it, and the value.
A risk output without evidence lines is rejected in this module and fails the rubric, because a
flag you cannot argue with is a flag a rep will simply ignore.
Trust ordering, and it is not a detail
The three inputs are not equally true, and an agent that treats them as though they were will produce
confident wrong flags.
Input
What it is
What it is not
pipeline_snapshots.csv
Facts about what the record said on a date
Facts about the deal
activities.csv
Facts about logging behaviour
Facts about the relationship. An unlogged call is indistinguishable from silence
notes/call-notes/
Unstructured, written for a human reader rather than a parser
Reliable. One of the 16 notes flatly contradicts the CRM row it describes
Files you need
pipeline_snapshots.csv, opportunities.csv, activities.csv, contacts.csv, accounts.csv,
notes/call-notes/ and DATA-DICTIONARY.md, in 10-data/talvera/, plus the workspace and the brief
method from LAB-04.
What comes next
LAB-08 applies this same pattern to the renewal book, which is the same shape of work against a
different object. TL-SELL-2 builds its deliverable on the note contradiction you meet here.
Module 6 · Pipeline inspection and forecast inputs
6 questions / pass at 70%
Checkpoint
Module 6 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
A week-over-week movement analysis, built by joining two consecutive snapshot weeks.
Movement, week ending 2026-06-15
Opening open pipeline 142 deals 8,914,000
created +11 deals + 604,000
advanced a stage 9 deals (no count or amount change)
slipped a stage 2 deals (no count or amount change)
close date pushed 14 deals (no count or amount change)
amount increased 0 deals + 0
amount decreased 3 deals - 118,000
Closing open pipeline 147 deals 8,588,000
Reconciliation
counts: 142 + 11 = 153 against a closing count of 147. Difference: 6 deals.
amounts: 8,914,000 + 604,000 - 118,000 = 9,400,000
against a closing amount of 8,588,000. Difference: 812,000.
Both lines fail. What is the most likely cause, and which check catches it?
Hint
List every way an opportunity can be in the opening set and not in the closing set. The movement table
has a category for all of them but one.
A. Incorrect. Mixed currency is a genuine hazard in this data and it explains only half of what you are looking at. Summing across currencies distorts amounts and cannot change a deal count, so it leaves the six-deal gap unexplained and you would still be looking for a second cause. Prefer the explanation that accounts for both failed lines with one mechanism, and check whether the amount gap is the value of the missing deals before reaching for a conversion error.
B. Incorrect. Changing the join makes the six rows reappear with null closing values, which is a useful debugging step and is not a diagnosis. You still have to say what those rows mean, and the answer is that they closed and left the file. Switching join types without adding a category for the result gives you six rows of nulls, an amount line still out by 812,000, and nothing to put in the movement table.
C. Correct. The table has a category for everything that changed and none for everything that left, and a snapshot file where closed deals age out the following week will always produce that gap. One cause explains both failed lines, which is the test of a diagnosis: six deals closed, taking 812,000 of open pipeline with them, so the count is out by six and the amount by their total. The six are not lost, they are unclassified. The check is a reconciliation that forces every opportunity in the opening snapshot to be accounted for on the closing side, including by having left, with the unexplained bucket required to be empty. That requirement is what turns a reconciliation from decoration into a check.
D. Incorrect. The opening figure comes from a snapshot, which is the correct source for movement, because it is a photograph of what pipeline looked like on that Monday. Recomputing it from the current opportunity file replaces a fact about the past with a fact about today, which is the mistake the snapshot file exists to prevent.
Question 2 of 6
Two rows from a risk output.
Illustrative excerpt. The records below are invented for this item and are not rows from the Talvera practice dataset.
OPP-2291 Grantham Foods expansion Risk score: 78/100 Priority: High
OPP-2318 Ravensworth Care renewal Risk score: 71/100 Priority: High
Why is this unusable, and what is the minimum fix?
Hint
Take one row and try to disagree with it out loud. Write down what you would need in order to say
something more specific than "I do not think that is right".
A. Incorrect. "The detail can be looked up in the CRM" is the sentence that undoes the workflow. If checking every flag means opening the record, you have moved the work rather than saved it, and you have added a ranking nobody can validate. Verification has to be cheaper than the work, and this design makes it more expensive.
B. Incorrect. The gap between 78 and 71 is not the problem, and recalibrating the scale would make it worse by adding precision to a number with nothing behind it. A composite score with no components is the trap: it looks quantitative and it cannot be checked, which is the same defect as an unexplained duplicate match score.
C. Incorrect. A recommended action is not what this output owes. Deciding what to do about a risk is judgment work that depends on the relationship, and the module keeps it human on purpose. What the output owes is the evidence that lets the human decide quickly.
D. Correct. A risk output without evidence lines is rejected in this module and fails the rubric, and the reason is operational rather than stylistic. An AE who disagrees with a flag has no way to say what is wrong with it, so they learn to ignore the list, and a list people ignore is worse than no list because it consumed the time to build. The evidence line is three things: which rule, which record or field, and the value.
Question 3 of 6
Your inactivity rule flagged a deal as dormant. You check, and the account has been engaged all month:
the conversation moved to a different contact at the same account, and the person your rule was watching
has not been involved since April. The rule was built on the contact-level last_activity_date field.
Three responses are proposed. Which is correct?
Hint
Ask which of the four responses would change what happens next Monday if you did nothing else. Only one
of them does.
A. Correct. You changed the rule rather than the output, and you named the specific defect: the rule was reading a field that measures one person's activity and using it as a proxy for a deal's activity. The reconciliation clause is the part most people miss. Once you join activities to opportunities you inherit whatever rows do not resolve, and a naive inner join drops them without comment, which trades one silent error for another.
B. Incorrect. This makes today's pack correct and guarantees the same flag next Monday, and the Monday after. It is also the most common thing people actually do, which is why the module names it: editing output is a patch, editing the rule is a fix, and only the fix survives.
C. Incorrect. An exception list for a rule you know is wrong is a growing maintenance liability that hides the defect rather than removing it. Every week you add another deal, the rule keeps producing wrong flags on deals nobody has caught yet, and the exception list becomes an undocumented second rule that nobody wrote down.
D. Incorrect. This is honest and it keeps the error. A flag that ships with a note explaining why it is wrong is a flag that should not have been raised, and the reader now has to read two things to learn nothing. Explaining a false positive is not the same as removing the cause of it.
Question 4 of 6
Two versions of the same weekly pipeline pack are on your manager's desk.
Pack A. Movement table, top ten flagged deals as cards, an executive summary, a stage conversion
chart and a forecast confidence indicator. No statement about what it covers.
Pack B. Movement table, top ten flagged deals as cards, three lines of what changed, and this
section:
What this pack does not cover. It does not judge whether a deal will close. It excludes
opportunities below 5,000 in account currency (41 deals, 2.1% of open amount). It uses logged activity
only, so engagement in channels we do not log is invisible to it. One snapshot week is excluded from
the movement figures and the reason is stated in the movement table footnote.
Which is more trustworthy, and why?
Hint
Assume both packs have entirely correct numbers. Then ask what a busy reader would wrongly conclude from
each one.
A. Correct. An exclusions section is a credibility instrument rather than a disclaimer. It converts the reader's uncertainty about scope into a stated fact, and it does the harder thing of quantifying the exclusion (41 deals, 2.1% of open amount) so the reader can judge whether it matters. You also caught the forecast confidence indicator, which is the real defect in Pack A: it is a number about something the pack cannot know, sitting next to numbers the pack does know, which is how an unearned judgment borrows credibility from an earned one.
B. Incorrect. Pack A has more content and less information about itself, and the difference matters the first time somebody acts on it. "A manager can ask about gaps" assumes the manager knows a gap exists, and the whole problem with an unstated exclusion is that it is invisible. Silence about scope reads as full coverage to every reader who has not built the pack themselves.
C. Incorrect. You chose the right pack for a reason that would lead you wrong next time. Length is not the variable. A short pack with no exclusions section would be worse than either of these, and a long pack with a rigorous one would be better than both. What is being scored is whether the document states what it does not do.
D. Incorrect. Correct numbers are necessary and they are not what this item is about. Two packs with identical correct numbers still differ in what a reader is entitled to conclude from them, and the pack that does not state its exclusions invites conclusions its numbers do not support. Trustworthy means "a reader who relies on this will not be misled", which is a property of the framing as well as the arithmetic.
Question 5 of 6
A deal card as produced.
Illustrative excerpt. The records below are invented for this item and are not rows from the Talvera practice dataset.
OPP-2291 Grantham Foods, expansion
Stage: Technical Validation Amount: 96,000 EUR
Age in stage: 141 days Days since last logged activity: 38
Next step: (empty) Owner: R. Aldbury
Flags
stage-stagnation: 141 days in Technical Validation against a Mid-Market norm of about 32 days
next-step-missing: next_step is empty
no-activity-30d: last logged activity 2026-06-29
The card has no "what the data does not tell you" line. What risk does this create when a manager reads
it?
Hint
Read the card as a manager who has thirty seconds and no other context, and write down the sentence you
would say out loud in the deal review. Then check whether the data supports that sentence.
A. Incorrect. "Accurate" and "not misleading" are different standards, and a card that meets the first and not the second is the most dangerous artifact in this module, because nothing on it can be challenged line by line. The honesty line does not weaken the card. It is what allows the rest of the card to be taken at face value.
B. Incorrect. Priority ranking is not what the card is for, and adding one would make the problem worse by putting a judgment on top of unqualified observations. The card exists so a human can form the judgment, which is why the flags carry their evidence and why the unknowns need naming.
C. Incorrect. This is a real consequence and it is downstream of the actual defect. The reason the owner is exposed to an unfair inference is that the card presents logged activity as though it were all activity. Fix the framing and the unfairness goes with it. Treating the fairness problem as the primary issue leads to softening the flags, which removes the useful part.
D. Correct. Every line on this card is true and the card as a whole says something that is not. "Days since last logged activity" is a measurement of logging behaviour, and without the honesty line it is read as a measurement of the relationship. The line is a gate against the card being read as a judgment: it names the three or four things this data structurally cannot see, so the manager arrives at the deal review with the facts and the assumptions separated.
Question 6 of 6
Four tasks inside the weekly review. Two of them must stay human.
Produce the week-over-week movement table with a reconciliation.
Decide whether the quarter is called at commit.
Flag deals that meet the written risk rules, with the evidence line for each.
Decide whether the champion on the largest open deal is real.
Which two stay human, and which fit-test factor vetoes them?
Hint
Score all four on Checkability before anything else, and be strict about what counts: a source of truth
is a record you can open, not a person you could ask.
A. Incorrect. Task 3 is the trap. Applying written risk rules and emitting the evidence line for each is exactly the work to hand over: the rules are yours, the output is checkable against the records that fired them, and the volume is real. Judgment density is also the wrong factor for task 2: the reason the commit call stays human is not that it is hard, it is that no output of it could be shown to be right or wrong from a file.
B. Incorrect. Cost of being wrong is high for both, and it is not a veto: high cost is managed with a review gate, a named approver and a reversal plan, which is how every draft-only workflow in this course works. If cost alone vetoed hand-over, nothing that mattered could ever be delegated. The factor that admits no mitigation is the one you want.
C. Correct. You picked the right two and the right factor. The commit call and the champion judgment both fail the same way: there is no record you could sample against to find out whether the output was right, which makes the Check letter impossible and vetoes the hand-over whatever the volume or the specification cost look like. This is also the form of the answer you can repeat to your manager, which is what O6.6 asks for.
D. Incorrect. "A human approves the final number" is a review gate, and a review gate does not create checkability. The approver would be approving a judgment they cannot verify, produced by a process with no source of truth behind it, which means the gate is a signature rather than a check. That is the shape of the failure the fit test is designed to catch.
M7: Demand and targeting, research, ICP, lists and outbound assets
Last verified: 2026-08-06.
Product claims trace to course/reference/COWORK-FACTS.md (FACTS §x.y). Dataset claims trace to
data/DATA-DICTIONARY.md and answer-keys/PLANTED-ISSUES.md in your practice pack for the Talvera
dataset generated with seed 20260801.
The module at a glance
There are 1,209 leads in the shipped dataset and no scoring anybody can explain. An SDR spends most
of a working block on account research that mostly recovers facts already sitting in your own CRM. A
campaign brief takes long enough that the first version gets thrown away, because the segment
definition was ambiguous.
And there is a danger here that does not exist in Modules 5 and 6. Hygiene work and pipeline work go
wrong in ways you can catch by reconciling against a source. Research goes wrong in a way that reads
beautifully: a funding round that did not happen, a tech stack nobody confirmed, a customer's own
unaudited number restated as something you measured. That sentence then goes into an email to a real
prospect with your name on it.
This module spends a disproportionate share of its time on that problem, because that is where the
work actually fails.
Lesson
Title
TRACK
Minutes
Module-level piece
L7.1
Design the check before you design the work
T, R
11
L7.2
ICP as testable rules, with the reason attached
A
11
Worked exercise
L7.3
The research brief, and the fabrication hunt
A, C
13
Failure lane
L7.4
Lists that reconcile, drafts that get reviewed
C, A, K
13
Discrimination pair
Total lesson time: 48 minutes. Module budget is 45 (PEDAGOGY.md §3.1), tolerance plus or minus
15%.
Then: checkpoint CP-07, taken as the module quiz in the
interactive app (about 10 minutes, and it carries the highest
weight in the applications track because this failure mode reaches customers), and LAB-07, the
exercise sheet in labs/LAB-07 of your practice pack (start with its README; Part A 50 minutes,
Part B Full 40 or Part B Lite 15).
Outcomes
O7.1 Write an ICP definition as testable rules, apply it to score accounts and leads, and make
the reason for every score visible in the output.
O7.2 Produce an account research brief a seller can act on, in which every claim carries a
source and every unknown is marked unknown rather than filled in.
O7.3 Build a segment list with stated inclusion and exclusion rules and a count reconciliation
that proves nothing was silently dropped.
O7.4 Produce draft-only outbound assets with the review gate stated and enforced.
O7.5 Detect unsupported claims in generated research using a repeatable procedure, and change
the brief so the claim class cannot recur.
O7.6 State the review gate that applies before any generated asset reaches a prospect,
including who reviews what.
The artifact you leave with
The targeting pack, in your own revenue-workspace/:
File
What it is
Built in
40-reference/icp-rules.md
Weighted, testable ICP rules with tier bands and the reason-string requirement
L7.2
40-reference/suppression-rules.md
Exclusion rules that apply to any list, ever
L7.4
40-reference/review-gate.md
Who reviews what before an asset leaves the folder
L7.4
20-briefs/account-research.brief.md
The research brief, with source discipline built in
Templates are in templates/M7/ of your practice pack.
icp-rules.md is the artifact a marketing and sales leadership pair argue over once and then use for
a year. review-gate.md is the document that makes a manager comfortable letting a team do this at
all. Both are portable to any tool.
Blast radius
Draft-only, everything. Every asset produced in this module is written by an agent and sent by a
human. Nothing generated here reaches a prospect without a person reading it, and the review gate
names who that person is for each kind of claim.
The one thing that is not merely draft-only in its consequences is the list. A segment that includes
an opted-out contact is a compliance problem the moment somebody presses send, which is why
suppression is scored as Safety in LAB-07 and not as a completeness deduction.
Prerequisites
Modules 1 to 4. Module 5 helps: scoring a duplicated account set double counts, and the shipped file
has eight documented duplicate accounts.
What this unlocks
The AE/SDR and GTM marketing capstones, and Module 8's expansion work, which reuses the account
brief pattern against existing customers.
A note on browser research
Live web research is available conditionally through Claude in Chrome, a Chrome extension in beta on
all paid plans that lets Claude read pages, click, type, fill forms and move between tabs, with tasks
launchable from the Chrome side panel or from Cowork or Claude Code (FACTS §5.9). It draws on the
same account usage allowance as everything else (FACTS §5.9, §6).
It is taught here as an option with its own risks, never as a requirement. Every lab in this module
is completable without it, and the external-source half of the account brief can be filled by hand
or left explicitly unknown. That is not a limitation of the course. Half of a good account brief is
recovering what your own company already knows and never wrote down.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
M7 lesson / about 11 min
Concept
Design the check before you design the work
Do this nowIndex the notes folder60s
Point a session at 10-data/ and ask it to list every file under notes/.
For each file ask for the filename and the account_id or opportunity_id values it mentions.
Do not ask for a summary of any of them.
You know it worked when: Nineteen files come back, sixteen under call-notes/ and three under qbr/, and one of them is named "vireo-retail renewal - notes.txt" with spaces in it.
What you type
Produce a short account brief for ACC-03089, Quarrybank Hospitality Group SL.
PERMITTED SOURCES
10-data/ (all files). Web research: not enabled.
Every claim carries a "Source:" line naming the file and the record id.
A claim with no source is not written as prose. It goes to the Unknown section
as a question naming who could answer it.
Do not use general knowledge about this company, its industry or its people.
What comes back
30-output/demand/account-briefs/ACC-03089.md
SMB, Hospitality, Spain, EMEA, 181 employees. Customer. ARR 27,000 USD, billing
currency EUR, Tier 3. Health score 86. Renewal 2027-06-24.
Source: accounts.csv, ACC-03089.
New business closed won on 2025-06-24, amount 25,000, currency label USD, owner
Franka Mertens.
Source: opportunities.csv, OPP-007089.
## Unknown
- Which currency OPP-007089 is actually in. The row says USD, the account bills
in EUR, and 25,000 at 1.08 is 27,000, which is the account's arr_usd exactly.
Source for the ambiguity: opportunities.csv OPP-007089 and accounts.csv
ACC-03089. Franka Mertens owns both records and could settle it.
Notice: Not one line is a paragraph about a hospitality company. Every claim names a file and a record, which is the only reason the Unknown section could exist at all.
Why this matters
An SDR on your team spends forty minutes researching an account before a first call. Most of what
they find is already in your CRM, in a call note nobody re-read, or in a support ticket from
February. The rest is from the web, some of it three years old.
You are about to hand that job to an agent. Pause first. Run it like the hygiene pass and you will
produce something worse, faster, at higher volume.
Everything this lesson uses came in the practice pack from the course page; if you have not
downloaded it, do that first.
What you will be able to do
Score account research on the fit test and name the factor that nearly disqualifies it.
Explain the one decision in the brief that makes it safe to run.
Say what counts as a source in your workspace, and what does not.
The factor that nearly kills this workflow
Factor
Account research
Why
Volume
5
Every account, every quarter, and again before every first meeting
Judgment density
3
Assembling is mechanical. Deciding what matters is not.
Cost of being wrong
5
It goes to a customer, with your name on it
Specification cost
3
An hour to write the brief properly
Checkability
1, unless you force it
A generated paragraph about a company has nothing to check it against
A low Checkability score is a veto: by the fit test, this job should not be handed over at all. It
passes for one reason: the brief manufactures checkability. Every claim carries a source line
naming where it came from; anything without one is written as "unknown". That turns an unverifiable
paragraph into statements you can each check in under a minute.
The move generalises, and it is worth more than the workflow: when a task fails on Checkability,
do not abandon it and do not ignore the score. Change the output shape until it is checkable. You
are not making the agent more accurate. You are making its output falsifiable.
What counts as a source
Decide up front what the agent may draw on, and name it in the brief. In order of trust:
Your own supplied data. Here: accounts, contacts, opportunities, activities, product usage,
support tickets, call notes, QBR documents. Highest trust, usually half of a good brief. Cowork
reads, creates and modifies files only inside the folders you grant, so 10-data/ is exactly the
universe.
Documents you explicitly provide. A press release the customer sent, an RFP, a security
questionnaire.
Live web research, if you enable it. More below.
Connectors, if authorised. Connectors are MCP servers wired to your tools. Gmail, Google
Drive, Slack and a Zoom connector are the ones we can name; check what yours can actually read
before you list it as a source.
Not a source, and this is the important line: the model's own recollection of a company. It will
produce fluent, specific, plausible statements from nothing at all, and they sound exactly like the
ones from your CRM.
The rule goes in the brief verbatim:
PERMITTED SOURCES
10-data/ (all files), and documents in 15-provided/.
<web research: enabled | not enabled>
Every claim in the output carries a source line: file or URL, plus the record id
or the section.
A claim with no source is not written as prose. It goes in the "unknown" section.
Do not use general knowledge about any company, product or person.
Browser research: an option with a cost
Claude in Chrome is a Chrome extension that lets Claude read pages, click, type, fill forms and move
between tabs. It is in beta, free to install, available on all paid plans. Tasks launch from the
Chrome side panel or from Cowork or Claude Code. It is officially supported on Chrome and may work
on Chromium browsers such as Edge and Brave. It draws on your account's usage allowance, one pool
shared across chat, Claude Code and Cowork.
Three honest things:
It changes the failure mode, it does not remove it. A sourced claim with a URL is checkable, a
real improvement. It is not the same as a true claim. Pages are stale, republished and wrong.
It is a beta with its own setup. Everything in this module and in LAB-07, the exercise
sheet in labs/LAB-07 of your practice pack (start with its README), works without it.
No browser? Fill the external-context section by hand, or mark it unknown. That is a legitimate
finished state.
It costs allowance, from the same pool as everything else. Research across forty accounts is a
triage decision first.
If you use it, one extra rule in the brief: the source line for a web claim is the URL and the
date it was read. A URL with no date is a claim about a page that may no longer say that.
Try this now
Set up the research workspace, and make each choice on purpose.
Put the CRM exports plus notes/, all from the data/ folder of your practice pack, into
10-data/. Include notes/call-notes/ and notes/qbr/:
the highest-value internal facts live there, unread since they were written. Leave out finance
files and anything with pricing approval: whatever is in the folder can end up quoted in a draft
that goes to a prospect.
Run the cheapest possible probe before writing any brief:
List every file in 10-data/notes/. For each: filename, size in lines, and the
first line verbatim. Then report which account_id or opportunity_id values
appear in each file. Do not summarise the contents. Do not write any file.
What you should see: an index of sixteen call notes and three QBR documents mapped to the
accounts they concern. You will need it in L7.3. It also shows what a summary would hide:
inconsistent filenames, one with spaces, .txt and .md mixed (documented in
data/DATA-DICTIONARY.md in your practice pack). If it starts
summarising instead, say: "Stop. Do not summarise. Only list filenames, line counts, first lines,
and the ids in each file."
Worth knowing: a file missing from that index is missing from every brief you write afterwards.
Better to find out now than in a meeting.
What you keep
Two blocks at the top of 40-reference/icp-rules.md, which you start in the next lesson:
Permitted sources, by name, with web research marked enabled or not enabled.
The source rule: every claim carries a source line; anything without one goes to "unknown".
Check yourself
Your manager says research is "low risk because it is only a draft". Answer using the two factors
in the fit test that actually apply here.
An account brief contains: "They recently expanded into the Nordics." What is the minimum you
need before that sentence is allowed to exist, and what happens to it if you cannot get it?
You have browser research enabled. A brief comes back with a claim sourced to a URL. Name two
ways it can still be wrong.
Where this goes next
L7.2 writes the ICP as rules you can test and argue with, not adjectives.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
M7 lesson / about 11 min
Do it
ICP as testable rules, with the reason attached
Do this nowScore one account by hand90s
Open accounts.csv and read segment, region, industry and employee_count for ACC-02831.
Add 10 if opportunities.csv holds a Closed Lost row on that account, and name it.
You know it worked when: Westmoor Logistics GmbH comes out at 90, Tier A, and the last +10 traces to OPP-007384.
What you type
OUTCOME
Score every account against the ICP rules and output one row per account with
the score, the tier and the reason string.
INPUTS
10-data/accounts.csv, 10-data/opportunities.csv
Rules: 40-reference/icp-rules.md, section N (normalisation) and I1 to I5, tier
bands in section T.
OUTPUT SHAPE
30-output/demand/icp-scores.csv, columns exactly:
account_id, account_name, score, tier, reason, suppressed, suppression_reason
reason is one string of the form "I1 <condition> +<n> | I2 ... ", naming every
rule including the ones that contributed zero.
Then, printed in the chat and not written to file: the rule fire-rate table over
prospect accounts with no open opportunity.
CONSTRAINTS
Apply the normalisation in section N before any string comparison.
Scoring and suppression are separate steps: score every account, then set
suppressed = TRUE where a suppression rule applies. Do not remove suppressed
rows from the file.
DEFINITION OF DONE
Every account appears exactly once. Row count in equals row count out.
What comes back
30-output/demand/icp-scores.csv written. 308 rows in, 308 rows out.
account_id,account_name,score,tier,reason,suppressed
ACC-02831,Westmoor Logistics GmbH,90,A,I1 Mid-Market +30 | I2 EMEA +15 | I3 Logistics +20 | I4 617 employees +15 | I5 prior Closed Lost OPP-007384 +10,FALSE
ACC-02995,Selby Supply Chain BV,80,A,I1 Mid-Market +30 | I2 EMEA +15 | I3 Logistics +20 | I4 1154 employees +15 | I5 no prior loss +0,TRUE
ACC-02885,Ardleigh Works BV,70,A,I1 Mid-Market +30 | I2 EMEA +15 | I3 Manufacturing +0 | I4 1385 employees +15 | I5 prior Closed Lost OPP-007295 +10,FALSE
ACC-03112,Ardleigh Industries Inc,20,C,I1 SMB +5 | I2 NA +5 | I3 Manufacturing +0 | I4 138 employees +0 | I5 prior Closed Lost OPP-007338 +10,FALSE
Rule fire rate, prospect accounts with no open opportunity (163):
I5 prior Closed Lost 154 of 163 94.5%
Notice: I5 fires on 94.5 percent of the pool, so it adds ten points to nearly everything and changes no ranking, and the reason string is the only thing that made that visible.
Why this matters
Ask five people to define the ideal customer profile and you get five sentences that all sound
right. "Mid-market operations-heavy businesses with a growing hourly workforce." Nobody disagrees,
because nobody can.
Then somebody builds a target list from it, sales says the list is bad, marketing says it is fine,
and the argument runs for a quarter, because there is nothing either side can point at.
What you will be able to do
Turn an adjectival ICP into weighted rules on fields that exist in your data.
Require a reason string on every score, and say why a score without one cannot be improved.
Spot a rule that fires on almost everything, and delete it.
Rules on fields that exist
"Growing sales team" is not a rule: there is no column called growing. Write conditions on fields
you have. Here, accounts.csv (from the data/ folder of your practice pack, copied into
10-data/ in L7.1) gives you segment, employee_count, industry, country,
region, lifecycle_status and primary_lead_source; opportunities.csv gives you history;
product_usage.csv and support_tickets.csv give you behaviour on accounts you already serve.
One warning first. accounts.region was derived from country at creation and never dirtied: the
reliable geography field. accounts.country and accounts.industry are not clean everywhere
(data/DATA-DICTIONARY.md in your practice pack): the file holds united states, DE, netherlands, The Netherlands,
Retail, logistics and Hospitality & Leisure. Match industry = 'Retail' exactly and you
silently drop ACC-02948, recorded as ' Retail' with a leading space.
So the rule file has a normalisation section before it has rules, and the brief applies it. Same
discipline as the match rules in Module 5: a comparison is only as good as the strings going in.
Weights, tiers, and the reason string
Give each rule a weight, sum to a score, and cut the score into named tiers. Then the requirement
that makes it auditable:
Every scored record carries a reason string listing which rules fired and what each contributed.
ACC-02831, score 90, Tier A
reason: I1 segment Mid-Market +30 | I2 region EMEA +15 |
I3 industry Logistics +20 | I4 employees 617 +15 |
I5 prior closed-lost +10
A score with no reason string cannot be argued with, so it cannot be improved. When somebody says
Tier A is wrong for an account, the reason string makes it a conversation about a weight, ending in
a rule change. Without it, it ends in "the scoring is broken" and you learn nothing. The reason
string is also how you find the rule that does nothing. That is next.
The rule that fires on everything is not a rule
Score the shipped prospect pool with the exercise rules below and one thing stands out. Of 163
prospect accounts with no open opportunity, 154 have a prior closed-lost opportunity. The "they
have evaluated us before" rule, worth +10, fires on 94% of the population, adds ten points to nearly
everything, and separates nothing. The score looks richer; the ranking is identical. Delete it, or
sharpen it: closed-lost in the last 12 months, or closed-lost with loss_reason = Price when
your packaging has since changed.
So the last step of writing an ICP is not writing it. It is running each rule against the population
and printing the fire rate:
For each rule, report: rule id, records matched, percentage of the population.
Any rule above about 80% or below about 5% deserves a second look. The first is a constant, the
second a special case you will forget you wrote.
The scoring brief, run for real
OUTCOME
Score every account in the export against the ICP rules and output one row per
account with the score, the tier and the reason string.
INPUTS
10-data/accounts.csv, 10-data/opportunities.csv
Rules: 40-reference/icp-rules.md, sections N (normalisation) and I1 to I5,
tier bands in section T.
OUTPUT SHAPE
30-output/demand/icp-scores.csv, columns exactly:
account_id, account_name, score, tier, reason, suppressed, suppression_reason
reason is a single string of the form "I1 <condition> +<n> | I2 ... ".
Every scored row has a non-empty reason.
Then, printed in the chat, not written to file: the rule fire-rate table.
Columns: rule_id, records_matched, pct_of_population.
CONSTRAINTS
Apply the normalisation in section N before any string comparison.
Scoring and suppression are separate steps: score every account, then set
suppressed = TRUE with a reason where a suppression rule applies. Do not remove
suppressed rows from the file.
Do not invent a rule that is not in icp-rules.md. If a record cannot be scored
against a rule, write it to exceptions.csv rather than scoring it as zero.
DEFINITION OF DONE
Every account in the input appears exactly once in the output.
Row count in equals row count out.
Two lines in that brief do real work. "Scoring and suppression are separate steps": filter
suppressed accounts out during scoring and you can no longer answer "how much of our best-fit
universe is already in play", one of the more useful numbers a marketer can produce. "Row count in
equals row count out": the reconciliation. It catches the run that silently drops the two accounts
with a dirty industry string, without you having to notice which two.
Try this now: score four accounts by hand, then check the rules
About seven minutes. Do the arithmetic yourself. It is the last time you will, and it is why you
will trust the output afterwards.
You are settling the ICP argument before territory planning, with accounts.csv and
opportunities.csv. Sales says logistics and hospitality in EMEA. Marketing says anything
mid-market.
Create 40-reference/icp-rules.md from templates/M7/icp-rules.md in your practice pack.
Write these five rules with these weights, so your numbers match the answer below:
I1 segment Mid-Market +30, Enterprise +25, SMB +5
I2 region EMEA +15, NA +5
I3 industry Retail, Hospitality, Logistics or Healthcare +20
Facilities Services +10, Manufacturing +0
(after normalisation: trim, title case)
I4 employees 200 to 1500 +15, above 1500 +10, below 200 +0
I5 history any prior Closed Lost opportunity +10
Tiers A at 70 and above, B 50 to 69, C below 50
Write one suppression rule: an account with an open opportunity is suppressed, with the
reason recorded. It is still scored.
Score these four accounts by hand, with a reason string for each:
ACC-02831, ACC-02885, ACC-03112, ACC-02995.
Run the fire-rate check on rule I5 across all prospect accounts with no open opportunity, and
decide what to do about it.
Starter brief (fill the gaps marked <<...>>):
OUTCOME
Score the four named accounts and print score, tier, reason, suppressed.
INPUTS
10-data/accounts.csv, 10-data/opportunities.csv
Rules: 40-reference/icp-rules.md
OUTPUT SHAPE
A table in the chat. Columns: account_id, account_name, score, tier, reason,
suppressed, suppression_reason.
The reason string format is <<write the exact format you want, with an example>>.
CONSTRAINTS
Normalise industry before comparing by <<which operations, in which order?>>.
An account is suppressed when <<state the condition, and say which file tells you>>.
<<does suppression change the score? say so explicitly>>
DEFINITION OF DONE
Four rows. Every row has a non-empty reason naming every rule that fired.
EXCEPTIONS RULE
<<what happens to an account whose industry value is not in your list at all?>>
If you get stuck
Hint 1: the idea
Two of the four accounts will surprise you, and both times the rules are working correctly. Before
you call a score wrong, decide whether you disagree with the score or with the weights. Only one of
those has a fix.
Hint 2: the move
Suppression is not in accounts.csv. Look in opportunities.csv for a row on that account whose
stage is not Closed Won or Closed Lost. Do the same lookup for I5, with the opposite test.
One of the four scores well and is suppressed. That combination is the most useful row in the
output, not an error.
Hint 3: the shape
A correct answer spreads the four across all three tiers, and exactly one carries
suppressed = TRUE while still holding a score. Two are dragged down by the same rule: that is the
weight doing most of the work.
The fire-rate check on I5 should come back high enough that the rule separates almost nothing. When
it does, change the rule. Do not be pleased about the coverage.
Fire-rate check, rule I5: 154 of 163 prospect accounts with no open opportunity have a prior
Closed Lost opportunity. 94.5%.
Your answer may legitimately differ: different weights give different totals, and the exercise only
asks that your arithmetic matches your own rules. What must match is the structure: four rows, every
reason string naming every rule including the +0 ones, and ACC-02995 scored and suppressed rather
than dropped. Omit the +0 rules and you cannot see which rule does nothing. If ACC-02995 is
missing entirely, your brief conflated scoring with filtering.
If ACC-02885 came out at 90 rather than 70, check whether you gave Manufacturing an industry
weight. That is a genuine choice, and it decides whether Ardleigh Works BV is on next quarter's
list.
Worth knowing: the rule that felt most insightful when you wrote it, "they have evaluated us
before", fires on 94% of the population and changes nothing about the ranking. The rule doing the
real work took ten seconds to write.
What you keep
40-reference/icp-rules.md: normalisation, five rules with weights, tier bands, the reason-string
requirement, the suppression rule, and a fire-rate line per rule with the date it was last checked.
Check yourself
Which of these is usable as an ICP rule, and rewrite the other: (a) "operations-led businesses
with complex scheduling needs"; (b) employee_count between 200 and 1500 AND
region = EMEA.
Your scoring output has a score, a tier and no reason. Sales says Tier A is wrong. What can
you do, and what can you not do?
A rule fires on 3% of your population and contributes 25 points when it does. Problem? Say what
you would check before answering.
Where this goes next
L7.3 writes the account research brief, then breaks it on the way research actually fails: a
number that was true, from a source that said it was unverified, restated as a fact.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
M7 lesson / about 13 min
Failure mode
The research brief, and the fabrication hunt
Do this nowFind the marking on the numbers60s
Open notes/qbr/business-review-lyndhurst-retail.md.
Find the section "What the customer reported" and read the line above the bullets.
Copy that line somewhere you will see it again.
You know it worked when: You are holding a sentence saying the figures belong to the customer, that Talvera has not audited any of them, and that they must not be quoted externally without written permission.
What you type
Produce an account research brief for ACC-03128, Lyndhurst Retail Group Inc,
that a seller can act on before an expansion conversation.
INPUTS
10-data/accounts.csv, opportunities.csv, product_usage.csv, support_tickets.csv,
and 10-data/notes/. Web research: not enabled.
Sections, in this order: what we already know | opportunity history | usage and
support | external context | why now (hypothesis) | unknown.
Every claim is followed by a "Source:" line naming the file and the record id or
the document section. Do not use general knowledge about this company.
What comes back
30-output/demand/account-briefs/ACC-03128.md
## What we already know
Enterprise, Retail, Canada, NA, 6,944 employees. Live since 2 December 2023 at
450,000 USD. Renewal 2026-12-02.
Source: accounts.csv ACC-03128, notes/qbr/business-review-lyndhurst-retail.md.
Lyndhurst has 192 corporate properties live with approximately 84 hourly staff
each. Labour cost is down 1.5% since go-live, and schedule publication has moved
from a Thursday scramble to a Monday routine.
Source: notes/qbr/business-review-lyndhurst-retail.md.
## Why now (hypothesis)
The 140 franchise managed properties are the remaining prize. OPP-007302 is open
at 140,000 USD in Solution Review, close 2026-10-25, and the QBR states it covers
the master agreement alone.
Source: opportunities.csv OPP-007302, notes/qbr/business-review-lyndhurst-retail.md.
Notice: Nothing is fabricated and every figure is transcribed correctly. What was dropped is the sentence saying the numbers belong to the customer and are unaudited, so the brief now reads as something Talvera measured.
Why this matters
You walk into a first expansion conversation at a Tier 1 account. Your brief says the customer
reduced labour cost by one and a half percent since go-live.
You say it in the room. The VP of Operations looks at you and says: that was our estimate, we never
published it, and our finance team disagreed with it. The rest of the meeting is about whether you
are careful.
Nothing in your brief was invented. Every number was in a real document. And the brief is still the
reason the meeting went badly.
Everything this lesson uses came in the practice pack from the course page; if you have not
downloaded it, do that first.
What you will be able to do
Structure an account brief so internal certainty comes before external inference.
Put a source line on every claim, and an "unknown" section that is allowed to be long.
Run a repeatable fabrication hunt, and fix the brief rather than the output.
Internal facts first, on purpose
Fixed sections, in this order:
1. What we already know from our own data (usually half the brief)
2. Opportunity history and outcomes
3. Relationship map, from contacts and activities
4. Product usage and support signals, if we are already in the account
5. External context, every claim with a source
6. Why now: the hypothesis, labelled as a hypothesis
7. Unknown: what could not be established
The ordering is the point. A seller reads top to bottom and stops when time runs out, so the most
certain material goes first. It is usually also the most useful: here, section 4 tells you Wexley
Care Group runs at 92% seat utilisation with the payroll sync module switched off, a better reason
to call than anything on a website.
Section 6 is labelled a hypothesis, in the output, in that word. Section 7 is not a failure state. A
brief with a long unknown section and six sourced facts beats one with twenty confident sentences,
and the second kind is what you get by default.
One source line per claim
Every claim says where it came from:
Claim: Expansion opportunity open at 74,000 EUR for payroll sync and extended
compliance reporting.
Source: opportunities.csv, OPP-007532, stage Proposal, close_date 2026-09-20.
Claim: The payroll sync module is not switched on.
Source: product_usage.csv, ACC-02830, 2026-07, feature_payroll_sync = FALSE.
Claim: Their payroll platform is on an old release with an upgrade planned for
the fourth quarter, and their IT director will not integrate twice.
Source: notes/qbr/business-review-wexley-care.md, "What stands between us and
74,000 EUR". Reported by the customer on the QBR call, not verified by us.
Notice the third one: sourced, and still not a fact about the world, only a fact about what somebody
said on a call. The source line says so. That distinction is the whole of the break below.
The fabrication hunt
Run this on every research output until it stops finding things. About four minutes per brief.
No source line, no claim: delete it, do not investigate it. The fastest step, and it removes
most of the problem. Checking whether an unsourced claim happens to be true rewards the behaviour
you are trying to remove.
Verify five sampled claims at source by hand. Open the file, find the record, read the value.
Search for the tells. Precise numbers with no source. Recent events with no date. Named
people in named roles. Tooling claims ("they use X"). Any sentence that would be embarrassing if
wrong.
Change the brief, not the output, then re-run. Make that class of claim require a source or
be out of scope. Fix only the output and the same class returns next week, in a brief you have
less time to check.
Step 4 is the one people skip, and the only one that compounds.
Try this now: run the research brief
Save this brief as 20-briefs/account-research.brief.md; a seed to start from is at
templates/M7/account-research.brief.md in your practice pack.
OUTCOME
Produce an account research brief for ACC-02830 that a seller can act on before
an expansion conversation.
INPUTS
10-data/accounts.csv, contacts.csv, opportunities.csv, activities.csv,
product_usage.csv, support_tickets.csv
10-data/notes/ (call notes and QBR documents)
PERMITTED SOURCES: the files above only. Web research: not enabled.
OUTPUT SHAPE
30-output/demand/account-briefs/ACC-02830.md
Sections in this order: what we already know | opportunity history |
relationship map | usage and support | external context | why now (hypothesis) |
unknown.
Every claim is followed by a "Source:" line naming the file and the record id or
the document section.
CONSTRAINTS
Do not use general knowledge about this company, its industry or its people.
A claim with no source is not written as prose: it goes in the unknown section
as a question.
Where a source marks a figure as customer-reported, unaudited or unverified,
carry that marking into the output in those words. Do not paraphrase it away.
Anything that reads as an event ("recently", "just announced") must carry a date
or be dropped.
DEFINITION OF DONE
Every paragraph in sections 1 to 5 has at least one source line.
The unknown section is not empty.
The why-now section is labelled "Hypothesis" in the output.
EXCEPTIONS RULE
Any claim you cannot source goes to the unknown section as a question naming who
could answer it. Do not delete it silently and do not assert it.
What you should see, in section 4 and the unknown section, in this shape:
## Usage and support
Seat utilisation 92.4% in 2026-07 (3,008 active of 3,254 licensed), flat trend,
adoption score 86.
Source: product_usage.csv, ACC-02830, 2026-07.
Forecasting module in use. Payroll sync module not in use.
Source: product_usage.csv, ACC-02830, 2026-07, feature_forecasting = TRUE,
feature_payroll_sync = FALSE.
Nine support tickets on record, one currently open, four escalated beyond first
line.
Source: support_tickets.csv, filtered on ACC-02830.
## Unknown
- Whether the four escalations were the same root cause. The ticket categories
are recorded but no linkage field exists. The support agent named on them
could answer this.
- Who signs a 74,000 EUR expansion. Two contacts are recorded with persona
Economic Buyer (CON-008577, CON-008609) and the QBR names one of them as the
budget owner. Which of the two holds this budget is not established.
Source for the ambiguity: contacts.csv and
notes/qbr/business-review-wexley-care.md.
That last unknown is the brief working as intended: two records say Economic Buyer, the QBR names
one person, and the honest output is a question, not a guess. An unguarded brief picks one, and the
seller addresses the wrong person.
When this goes wrong
Break it
Run the same brief against ACC-03128 (Lyndhurst Retail Group Inc) without the constraint line
that carries unverified markings through. Web research is still off: this failure is entirely
internal.
The tell
The brief comes back reading, in confident prose:
Lyndhurst has 192 corporate properties live with approximately 84 hourly staff each. Labour cost
is down 1.5% since go-live, and schedule publication has moved from a Thursday scramble to a
Monday routine.
Now open the source, the QBR note at data/notes/qbr/business-review-lyndhurst-retail.md in your
practice pack, section "What the customer
reported":
Figures below belong to the customer and were presented by Liesbeth. Talvera has not audited any
of them and they must not be quoted externally without written permission.
The underlying call note, data/notes/call-notes/lyndhurst-retail-franchise-workshop.txt, independently
says the same:
Their numbers, which are theirs and not ours, and which we should not repeat outside without
asking.
Two documents, by two different people, both flagged these figures. The brief stripped both markings
and kept the numbers.
Nothing was fabricated. Every figure is real and correctly transcribed. What was removed was the
sentence saying whose figures they are and what may be done with them. The claim now reads as
something Talvera measured, and the next place it appears is a slide.
Next to the source, the laundering is obvious. Alone, as a seller reads it at 8:50am, it is
invisible. That is why this failure survives review by the person who wrote the brief.
A second instance in the same dataset has a different shape.
data/notes/call-notes/call-selby-supply.txt records a champion's estimate that depot supervisors spend
two to three hours a week each rebuilding rotas across 18 depots. The rep who wrote it added: "His
number, not ours. I have never validated it and I should." An unguarded brief produces "depot
supervisors lose two to three hours a week to manual rota building", and now Talvera is asserting
it.
The guard
Three rules in the brief, and a fourth habit:
Every claim carries a source line. Already in the brief.
Anything the source marks as unverified, unaudited, customer-reported or not to be quoted stays
marked in the output, in those words. Not paraphrased, not footnoted, not dropped because it
interrupts the sentence.
Anything with no source goes to the unknown section as a question, never into prose.
Re-run and confirm the sentence survives with its marker attached. A guard you have not
watched fire is not a guard. The corrected output reads:
The customer reported, on the QBR call and unaudited by Talvera, labour cost down 1.5% since
go-live across 192 corporate properties. Source:
notes/qbr/business-review-lyndhurst-retail.md, "What the customer reported", which states the
figures belong to the customer, have not been audited by Talvera, and must not be quoted
externally without written permission.
Then write one sentence, by hand, on what would have happened if the first version had reached a
pricing conversation. Keep it. It is why you will not remove the guard when it makes a brief less
punchy.
Worth knowing: the call notes here are messy in form but agree with the CSV rows, with one
deliberate exception, documented in answer-keys/PLANTED-ISSUES.md in your pack, which you open
only after attempting the lab, under note-crm-contradiction: opportunity
OPP-007372 on account ACC-02826, where a note states an amount, a close date and a stakeholder
role the CRM rows disagree with, and claims the record has been updated when it has not.
Laundering (this lesson) is a marked claim losing its marking; a source-marking rule catches it.
Contradiction is two sources disagreeing about a fact; you catch it by reconciling sources and
refusing to pick, the deal-card corroboration work in Module 6.
What you keep
20-briefs/account-research.brief.md, with the permitted-sources block, the source-line rule, the
unverified-marking rule and the unknown-section rule.
30-output/demand/account-briefs/<account-id>.md, at least one.
A 50-runs/RUN-LOG.md entry recording the fabrication hunt: claims checked, claims removed,
claims moved to unknown, and the brief change you made as a result.
Check yourself
A brief contains: "They are investing heavily in workforce technology this year." What is wrong,
and what are the two acceptable outcomes for that sentence?
Your source is a QBR document that flags a figure as customer-reported. Your seller says the
marking makes the brief read weakly. What do you say, and what do you offer instead?
A fabrication hunt finds three unsourced claims of the same kind. Name the two changes you make,
and which one comes first.
Your unknown section has eleven items and your sourced facts number six. Working or failing? What
would tell you?
Where this goes next
L7.4 builds the list these briefs feed: suppression, a reconciliation that proves nothing was
dropped, and the review gate between a draft and a prospect.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
M7 lesson / about 13 min
Do it
Lists that reconcile, drafts that get reviewed
Do this nowCount the opt-outs45s
Open contacts.csv and count the rows where opted_out is TRUE.
Check that each of those rows also carries an opted_out_date.
Write the count into 40-reference/suppression-rules.md.
You know it worked when: 75 of 911 contacts are opted out, and your suppression file now holds a number rather than an intention.
What you type
OUTCOME
Build the target segment: EMEA Mid-Market and Enterprise prospects in Retail,
Hospitality, Logistics or Healthcare, with contacts at Manager level or above.
INPUTS
10-data/accounts.csv, 10-data/contacts.csv, 10-data/opportunities.csv
Rules: 40-reference/suppression-rules.md
OUTPUT SHAPE
30-output/demand/segment-2026-08-06.csv, plus a reconciliation ladder printed
above it. The ladder starts at the full file count, subtracts one line per
exclusion rule with its own count, and ends at the final count. Accounts first,
then contacts.
CONSTRAINTS
Normalise industry (trim, then title case) before any comparison.
Apply every rule in suppression-rules.md, not only the ones this brief names.
DEFINITION OF DONE
The starting count minus every exclusion count equals the final count exactly.
If it does not, stop and report the step where it broke.
What comes back
30-output/demand/segment-2026-08-06.csv written. 61 accounts, 142 contacts.
ACCOUNTS
308 accounts in file
- 104 existing customers and churned logos
- 41 accounts with an open opportunity
- 64 SMB segment, out of scope for this campaign
- 29 region is not EMEA
- 9 industry not in Retail, Hospitality, Logistics, Healthcare
= 61 accounts
CONTACTS
188 contacts at those 61 accounts
- 12 opted_out = TRUE
- 4 role-based mailbox
- 1 malformed or untrimmed email address
- 29 seniority below Manager
= 142 contacts
Notice: 308 minus the five exclusion counts is exactly 61, which is the only reason you can say nothing was dropped by accident instead of hoping so.
The segment ladder on the shipped data goes 308 accounts to 61 and 188 contacts to 142 with every excluded record counted, so a record lost between steps shows up as arithmetic that does not close.
Why this matters
A campaign goes out. Forty recipients are existing customers, two are in an active negotiation with
your own AE, and one asked not to be contacted fourteen months ago.
Nobody chose any of that. The list came from a correct rule, and the records that should have been
removed were never in anybody's head as a category. With no reconciliation, nothing announced the
skipped step.
What you will be able to do
Write suppression as first-class rules rather than a final tidy-up.
Produce a count reconciliation that proves nothing was silently dropped.
Tell a sourced personalisation line apart from a plausible one.
Write the review gate between a draft and a prospect, and name who owns each part.
Suppression is a rule set, not a filter you remember
40-reference/suppression-rules.md, which you write once in this lesson (a seed is at
templates/M7/suppression-rules.md in your practice pack) and apply to every list you will ever
build:
Rule
Why it exists
Existing customers, unless the campaign is for customers
The campaign that offers a new-logo discount to a renewing account
Accounts with an open opportunity
The AE finds out from their prospect
Do-not-contact and opt-out
Legal, and it is the one people find out about publicly
Recently contacted, within N days
Sequence collision across teams
Competitors and partners
Obvious afterwards
Bad or role-based email addresses
Deliverability, which you own even when a tool sends
Three of those are measurable in this dataset right now: 87 accounts are customers and 17 are
churned; 62 open opportunity rows sit across 62 distinct account references, 61 of which
resolve to a real account (one row carries an account_id that does not exist, which is why the
suppression step counts references rather than trusting the join); and 75 of 911 contacts carry
opted_out = TRUE with an opted_out_date.
And the failure is already in the data. answer-keys/PLANTED-ISSUES.md in your pack, which you
open only after attempting the lab, documents a compliance-optout class of
22 rows in activities.csv: emails recorded as sent after the contact's opt-out date.
ACT-0041324, for instance, records an email sent on 2026-03-08 to CON-009112, who opted out on
2025-05-01. Somebody built a list, the suppression step was not in it, and the send happened.
That is why, in LAB-07, the exercise sheet in labs/LAB-07 of your practice pack (start with its
README), a segment holding an opted-out contact counts as a safety failure, not a
completeness deduction. It is not an incomplete list. It is a list that does harm.
The reconciliation ladder
A segment list states its inclusion rules, its exclusion rules, and a ladder counting every excluded
record. If the numbers do not add up, something was dropped without anybody deciding it. Here is a
real ladder from the shipped data, targeting EMEA mid-market and enterprise prospects in the four
core industries:
ACCOUNTS
308 accounts in file
- 104 existing customers and churned logos
- 41 accounts with an open opportunity
- 64 SMB segment, out of scope for this campaign
- 29 region is not EMEA
- 9 industry not in Retail, Hospitality, Logistics, Healthcare
= 61 accounts
CONTACTS
188 contacts at those 61 accounts
- 12 opted_out = TRUE
- 4 role-based mailbox (sales@, admin@, accounts@, rota@)
- 1 malformed or untrimmed email address
- 29 seniority below Manager
= 142 contacts
Three things this ladder does that a total does not.
It proves nothing vanished. 308 minus the five exclusion counts equals 61, exactly. If not, a
record was lost between two steps, and you go find out where.
It makes the expensive rule visible. The largest exclusion is 104 existing customers, then 64
for segment. Whoever wants a bigger list sees which rule to argue about, and "SMB is out of scope"
is a business decision somebody can reverse in a meeting.
It catches the silent drop. The industry step is where this dataset bites: accounts.industry
holds ' Retail' with a leading space, 'logistics' in lower case and 'Hospitality & Leisure'
(data/DATA-DICTIONARY.md in your practice pack). A rule matching industry = 'Retail' exactly
drops ACC-02948 and never
mentions it. Normalise first, and print how many records changed under normalisation. That turns a
silent drop into a line in the ladder.
A sourced line and a plausible line
Two personalisation lines for the same account, ACC-02830, Wexley Care Group GmbH. Both read well.
One can be checked by a reviewer in forty seconds without opening the CRM.
Line A: sourced
You are running Talvera Shift at 92% seat utilisation across your licensed base with the payroll
sync module switched off, and your team costed the fortnightly manual payroll export at two
full-time-equivalent days per cycle.
Evidence: product_usage.csv, ACC-02830, 2026-07, seat_utilization 0.924,
feature_payroll_sync = FALSE. Second claim reported by the customer on the QBR call and not
verified by Talvera: notes/qbr/business-review-wexley-care.md.
Two claims, two source lines, and the second carries its marking through, because the source marks
it as the customer's own costing. A reviewer reads the evidence block, opens one file, and approves.
Line B: plausible
As a healthcare group scaling across multiple sites, you are likely wrestling with the compliance
reporting overhead that comes with fragmented rostering data.
Nothing in that sentence is false. Nothing in it is checkable either. It was not read out of any
file; it was inferred from the account being in healthcare with more than one site. It will pass a
skim review, land as generic, and if the prospect answers "actually our compliance reporting is
fine, that is the one thing we sorted last year", the seller has nothing to fall back on.
The test is one question: which file and which record would I open to check this? No answer
means a plausible claim, and it does not go in a customer-facing asset as a statement about the
customer.
What a good reviewer says about line B
The line is fluent, on-brand, and not usable, and not because it is wrong. It may well be true.
Nobody can establish whether it is true, so it is not personalisation, it is a segment-level
assumption in the second person.
Which file and record would you open to check "wrestling with compliance reporting overhead"? No
answer. Meanwhile this account's own folder holds, with no external research: seat utilisation at
92.4% in the most recent month, the payroll sync module switched off, an open 74,000 EUR expansion
on payroll sync and compliance reporting, and a QBR document in which the customer's own
operations director costed the manual payroll workaround. Four checkable, specific, unused facts.
An inferred line is also invisible to a review gate: a claim with no source presents nothing to
check, so it passes. The gate fails silently, which makes this class of line the one most likely
to reach a prospect.
Fix the brief, not the draft. Add: "Every personalisation line carries an evidence block naming
the file and the record it came from. A line that cannot carry one is not written." Re-run and
count how many drafts lose their opening sentence. That count is how much of your personalisation
was ever personal.
If you want the industry framing, keep it and label it. "Teams like yours often tell us X" is a
claim about your other customers, not this one, and checkable against your own data if you state
the number.
Draft-only, and what makes a draft reviewable
Sequence drafts and campaign briefs are draft-only by definition, stated in the output header,
not assumed:
BLAST RADIUS: draft-only
Reviewed by: <name> (factual claims), <name> (messaging), <name> (product and
regulatory claims)
Not to be sent until all three have signed.
A draft is reviewable when the evidence travels with it: every personalisation line carries its
evidence block underneath, so a reviewer checks the claim without opening the CRM. Strip the blocks
before sending; keep the version with them in 30-output/demand/.
The review gate, written down
40-reference/review-gate.md, which you write now (a seed is at templates/M7/review-gate.md in
your practice pack). This is your own policy, and the file says so; the product does not
enforce it.
What is being reviewed
Who owns it
What they are checking
Factual claims about the account
The seller who owns the account
Every evidence block resolves, and nothing is stated that the record does not support
Messaging and positioning
Marketing or brand
It sounds like the company, and it does not promise something the product does not do
Product claims and regulated segments
Legal or compliance
Anything making a product claim, and anything touching a regulated segment or region
Suppression
Whoever owns the list
The ladder reconciles, and the opt-out count is not zero when it should not be
Two lines belong underneath the table:
Nothing generated in this module reaches a prospect without a human reading it. Not a sample.
Every asset.
Deliverability, consent and regional contact rules are outside the scope of this course and are
your own obligation. We teach the suppression discipline. We do not tell you what the law is
where you operate.
Keep the rules, not the output
The segment expires in a month. The briefs expire the day after the meeting. What survives:
40-reference/icp-rules.md, with its fire-rate check dated.
40-reference/suppression-rules.md.
40-reference/review-gate.md.
20-briefs/account-research.brief.md.
Each is a document most GTM teams have never written down, each is portable to any tool, and none
depends on this product. Cowork carries no memory between sessions, so these files are not
documentation of your process. They are your process.
Try this now
Write 40-reference/suppression-rules.md from the table above, one row per rule.
Build the segment from your ICP scores, with the reconciliation ladder alongside the list: every
exclusion counted, normalisation changes counted too.
Check the arithmetic. If the ladder does not add up, say: "A record was lost between two steps.
Find which step, and show me the records."
What you keep
40-reference/suppression-rules.md and 40-reference/review-gate.md.
30-output/demand/segment-YYYY-MM-DD.csv with its reconciliation ladder in a header comment or a
companion markdown file.
At least one draft asset with an evidence block under every personalised claim, and its blast
radius stated in the header.
Check yourself
Your segment ladder starts at 1,209 leads and ends at 300, and the exclusions add to 880. What
do you do before you send anything?
Marketing asks you to drop the open-opportunity suppression because it costs 41 accounts and the
list is small. Give the answer, and the alternative that meets them halfway.
A reviewer approves a sequence in four minutes across twelve accounts. What can they not have
checked, and what change to the draft format would have forced it?
Which of these needs legal review before it leaves the folder, and why: (a) a subject line;
(b) "cuts payroll reconciliation time by half"; (c) a personalisation line quoting the prospect's
own published headcount; (d) a follow-up to a contact in a regulated healthcare segment.
Where this goes next
LAB-07 runs the whole pack on the shipped dataset with a detection scorecard that scores your
suppression, then on your own data behind the redaction gate.
M8 turns to the accounts you already have: the same brief pattern becomes a QBR pack, and the
work starts repeating on a cadence you have to design.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
Lab / ? min / draft-only
Lab
Targeting without fabrication
LAB-07: Targeting without fabrication
The situation
There are 1,209 leads in the shipped Talvera dataset and no scoring anybody can explain. An SDR
spends most of a working block on account research that mostly recovers facts already sitting in your own
data. A campaign brief takes long enough that the first version is thrown away, because the segment
definition was ambiguous.
And there is a danger in this module that does not exist in the others. Hygiene work and pipeline work
are wrong in ways you can catch by reconciling against a source. Research output is wrong in a way that
reads beautifully: a number nobody measured, an event stated without a date, a job title from three
roles ago. That claim then goes into an email to a real prospect with your name on it.
This lab spends a disproportionate amount of its time on the fabrication problem, because that is where
this work actually fails.
What you will produce
File
What it is
40-reference/icp-rules.md
Scored, weighted, testable ICP rules with tier bands and the reason-string requirement
40-reference/suppression-rules.md
Exclusion rules for any list, with the reconciliation they must produce
40-reference/review-gate.md
Who reviews what before an asset leaves the folder
20-briefs/account-research.brief.md
The research brief, with source discipline built in
30-output/demand/segment-2026-08-02.csv
The segment, with a reason string per record and a count reconciliation
30-output/demand/account-briefs/<account-id>.md
Five account briefs, each with source lines and an "unknown" section
30-output/demand/fabrication-hunt.md
Claims checked, claims removed, and the brief change that stops each class recurring
Outcome accounting
Manual baseline: 26 minutes per account brief.
Measured. The lab author built one brief by hand for a single account: pull the
account row, its opportunities and outcomes, its contacts and their activity,
its usage and ticket history, read the call notes and the QBR document, then
write the page. Timed once, on one account. Not extrapolated.
Target with Cowork: Five briefs in 22 minutes of running, observed. That is about 4.4 minutes each.
Plus 9 minutes of fabrication hunt across the five, which is 1.8 each.
Effective: about 6.2 minutes per brief, checking included.
Verification cost: 9 minutes across five briefs, counted separately and inside the number above.
Source line present on every claim, mechanically 2 min
Five claims sampled and verified at source by hand 5 min
Fabrication tell search across all five 2 min
MODULES/07 requires the fabrication hunt time to be inside the net saving, and
it is. A research workflow whose saving is quoted without the hunt is quoting a
number for a process nobody should run.
Net: 26 - 6.2 = about 20 minutes saved per brief.
Five briefs: 130 minutes by hand against 31. Saves about 99 minutes.
Your actual: _____ your own by-hand time for ONE brief, measured before you start
_____ five briefs _____ fabrication hunt
Measure one brief by hand before you start. We are not printing a figure for your data that we have not
measured on your data.
Blast radius
Draft-only, and this is the module where the label does the most work. Sequence drafts and campaign
briefs are draft-only by definition and the class is stated in the output header.
Obligation
This lab
Review gate
The seller reads every factual claim about the account; the fabrication hunt runs before review, not after
Approver
The seller for factual claims, marketing or brand for messaging, legal or compliance for product claims and regulated segments
Reversal plan
Do not send
Nothing generated in this module goes to a prospect without a human reading it.
One rule in this lab is a hard fail rather than a deduction. A segment that includes an opted-out
contact fails the Safety row outright (MODULES/07). contacts.csv carries opted_out and
opted_out_date, so the suppression rule is writable and there is no excuse for missing it.
activities.csv is in the file scope because all 22 compliance-optout rows are activity rows:
emails sent to a contact after the date they opted out. Without the file the class would be declared
with zero rows, and it is the class that carries the hard-fail rule above.
period-attribution belongs in a demand lab because bookings shifted across a period boundary distort
which campaign and which segment appear to convert. It is a demand-side error before it is a finance
one, and a source-performance report built on the raw close dates will credit the wrong quarter.
Hard-class coverage requires an assertive identification and accumulates across the four scored core
labs toward the 3-of-6 bar (SCORING.md §2, §7).
How you verify your own output
Source line on every claim, checked mechanically. Any sentence in an account brief without a
source line is deleted, not investigated. That is the rule, and it is cheap because it does not
require judgment.
Sample five claims and verify them at source by hand. Five across the whole set, not five per
brief. State that you did.
Search for the fabrication tells: precise-sounding numbers with no source, recent events stated
without a date, named people in named roles, tooling claims, and any sentence that would be
embarrassing if wrong.
The segment reconciliation. Total records, excluded by each rule with counts, remaining. If the
numbers do not add up, something was dropped silently, and in a segment build a silent drop means
somebody did not get contacted or somebody did.
When you find a fabrication, change the brief so that class of claim requires a source or is out of
scope. Then re-run. Do not fix the output.
What ships and what does not, on browser research
Claude in Chrome is available conditionally: a Chrome extension in beta on all paid plans that lets
Claude read pages, click, type, fill forms and move between tabs, and tasks can be launched from the
Chrome side panel or from Cowork or Claude Code (FACTS §5.9). Officially supported on Chrome, and it may
work on other Chromium browsers (FACTS §5.9). It draws on the same account usage allowance
(FACTS §5.9, §6).
It is optional and this lab is completable without it. Part C in part-a.md is the extension, and it
is explicitly optional with the usage cost noted. No enrichment vendor, data provider or specific
connector capability is claimed anywhere in this lab, because our fact base does not cover them.
A requirement this lab states rather than claims
MODULES/07 beat 5 requires two teaching defects for the fabrication hunt: a note that contradicts the
CRM, and an outdated job title in a note.
The note contradiction ships. Key 20260801 documents it as note-crm-contradiction, and it is in
LAB-06's declared scope. This lab uses it as a second worked example rather than as scored scope.
The outdated job title does not exist in key 20260801. Part 3 does not own
course/labs/data/, so this remains a requirement on the dataset owner, and this lab does not assert
that such a defect ships. The fabrication hunt runs on the laundering case, which does ship and is
the sharper lesson anyway.
LAB-08 reuses the account brief pattern against existing customers for expansion, and the review gate you
write here is one of the three governance files every track produces.
Module 7 · Demand and targeting, research, ICP, lists and outbound assets
6 questions / pass at 70%
Checkpoint
Module 7 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 6
Three sentences from a generated account brief.
(1) Talvera has two open opportunities with this account totalling 118,000 EUR, both created in March
2026 [source: opportunities.csv, OPP rows listed in appendix].
(2) The account raised eleven support tickets in the last two quarters, four of them escalated
[source: support_tickets.csv, filtered by account_id].
(3) They recently completed a Series B and are expanding their operations team across three new
markets.
Which sentence is the fabrication, and what is the tell?
Hint
Do not ask which sentence is least plausible. Ask which sentence you could not check in the next two
minutes, and what about its wording tells you that before you try.
A. Incorrect. Sentence 2 carries a source line naming the file and the filter, which makes it checkable in under a minute. A join can indeed be wrong, and the response to that is verification (pull the rows and count them), not suspicion. A claim you can check is not the risk in this module.
B. Correct. You identified the fabrication by structure rather than by plausibility, which is the only method that scales. The tells are all present at once: a recent event with no date, a named business fact with no source, and a level of specificity ("three new markets") that implies a source the brief does not have. The second half of your answer is the part that stops it recurring: fixing the output leaves the same class of claim free to appear next time, and fixing the brief removes the class.
C. Incorrect. This is a real precision issue and it is not fabrication. The currency question has a determinate answer in the data, so the sentence is checkable, and the fix is a brief that requires the currency to be named. Grading it as the fabrication would train you to look for imprecision rather than for unsupported assertion, and those fail in different ways.
D. Incorrect. It may well be true, and that is exactly why the rule is what it is. You have no way to know, the seller reading the brief has no way to know, and the claim is heading for an email with their name on it. The rule is that claims without a source are deleted rather than investigated, because investigating each one individually costs more than the brief saves, and "probably true" is not a standard you can defend to a prospect.
Question 2 of 6
Two ICP definitions.
Definition 1. Mid-market SaaS companies with a growing sales team, a modern tech stack, and a real
appetite for change.
Definition 2.
Weight Condition
+3 employee_count between 200 and 1500
+2 industry in (Retail, Hospitality, Logistics)
+2 region = EMEA
+2 at least one closed-won opportunity in the last 24 months
+1 seat_utilization above 0.6 in the most recent usage month
-3 lifecycle_status = Churned
Tier A: 8 or more. Tier B: 5 to 7. Tier C: below 5.
Output carries a reason string listing every rule that fired.
Which is usable, and what specifically makes the other one untestable?
Hint
Take one condition from each definition and ask what you would type to apply it to an export. One of them
converts to a filter. The other one converts to an argument.
A. Incorrect. This is the most comfortable answer and it is how the argument never gets settled. Definition 1 is not the strategy behind definition 2, because there is no way to check whether definition 2 implements it: no set of field conditions can be shown to be a faithful rendering of "appetite for change". Keeping both means keeping a claim that can be used to overrule the rules whenever somebody dislikes the output.
B. Incorrect. Right answer, wrong principle, and the principle is what you carry forward. Numbers are not what makes definition 2 usable. A rule reading "score above 80 on our fit model" is entirely quantitative and entirely untestable, because it names no field. What matters is that a second person can apply the definition to the same data and get the same answer.
C. Incorrect. The critique of definition 2 has real force and it is an argument for changing the conditions, not for keeping definition 1. If the fields in your CRM do not predict a good customer, the work is to find fields that do, or to write down the missing signal and go and collect it. A definition nobody can apply predicts nothing at all.
D. Correct. The test is reproducibility, not tone. Definition 2 names fields, thresholds and weights, so two people applying it to the same export get the same tiers, and the reason string makes every score arguable and therefore improvable. Definition 1 fails on all four conditions, and you picked out the worst of them: "appetite for change" has no record behind it, so anybody scoring it substitutes their own proxy, and the resulting list cannot be defended in the meeting where marketing and sales disagree.
Question 3 of 6
A segment build for a new-logo campaign.
Segment: EMEA operations leaders, new logo, Q3
Total contacts in source 911
excluded: account is an existing customer -212
excluded: open opportunity on the account -46
excluded: region is not EMEA -381
excluded: seniority not in target set -94
-------
Remaining 178
The numbers do not reconcile, and one exclusion rule is missing. Name both problems.
Hint
Take one imaginary contact who works at an existing customer, outside EMEA, in the wrong seniority band.
Count how many times the report just removed them. Then look for the field in `contacts.csv` that nobody
filtered on.
A. Incorrect. Exclusion order changes the per-step counts and does not change the final set or fix the reconciliation. Ordering by volume is a readability preference. Whatever order you choose, the report still has to say whether each count is measured before or after the previous exclusions, which is the actual defect.
B. Correct. You caught both halves. The arithmetic looks right precisely because it was computed as a subtraction rather than counted, and overlapping exclusions mean a sequential subtraction double-counts every contact that matched two rules. The honest form reports each rule's count as "removed at this step" after the previous steps, plus a distinct total removed, plus the remaining count verified by counting rather than by arithmetic. The missing rule is the one with a consequence outside your own reporting: an opted-out contact in a campaign is a compliance failure, and the field is right there in the source.
C. Incorrect. Recently-contacted suppression is a real rule and it is the second most important of the ones missing here. Opt-out is first, because it is a stated wish from the contact rather than a courtesy, the data carries it explicitly, and a segment that includes an opted-out contact is a hard Safety failure rather than a deduction. You also missed the reconciliation problem entirely.
D. Incorrect. "The numbers subtract to the stated remainder" is the trap, and it is the same trap as a consistency check standing in for a verification. Arithmetic that is internally correct tells you nothing about whether the remaining set is the right set. Count the remainder independently and compare it to the arithmetic: if they differ, exclusions overlapped, which is the normal case.
Question 4 of 6
A sequence draft comes back with this opening line:
"Given how quickly you are scaling your store estate, keeping shift coverage predictable across sites
must be getting harder."
You check. Nothing in the account record, the activity history or the notes says the store estate is
growing. The agent inferred it from the account being in the Retail industry with an employee count that
sits at the top of the mid-market band.
Which review gate should catch this, and who owns it?
Hint
Ask what kind of claim the sentence is: a claim about the customer, a claim about your product, or a
claim about your positioning. The kind of claim tells you who has to sign it off.
A. Correct. The gate follows the kind of claim. This is a claim of fact about the account, so it belongs to the person who owns the account and would be in the room when the prospect corrects it. Your second half is the durable fix: personalisation carries its evidence, which is what makes the review cheap. A reviewer who has to open the CRM to check one line will stop checking, and the gate quietly stops existing.
B. Incorrect. Marketing owns messaging and positioning, and this line is neither. Routing a factual claim about a named account to a brand reviewer sends it to somebody with no way of knowing whether it is true. Getting the routing wrong is worse than having no gate, because it produces a signature that looks like a check.
C. Incorrect. Legal and compliance review anything that makes a product claim or touches a regulated segment, and an inaccurate flattering guess about a prospect's growth is not that. Escalating everything to legal is how a review gate becomes a queue, and a queue people bypass.
D. Incorrect. The hedge is what makes it worse rather than better. "Must be getting harder" reads as informed, so the prospect assumes you know something about them, and the first reply corrects you. The specific anti-pattern is personalisation that references something the agent inferred rather than something it read, and softening the verb does not change what it references.
Question 5 of 6
An account research brief, in outline.
Account brief: Grantham Foods
1. What we already know from our own data (14 lines, every line sourced)
2. Opportunity history and outcomes (6 lines, sourced)
3. Relationship map (9 contacts, roles, last touch, sourced)
4. Product usage and support signals (7 lines, sourced)
5. External context (4 lines, 3 with source URLs, 1 without)
6. Why now (3 lines of hypothesis, labelled as hypothesis)
There is no "unknown" section. What risk does that create?
Hint
Read the outline as a seller with ten minutes before the call, and write down the three questions you
would still need answered. Then ask where in this brief you would have learned that those questions were
open.
A. Incorrect. Section 6 is already labelled as hypothesis, which is the control for that risk and it is present. Marking a hypothesis and listing unknowns are two different jobs: the first says "this claim is uncertain", the second says "this question has no claim at all". The brief does the first and not the second.
B. Incorrect. Length is not the issue and an unknown section adds to it rather than reducing it. The brief's structure is sound, including putting internal facts first so the seller reads what is certain before what is inferred.
C. Correct. A brief without an unknown section makes silence ambiguous, and ambiguity resolves in the reader's favour: a question with no answer in the brief reads as a question with no problem behind it. The unknown section is what converts "we did not find out who signs" into a fact the seller carries into the meeting as their first question. You also connected it to the unsourced line in section 5, which is where the missing section would have relocated that claim.
D. Incorrect. Source lines control one failure mode, which is claims that are not true. The unknown section controls the other, which is questions that were never asked and therefore never appear. A brief can be perfectly sourced and still leave a seller unaware that nobody established who holds budget.
Question 6 of 6
A colleague proposes: "The sequence drafts are good now. Let us have the workflow send them directly and
review a sample of 20 afterwards each week. It is the same quality control and it saves the bottleneck."
What is the blast-radius class, what does it oblige, and what is wrong with the sample argument?
Hint
Ask when the human sees the message relative to when the prospect does. Then ask what the reversal plan
would be for the messages sent between the two.
A. Incorrect. An email to a prospect is the most consequential output in this entire course. It reaches a person outside your company, with your name on it, and it cannot be unsent. "Not a system of record" is true and irrelevant: the classes are about what breaks, and reaching a customer is named in the definition of the top class.
B. Incorrect. They are drafts right up until the moment the proposal removes the human, which is the whole content of the proposal. Draft-only means a human reads it before it leaves the folder, and "send, then sample" is a description of not doing that. This is the distractor the module names explicitly, because it is the version of the argument that gets made in real teams by people who are not being careless.
C. Correct. You classified on the destination, named both obligations, and identified the specific flaw in the proposal. A post-hoc sample is a monitoring instrument, not a gate: it is what you use to measure a process you have already decided to trust. The line that ends the argument is that there is no reversal plan for a sent email, and writes-to-system requires one that is written and tested before the first write.
D. Incorrect. You classified correctly and then answered a different question. Raising the sample size improves the quality of your monitoring and does not turn monitoring into a gate: the first 500 emails still went out unread. The missing obligation is not a bigger sample, it is a human read before sending, plus the approver and the reversal plan that the class carries.
Last verified: 2026-08-06.
Product claims trace to course/reference/COWORK-FACTS.md (FACTS §x.y). Dataset claims trace to
data/DATA-DICTIONARY.md and answer-keys/PLANTED-ISSUES.md in your practice pack for the Talvera
dataset generated with seed 20260801.
The module at a glance
Two problems, and they are the same problem.
You have a renewal book with a contract anniversary on every account, and you currently find out an
account is at risk when the customer tells you. By then the decision is mostly made. The signals were
in the data: accounts heading for a difficult renewal show six consecutive months of falling seat
utilisation before the renewal date (data/DATA-DICTIONARY.md and data/product_usage.csv in your
practice pack). Six months of
warning, unread.
Meanwhile the weekly executive report is a Sunday night job, every week, forever, and it is the same
report.
Both are work that repeats on a cadence over data you already have. This module builds the retention
pack, and then answers the question the whole course has been deferring: how does this happen again
next week without you, and what do you do about the fact that the scheduling is not guaranteed.
Lesson
Title
TRACK
Minutes
Module-level piece
L8.1
Health signals with evidence, and the threshold that lied
T, A
13
Failure lane 1
L8.2
The QBR pack, and the half the agent does not write
A
12
Worked exercise
L8.3
Scheduling, the failure plan, and the trigger you do not have
K
14
Discrimination pair, Failure lane 2
L8.4
The remote surface, the rollout brief, and knowing when to leave
K
12
Total lesson time: 51 minutes. Module budget is 50 (PEDAGOGY.md §3.1), tolerance plus or minus
15%.
Then: checkpoint CP-08, taken as the module quiz in the
interactive app (seven items, the longest in the course,
because this module carries the most limits), and LAB-08, the exercise sheet in labs/LAB-08 of
your practice pack (start with its README; Part A 50 minutes, Part B Full 40 or Part B Lite 15).
Outcomes
O8.1 Produce a renewal risk list from usage, ticket, activity and note signals, with the rule
and the evidence line behind every flag.
O8.2 Produce a QBR pack for one account from source data, including a stated "what we do not
know" section.
O8.3 Design a recurring report, schedule it, and write the failure plan for the run that does
not happen.
O8.4 Determine and record which execution mode your scheduled tasks use, and state how you
verified it.
O8.5 Design a cadence-plus-check pattern that substitutes for the event triggers Cowork does
not have, and state its detection latency.
O8.6 Produce a team rollout brief covering permissions, approvals, blast-radius policy, the
audit gap and the usage pool, at a level a manager can approve or reject.
O8.7 Decide when a workflow should move off Cowork to a workflow engine or a CRM automation,
and name the deciding factor.
The artifact you leave with
The retention and operations pack, in your own revenue-workspace/:
File
What it is
Built in
40-reference/health-rules.md
Signal families, thresholds, evidence lines, and the field profile that runs before any threshold
L8.1
20-briefs/renewal-risk.brief.md
The recurring risk pass
L8.1
20-briefs/qbr-pack.brief.md
The retrieval half of QBR preparation, with the human half fenced off
L8.2
30-output/retention/renewal-risk-YYYY-MM-DD.md
The dated risk list
L8.1
30-output/retention/qbr/<account-id>.md
One QBR pack
L8.2
ops/scheduled-tasks.md
The register. One row per scheduled task: cadence, output location, execution mode as verified and the date, detection latency accepted, failure plan, owner.
L8.3
ops/rollout-brief.md
The manager-facing approval document, one to two pages
L8.4
Templates are in templates/M8/ of your practice pack.
The register is the signature artifact of this module and arguably of the course. It is the only
place in this market where a learner is asked to write down what happens when the automation does
not run. The rollout brief is the artifact that gets a team licence approved.
Blast radius
Read-only and draft-only throughout. The risk list and the QBR pack read exports and write markdown.
The scheduled task is the one thing here that keeps operating when you are not watching, which is
why it carries obligations the other modules do not: a verified execution mode, a stated detection
latency, a positive run signal, and a written failure plan. A register entry with no failure plan
fails Safety at 1 in the LAB-08 rubric.
Prerequisites
Modules 1 to 4 for the method, Module 5 for account data you can reason about, and Module 6 for the
evidence-line pattern that the health rules reuse.
What this unlocks
The capstone, on every track, because the capstone requires a register entry and a failure plan
whatever workflow you build. This is the last core module.
Where to go next
When you finish this module you have the full loop, four times over. Two honest next steps:
Run one workflow twice, for real, on your own data, with the run log filled in both times.
Everything in this course is theory until a second run happens on a week you did not feel like it.
Take beat 9 of L8.4 seriously. Some of what you have built should eventually leave this tool.
Knowing which, and saying so, is the difference between an operator and an enthusiast.
Module 8 · Retention, expansion, and work that repeats
M8 lesson / about 13 min
Failure mode
Health signals with evidence, and the threshold that lied
Do this nowSort one CSM's health scores60s
Filter accounts.csv to csm equal to Tomasz Jonker.
Sort health_score ascending and read the first twelve values.
Do the same for any other CSM and compare the two lists.
You know it worked when: His book reads 4, 5, 6, 6, 7, 7, 8, 8, 9, 12, 19, 23, and no other CSM has more than two accounts at 10 or below.
What you type
Flag any account with health_score below 40 as critical.
What comes back
41 accounts flagged critical, out of 102 accounts carrying a health score.
By CSM:
Tomasz Jonker 15 of 23 accounts
Bo Fournier 10 of 21
Helena Zabala 8 of 23
Tobias Costa 4 of 19
Niamh Ricci 4 of 16
Lowest scores in the flagged list:
ACC-02938 Juniper Grocers Corp 4 Customer Tomasz Jonker
ACC-02834 Kingsmere Clinics Ltd 5 Customer Tomasz Jonker
ACC-02945 Marchmont Health Services PLC 6 Customer Tomasz Jonker
ACC-02984 Parkgate Facility Services GmbH 6 Customer Tomasz Jonker
ACC-03086 Follifoot Hospitality Group AB 7 Customer Tomasz Jonker
Notice: Nine of the forty-one are recorded on a 1 to 10 scale, all in one CSM's book, so a fifth of the at-risk list is wrong and the rule read a 9 as catastrophic.
Grouping health_score by the CSM who enters it shows nine of Tomasz Jonker's accounts crowded below 10 on a 1 to 10 scale, which is why 41 of 102 accounts fall under a threshold of 40.
Why this matters
Today you learn a renewal is in trouble when the customer tells you, about six weeks out, after the decision was made in a room you were not in.
The signals were there. In this dataset, accounts heading for a difficult renewal show six consecutive months of falling seat utilisation before the renewal date (data/DATA-DICTIONARY.md in your practice pack). Nobody read them, because nobody owned a view that put usage decline and renewal date on the same screen.
Everything this lesson uses came in the practice pack from the course page; if you have not downloaded it, do that first.
What you will be able to do
Score the renewal motion and name the part that is not agent work.
Write health rules across five signal families, each with an evidence line.
Profile a field before you set a threshold on it, and say why that step is not optional.
Three parts of the renewal motion
Assemble the account's history and signals
Flag risk on stated rules
Decide what to do about a risk
Volume
5. Every account, every month
5
2
Checkability
5. Every figure traces to a row
4. Each flag names the row that fired it
1. The source of truth is a relationship
Judgment density
1
2
5
Cost of being wrong
2
3. A wrong flag wastes a call
5
Specification cost
2
3
5
Verdict
Very high fit. Hand it over.
High fit, with evidence required.
Human.
You are building the first two thirds; the third is not a tooling problem. The QBR pack splits the same way, and L8.2 draws that line.
Five signal families, and one trap
Write 40-reference/health-rules.md before any run; a seed is at templates/M8/health-rules.md in your practice pack. Five families:
The trap: a composite score with no components is worse than no score at all. A CS lead cannot act on 36. They can act on "six months of declining seat utilisation, two escalated tickets, no executive contact since May". The composite sorts, and it removes the one thing that tells somebody what to do on Monday.
Already have a health score? Treat it as one input among five, never as the rule. The story below shows why, with numbers.
Two cheap checks that find money
Renewals missed, not lost. Compare accounts.renewal_date against opportunities where type = Renewal. In the shipped data, eight live customer accounts have a renewal date already past with no renewal opportunity ever created: ACC-02956, ACC-02958, ACC-03043, ACC-03097, ACC-03099, ACC-03107, ACC-03123, ACC-03174 (answer-keys/PLANTED-ISSUES.md in your pack, which you open only after attempting the lab; class missing-renewal-opportunity; narrative N8). Their combined arr_usd is 413,600 USD.
These are not losses. They are absences: nobody built the process, so nothing failed, so nothing raised an alarm. The most common finding in this class of work, and it takes one join.
The health score would not have found them. ACC-03097 scores 90, ACC-02956 scores 74: two of the healthiest accounts in the book, renewal date months in the past, no opportunity on record.
Usage decline against a near renewal. Join product_usage.csv to accounts.csv and look for a run of falling seat_utilization where renewal_date is inside the next five months (PLANTED-ISSUES.md, N4, names the cohort). ACC-02858 (Vireo Retail PLC) is the clearest case: seat utilisation falls every month from 0.633 in February to 0.412 in July, on a renewal dated 2026-09-17, with a renewal opportunity in Negotiation at 96,000 GBP.
Rank by exposure in USD. A CS lead's week is finite: separate "declining but large" from "declining but tiny".
Try this now
Run the renewal risk brief. Step 0 sits at the top on purpose; the story after it is why.
Save the brief as 20-briefs/renewal-risk.brief.md; a seed to start from is at templates/M8/renewal-risk.brief.md in your practice pack.
OUTCOME
Produce the renewal risk list for accounts with a renewal date in the next 150
days or already in the past.
INPUTS
10-data/accounts.csv, product_usage.csv, support_tickets.csv, activities.csv,
opportunities.csv, notes/
Rules: 40-reference/health-rules.md, families H1 to H5.
STEP 0, RUNS BEFORE ANY RULE
Profile every numeric field used in a threshold rule: minimum, maximum, count,
and the distribution grouped by the owning CSM. Report it. If any one owner's
range does not look like everyone else's, stop and report rather than continuing.
OUTPUT SHAPE
30-output/retention/renewal-risk-2026-08-06.md
One row per flagged account: account_id, name, ARR (USD), renewal_date, CSM,
flags with evidence lines, total signal families firing.
Sorted by ARR descending. Then a section "Accounts excluded and why".
CONSTRAINTS
Every flag prints its evidence line: rule | record | value.
Do not compute or output a composite score.
Notes are evidence, never a source of a number.
DEFINITION OF DONE
The field profile from step 0 is printed above the list. Every flag has evidence.
EXCEPTIONS RULE
An account that cannot be evaluated against a rule (no usage rows, no health
score) goes to exceptions.csv with the reason. Missing is not the same as
healthy.
Keep the exceptions rule's last line: no usage rows means unknown, not low-risk. Treat absence as safety and you lose the accounts that stopped logging in entirely.
The threshold that lied
The break. Build the list the way everybody does the first time: accounts.csv has a health_score column, so use it.
Flag any account with health_score below 40 as critical.
It returns 41 accounts. Looks sensible for a book of 102 scored accounts.
The tell. Nine of the 41 are not at risk at all, and all nine belong to the same CSM.
PLANTED-ISSUES.md documents this under SUBTLE scale-mismatch: nine accounts (ACC-02805, ACC-02834, ACC-02938, ACC-02945, ACC-02984, ACC-03007, ACC-03043, ACC-03086, ACC-03118) carry a health score on a 1 to 10 scale; every other account uses 0 to 100. All nine are in Tomasz Jonker's book. A threshold rule reads a 9 as catastrophic.
No single row shows it: ACC-03007 alone reads "health score 9, live customer" and nothing announces itself. The bug only appears when you group the field by who fills it in:
CSM
accounts scored
score 10 or below
11 to 20
21 to 39
40 and above
Bo Fournier
21
2
2
6
11
Helena Zabala
23
2
0
6
15
Niamh Ricci
16
0
0
4
12
Tobias Costa
19
1
1
2
15
Tomasz Jonker
23
9
2
4
8
Nine of one CSM's twenty-three accounts sit at 10 or below; nobody else has more than two. Sorted, his book reads 4, 5, 6, 6, 7, 7, 8, 8, 9, 12, 19, 23, ...: nine values crowded at the bottom of a 1 to 10 range, then a jump to 12 where the real 0 to 100 scores start.
Two consequences, and the second is worse:
22% of your at-risk list is wrong. A CS lead who calls three of those accounts and finds them fine stops reading the list.
ACC-03043 is on the list for the wrong reason, and it is genuinely in trouble. Ziegfeld Freight Inc, 116,000 USD ARR, health score 8 on the 1 to 10 scale, meaning healthy. Seat utilisation 0.743 and flat. But the renewal date was 2026-07-28, five days before the run date, with no renewal opportunity on record, and the last logged activity of any kind is 2025-07-11, over a year ago. The CSM wrote it down in the call note at data/notes/call-notes/csm-ziegfeld-freight.txt in your practice pack: "Renewal date is 2026-07-28. I raised it with Ursula in April, raised it again today, and there is still no renewal opportunity anywhere in the system. Not in the wrong stage, not with the wrong date, just absent."
A single-signal rule got the right account onto the list for a false reason and said nothing about the real one. Fixing the scale bug alone would have removed it entirely.
The guard. Profile the field before you threshold it. In the brief, permanently, as step 0:
STEP 0
For every numeric field used in a threshold rule, report minimum, maximum, count
and distribution, grouped by the owning user (CSM, AE, whoever populates it).
If one group's range does not look like the others', stop and report.
Worth knowing: the lesson is worth more than the bug. A threshold is only as good as the assumption that everybody fills the field the same way, and in a revenue system that assumption is false somewhere; you only find out where by looking. Five minutes per field, once, and it lives in the brief.
One more guard, from the ACC-03043 half: never flag or clear an account on one signal family. Print the number of families firing next to each account, and treat "one family, high confidence" as a prompt to check the other four, not a conclusion.
What you keep
40-reference/health-rules.md: five families, your thresholds, the evidence line format, the step 0 profile requirement, and a line recording the profile result and its date.
20-briefs/renewal-risk.brief.md.
30-output/retention/renewal-risk-YYYY-MM-DD.md, your first dated risk list.
Check yourself
Your health rule flags 41 accounts. What do you check first, before reading a single account name?
An account has no rows at all in the usage file. Your rule does not flag it. Is that correct? Say what your brief should do.
A CS lead says the risk list is "obviously wrong, I called three of these and they are fine". Name the two possible causes and the check that separates them.
Your VP asks for a single number per account so the book can be sorted. Give the answer, and give them something better.
Where this goes next
L8.2 assembles the QBR pack and draws the line between the half an agent can retrieve and the half only you can write.
Module 8 · Retention, expansion, and work that repeats
M8 lesson / about 12 min
Do it
The QBR pack, and the half the agent does not write
Do this nowRead seven months of seats60s
Filter product_usage.csv to account_id ACC-02858.
Read seat_utilization for the months 2026-01 through 2026-07 in order.
Write the seven numbers in a row and check the renewal_date on that account.
You know it worked when: 0.692, 0.633, 0.579, 0.548, 0.520, 0.471, 0.412: six consecutive falls, on an account whose renewal is 2026-09-17.
What you type
OUTCOME
Produce the retrieval half of the QBR pack for ACC-02858.
INPUTS
10-data/accounts.csv, product_usage.csv, support_tickets.csv, opportunities.csv,
activities.csv, notes/qbr/, notes/call-notes/
Rules: 40-reference/health-rules.md for the risk section.
OUTPUT SHAPE
30-output/retention/qbr/ACC-02858.md
Sections 1 to 6 as listed in 20-briefs/qbr-pack.brief.md. Each header carries
"[assembled]". Every figure carries a source line: file plus record id, or
document plus section heading.
CONSTRAINTS
Do not write a narrative, a summary, a recommendation, an ask or a next-steps
section. Those are written by the account owner. If the source material suggests
one, put it in section 6 as a question instead.
Usage trend is stated as a series of months with values, not as an adjective.
Where a source marks a figure as customer-reported or unverified, keep that
marking in the output in the source's own words.
DEFINITION OF DONE
Six sections. Every figure sourced. Section 6 is not empty.
What comes back
30-output/retention/qbr/ACC-02858.md written.
## 2. Usage and adoption [assembled]
Seat utilisation, monthly:
2026-01 0.692 2026-02 0.633 2026-03 0.579 2026-04 0.548
2026-05 0.520 2026-06 0.471 2026-07 0.412
Active seats 153 of 221 licensed in January, 91 of 221 in July.
Source: product_usage.csv, ACC-02858, months 2026-01 to 2026-07.
## 5. Open risks [assembled]
| seat utilisation down six consecutive months | product_usage.csv ACC-02858 | open |
| sponsor departed in May, successor started 8 June, no contact | notes/qbr/business-review-vireo-retail.md | open |
| two of four regions stopped publishing in March | notes/qbr/business-review-vireo-retail.md | open |
| 12 tickets on record, 2 escalated, 2 still open | support_tickets.csv ACC-02858 | open |
No narrative, summary, recommendation or ask written. Section 6 is yours.
Notice: Two of the four risks exist only in prose, which is the honest answer to why health_score 36 sat on this account for months and nobody moved.
Why this matters
QBR prep runs into hours per account, and the slowest parts are pure retrieval: what happened since last time, what the usage did, which tickets hurt, where the commercial position stands. The parts that matter, the narrative and the ask, take twenty minutes, and they are the reason the meeting exists. That is exactly backwards.
We will not claim a figure for how many hours yours takes; we have not measured it. LAB-08, the exercise sheet in labs/LAB-08 of your practice pack (start with its README), measures it.
What you will be able to do
Specify a QBR pack whose retrieval sections are agent work and whose narrative sections are fenced off in the brief itself.
Produce a "what we do not know" section that makes a customer-facing meeting safer.
Mark, in the document, which sections were assembled and which were written.
The sections, and the line through the middle
AGENT ASSEMBLES HUMAN WRITES
1. What happened since last QBR The narrative
2. Usage and adoption, with trend The ask
3. Support summary, two worst named The commercial proposal
4. Commercial position The recommendation
5. Open risks with evidence
6. What we do not know
State the split in the brief, not in your head; the drift is gradual and one-directional. Ask for "a QBR pack" and an agent writes a narrative, because that is what such documents usually contain. Ask for "sections 1 to 6 only" and it stops at the retrieval.
Say it in these words:
Do not write a narrative, a summary, a recommendation, an ask or a next-steps
section. Those are written by the account owner. If the source material suggests
one, put it in section 6 as a question instead.
Section 6 is where the agent's inferences live, converted into questions. That is no consolation prize: half the pack's value is a list of things somebody should have asked.
Mark the boundary in the document itself
Each section of the finished pack carries a marker: assembled, or written. Keep the markers even when nobody checks. A traceable document can be defended if a figure is challenged, and a team that marks the boundary learns, over a quarter, how much of its QBR effort is retrieval. Most teams guess low.
Try this now
Run the QBR brief for ACC-02858, then read one section of what comes back.
The brief below refers to 20-briefs/qbr-pack.brief.md, which you create first from the seed at templates/M8/qbr-pack.brief.md in your practice pack.
OUTCOME
Produce the retrieval half of the QBR pack for ACC-02858.
INPUTS
10-data/accounts.csv, product_usage.csv, support_tickets.csv, opportunities.csv,
activities.csv, notes/qbr/, notes/call-notes/
Rules: 40-reference/health-rules.md for the risk section.
OUTPUT SHAPE
30-output/retention/qbr/ACC-02858.md
Sections 1 to 6 exactly as listed in 20-briefs/qbr-pack.brief.md.
Each section header carries "[assembled]".
Every figure carries a source line: file plus record id, or document plus
section heading.
CONSTRAINTS
Do not write a narrative, summary, recommendation, ask or next steps.
Where a source marks a figure as customer-reported or unverified, keep that
marking in the output in the source's own words.
Usage trend is stated as a series of months with values, not as an adjective.
DEFINITION OF DONE
Six sections. Every figure sourced. Section 6 is not empty.
Section 2 of what came back:
## 2. Usage and adoption [assembled]
Seat utilisation, monthly:
2026-01 0.692 2026-02 0.633 2026-03 0.579 2026-04 0.548
2026-05 0.520 2026-06 0.471 2026-07 0.412
Six consecutive months of decline. Active seats 153 of 221 licensed in January,
91 of 221 in July.
Source: product_usage.csv, ACC-02858, months 2026-01 to 2026-07.
Payroll sync and forecasting modules both in use every month of the period.
Source: product_usage.csv, feature_payroll_sync = TRUE, feature_forecasting = TRUE.
The series rather than the adjective is deliberate. "Usage is declining" invites a debate. Seven dated numbers do not.
Worked exercise: the two sections that decide whether the pack is honest
About eight minutes. You produce sections 5 and 6 for ACC-02858, Vireo Retail PLC.
The setup
You are the CSM. The renewal is 2026-09-17, six weeks out, and the renewal opportunity sits in Negotiation at 96,000 GBP. Your VP will ask on Thursday whether this renews. You have accounts.csv, product_usage.csv, support_tickets.csv, opportunities.csv, the QBR note at data/notes/qbr/business-review-vireo-retail.md in your practice pack and the call note at data/notes/call-notes/vireo-retail renewal - notes.txt (the filename really contains spaces; the data dictionary warns about these).
Steps
Read the two note files. They are short, and the exercise fails if you only read the CSVs.
Fill the gaps in the starter brief below and run it to produce section 5, open risks with evidence.
Write section 6, "what we do not know", by hand. Three items minimum.
Compare against the reference below, then ask yourself which of the two sections would change what your VP does on Thursday.
Starter brief (fill the gaps marked <<...>>)
OUTCOME
Produce sections 5 and 6 of the QBR pack for ACC-02858.
INPUTS
10-data/accounts.csv, product_usage.csv, support_tickets.csv, opportunities.csv
10-data/notes/qbr/, 10-data/notes/call-notes/
Rules: 40-reference/health-rules.md
OUTPUT SHAPE
Section 5: one row per risk. Columns: risk | evidence line | source | state.
The evidence line format is <<state the three parts>>.
Section 6: a list. Each item names <<what two things must each item name?>>.
CONSTRAINTS
Notes are evidence, never the source of a number.
Where a source marks a figure as customer-reported or unverified, <<what happens
to it?>>
Do not write a recommendation or an ask.
<<which section, if any, may the agent write? state it explicitly>>
DEFINITION OF DONE
Every risk in section 5 has an evidence line pointing at a record or a named
document section.
Section 6 has at least three items.
Hints
Hint 1 (Concept)
The risks are not all one kind: some live in a column, some only in a sentence somebody typed, and one must be inferred from an absence. Sort them into those three kinds first; section 6 writes itself out of the third.
Hint 2 (Move)
Structured half: compute the run length of consecutive monthly declines in seat_utilization (do not eyeball it), count escalated tickets in support_tickets.csv for this account, and check the renewal opportunity's stage and amount in opportunities.csv.
Unstructured half: the QBR document has a section called "Four things moved against us this quarter". Two of those four appear in no CSV column. Those two are the most important risks in the account.
Hint 3 (Shape)
A correct section 5 has five or six rows, at least two sourced to a document rather than a table, and exactly one about something that is not there.
A correct section 6 does not contain "we do not know if this will renew". Every item names a specific unknown and the person who could resolve it. If an item is about a number the customer gave you, check whether the source marked it, and carry the marking.
Reference answer
## 5. Open risks [assembled]
| risk | evidence | source | state |
|---|---|---|---|
| Seat utilisation declining six consecutive months | H1 usage-decline \| ACC-02858 2026-02 to 2026-07 \| 0.633 -> 0.412 | product_usage.csv | open |
| Executive sponsor departed, successor has had no contact | H5 sponsor-loss \| QBR "Four things moved against us" \| left May, successor started 8 June | notes/qbr/business-review-vireo-retail.md | open |
| Two of four regions stopped publishing in March and have not returned | H5 stated-blocker \| QBR "Four things moved against us" \| trigger recorded as an overtime calculation defect | notes/qbr/business-review-vireo-retail.md | open |
| Contract value above the customer's procurement review threshold | H4 commercial \| contract above 50,000, review running \| 96,000 GBP renewal | notes/qbr/ + opportunities.csv OPP-006874 | open |
| Two escalated support tickets on record | H2 support \| support_tickets.csv, ACC-02858 \| 12 tickets, 2 escalated, 2 open | support_tickets.csv | open |
| No executive-level meeting secured for the renewal cycle | H3 engagement \| QBR commitments: 30 minutes with the incoming COO requested, agreed to try, not agreed to deliver | notes/qbr/business-review-vireo-retail.md | **not evidenced by any CRM field** |
## 6. What we do not know [written by Niamh Ricci, 2026-08-06]
- Whether the overtime calculation defect is actually fixed, and how the customer
would verify it themselves. The QBR records it as shipped and unverified by
them. Product liaison can answer the first half; only the customer can answer
the second.
- Whether the incoming COO has any awareness of us. The record shows no contact.
Absence of logged contact is not proof of absence of contact, and only Helena
Jonker can tell us which it is.
- What the two northern regions are using now, and whether returning is
technically possible before September. Nobody has asked.
- The payroll sync time saving of around a day and a half per fortnight is the
customer's own estimate, reported on the QBR call and not audited by Talvera.
It is the strongest value claim in the account and it is not ours to state.
- Whether the procurement review has a fixed date. The QBR records that a review
of contracts above 50,000 is running. It does not record when it concludes.
How a correct answer may differ. Different rule names and thresholds change the evidence strings, not the findings. A seventh risk for adoption score decline is defensible, slightly redundant with the first. Omitting the escalated-ticket row is defensible if your threshold is three rather than two and your rules file says so.
Not defensible: fewer than two document-sourced risks in section 5, because the two largest risks here exist only in prose; a section 6 shorter than three items on an account in this condition; or restating the day-and-a-half payroll figure without its marking, the Module 7 laundering failure in a different document.
Worth knowing: the two risks that will decide this renewal, a sponsor who left and two regions that stopped using the product, are absent from every column in every CSV. That is the honest answer to why the health score said 36 and nobody moved.
What you keep
20-briefs/qbr-pack.brief.md, with the fence in it.
30-output/retention/qbr/<account-id>.md, one pack, with [assembled] and [written by] markers on every section.
Check yourself
A QBR pack arrives with a well-written two-paragraph narrative about the customer's year. What is the boundary violation, and which fit-test factor is behind it?
Your pack's usage section says "adoption has been trending downward". Rewrite it, and say what the rewrite prevents.
Which of these belongs in section 6 rather than section 5: (a) two escalated tickets; (b) a suspicion that the customer is talking to a competitor; (c) a renewal date with no opportunity; (d) whether the champion still has the budget?
Your account owner asks you to remove section 6 before the customer sees the pack. What do you say, and what do you actually do?
Where this goes next
L8.3 answers the question that makes all of this recur: how does it run next month, what happens when it does not, and how would you know.
Module 8 · Retention, expansion, and work that repeats
M8 lesson / about 14 min
Concept
Scheduling, the failure plan, and the trigger you do not have
Do this nowStart the heartbeat pre-flight90s
Schedule a trivial recurring task that writes the current date and time to ops/heartbeat.md, once a day.
In ops/scheduled-tasks.md record the scheduled time, today's date, and what you are about to do to the machine.
Close the desktop app and let the machine sleep.
You know it worked when: ops/scheduled-tasks.md has a dated row reading "execution mode: under test", and you open ops/heartbeat.md at the start of your next session to finish the row.
What you type
Amend 20-briefs/renewal-risk.brief.md so that every run appends one dated entry
to 50-runs/RUN-LOG.md, including runs that flag nothing, and writes a dated
output file even when the findings are empty.
The entry must state: accounts evaluated, renewal dates in the next 150 days or
already past (split into past and upcoming), newly flagged this week, exceptions
with their reason, and the completion time.
Do not change the risk rules. Show me the last two entries the amended brief
would have produced.
What comes back
50-runs/RUN-LOG.md
## 2026-08-03 renewal risk pass
Accounts evaluated: 102 scored, 87 live customers
Renewal dates in the next 150 days or past: 47 (8 past, 39 upcoming)
Newly flagged this week: 0
Exceptions: 4 accounts with no rows in product_usage.csv
Run completed 07:04
## 2026-07-27 renewal risk pass
Accounts evaluated: 102 scored, 87 live customers
Newly flagged this week: 1 (ACC-03107)
Run completed 07:03
Nothing dated 2026-08-10 or 2026-08-17. The output folder holds no file newer
than renewal-risk-2026-08-03.md.
Notice: The 2026-08-03 entry proves 102 accounts were evaluated to produce a zero, and the two missing weeks below it are the only reason you can tell a quiet report from an absent one.
With no event trigger available, cadence plus saved state substitutes for it, and the price is a detection latency you state in writing rather than discover.A scheduled report that ran and found nothing and one that never ran produce the identical observation, and only a positive run signal written on every run tells them apart.
Why this matters
Your renewal risk pass works. Now it has to happen every Monday for two years, including the Mondays you are on holiday, the Monday your laptop is in a bag, and the Monday you forget it exists.
Most courses say "now schedule it" and move on. We stay, because the interesting question is what happens on the week it does not run, and how you would ever know.
What you will be able to do
State what scheduling in this product is, and what our fact base cannot tell you about it.
Run a pre-flight check that establishes which execution mode applies to you, and record it.
Write the four-question failure plan before creating any schedule.
Design a cadence-plus-check substitute for an event trigger, and state its detection latency.
What scheduling is, and the part that is unresolved
What it is: you describe a task once and Claude runs it on a recurring basis or on demand, delivering reports, briefings and summaries. For a weekly risk list, exactly the shape of the job.
Now the honesty, because you are about to make a business promise on top of it: our fact base carries an unresolved conflict about how scheduled tasks execute.
One reading, from post-July 2026 sources: scheduled tasks run remotely, so they run on their cadence even when your computer is asleep or the desktop app is closed.
The other, older reading: tasks execute only while the computer is awake and the desktop app is running.
The likely resolution is that the second describes pre-July local execution and the first describes post-July cloud execution. That is unverified, and this course will not assert it.
Whichever applies to you, one claim our fact base does make plainly: scheduling is convenience automation and should not be treated as a reliable mechanism for tasks that must run at an exact time without fail.
You are going to find out yourself: the answer for your machine, your plan and your build is the only one that matters, and a pre-flight you ran beats a sentence you read in a course.
Try this now
The pre-flight check.
1. Schedule a trivial task: write the current date and time to
ops/heartbeat.md, once a day.
2. Note the scheduled time, and the date you set it up.
3. Close the desktop app. Put the machine to sleep. Leave it.
4. After the scheduled time has passed, open ops/heartbeat.md.
5. Record in ops/scheduled-tasks.md: execution mode observed, the date you
verified it, what you did (app closed, machine asleep), and the result.
6. Re-run this check after any app update, and re-date the entry.
It takes elapsed time, not work time: start it at the beginning of Part A of LAB-08, the exercise sheet in labs/LAB-08 of your practice pack (start with its README), and read the result at your next session. And record the result even if it is inconvenient. A missing output teaches you more: your reporting depends on a machine being awake, and that belongs in the register and the rollout brief.
The failure plan: four questions before the schedule exists
Mandatory for every scheduled revenue task. Answer in writing, in the register, before you create the schedule.
Question
Why it is the question
1. How would I know it did not run?
Failure here is silent. There is no error in your inbox.
2. What breaks downstream if it does not run?
This is the one that decides whether the workload belongs here at all.
3. What is the manual fallback, and how long does it take?
If you cannot answer this in minutes, you do not have a fallback, you have a hope.
4. What is the maximum acceptable delay?
Turns "it should run weekly" into a number somebody can be held to.
The rule: if the answer to question 2 is "the business", this workload does not belong on a schedule you cannot guarantee. A commission run, a billing file, a regulatory submission, a customer-facing report with a contractual deadline: none of those. They go where a guaranteed scheduler lives, and L8.4 tells you when to move.
What does belong here: the weekly risk list, the Monday movement pack, the monthly hygiene audit. Work where a day late is an inconvenience and a human notices the absence within the acceptable delay, because you built something that makes them notice.
No event triggers, and the pattern that replaces them
There is no inbound email address and no native "when a deal moves to Stage 3, run this". This is the sharpest architectural limit in the product for revenue work, and the clearest boundary between this tool and a workflow engine or CRM automation.
The substitute is cadence plus check:
1. Run on a cadence.
2. Save the state you observed, dated, in 30-output/state/.
3. Next run, compare current state against the saved state.
4. Act on the difference. Save the new state.
That gives you "tell me when this changes", with a lag. The lag is called detection latency, and it is a number you choose and accept in writing:
Cadence
Detection latency
What that means on a Friday afternoon
Daily
up to 1 day
Friday change seen Saturday
Weekdays only
up to 3 days
Friday change seen Monday
Weekly, Mondays
up to 7 days
Friday change seen the following Monday
Monthly
up to 31 days
Do not use this for anything with a renewal date in it
Write it in the register in this form: "A renewal date passing without a renewal opportunity is detected within 7 days. The maximum exposure is a renewal that lapses on a Tuesday and is seen the following Monday."
Then the honest alternative, said out loud to whoever asked: if that latency is unacceptable, the trigger belongs in the CRM or a workflow engine. Wrong tool for the trigger, not for the analysis: keep the analysis here and put the trigger where triggers work.
The register entry
Write it in ops/scheduled-tasks.md, before the schedule is created; a seed register is at templates/M8/scheduled-tasks.md in your practice pack. This is the artifact.
| field | value |
|---|---|
| Task | Weekly renewal risk list |
| Brief | 20-briefs/renewal-risk.brief.md |
| Cadence | Mondays, 07:00 local |
| Output | 30-output/retention/renewal-risk-YYYY-MM-DD.md |
| Positive run signal | Appends a dated line to 50-runs/RUN-LOG.md on every run, |
| | including runs that flag nothing |
| Execution mode | Verified 2026-08-06: app closed and machine asleep overnight, |
| | output present next morning. Re-check after any app update. |
| Detection latency | Up to 7 days. A renewal date passing with no renewal |
| | opportunity is seen at the next Monday run. |
| How I would know it did not run | Monday routine: check that RUN-LOG.md has a |
| | line dated this week. Absence of the line, not absence of an alert. |
| What breaks if it does not run | One week of delay on a risk list. Nothing |
| | contractual, nothing customer-facing, no SLA. |
| Manual fallback | Run the brief by hand. 12 minutes measured on 2026-08-06. |
| Max acceptable delay | 7 days. Beyond that, escalate to the CS lead. |
| Owner | <name> |
Note the "positive run signal" row. The report writes a line every time it runs, including the weeks it finds nothing. That one design choice makes the next section possible.
Tell these apart: ran and found nothing, or did not run
In your inbox, your folder and your week, these two are identical. Nothing arrived.
It ran and found nothing
50-runs/RUN-LOG.md
## 2026-08-03 renewal risk pass
Accounts evaluated: 102 scored, 87 live customers
Renewal dates in the next 150 days or past: 47 (8 past, 39 upcoming)
Newly flagged this week: 0
Still flagged from previous weeks: 12
Field profile step 0: pass, no owner range anomaly
Exceptions: 3 accounts, no usage rows
Run completed 07:04
Zero new flags, and the log proves 102 accounts were evaluated to produce that zero. This is a result. You can act on it, quote it, compare it to next week.
It did not run
50-runs/RUN-LOG.md
## 2026-07-27 renewal risk pass
Accounts evaluated: 102 scored, 87 live customers
Newly flagged this week: 1 (ACC-03107)
Run completed 07:03
That is the last entry. Nothing for 2026-08-03, nothing for 2026-08-10. The output folder holds renewal-risk-2026-07-27.md and nothing newer. Two weeks in which nobody was told anything about the renewal book, and nobody noticed, because a report that finds nothing and a report that does not exist feel the same: a quiet inbox.
The difference is only visible if the report emits a positive signal on every run. Absence produces no evidence of itself; you manufacture the evidence, then build a habit of looking for it.
What honest feedback on this looks like
Here is the response a learner receives when their register's "how I would know it did not run" row reads: "I would notice, because I would not get the report."
That answer describes exactly the situation in which you will not notice.
In week one the report runs, evaluates 102 accounts, flags nothing: nothing arrives. In week two it does not run at all: nothing arrives. Your detection mechanism is "nothing arrived", which cannot tell them apart. The two are the same observation.
The failure of a scheduled report is silent. No error, no red mark, no retry notice. The absence is the error, and absence generates no signal by itself.
What would have caught it. A positive run signal: a dated line appended to 50-runs/RUN-LOG.mdevery run, including runs that flag nothing, plus a Monday routine that checks last week's line exists. You are then looking for the presence of something rather than the absence of something, which a human can actually do. Stronger still: write a dated file every run even when the findings are empty, so a gap in the filenames shows at a glance.
What to change. Two rows of your register: "Positive run signal" becomes a real mechanism, and "How I would know it did not run" names the check, the person, and a cadence. Then answer question 2 again, honestly. If "what breaks if it does not run" was serious, ask whether this workload belongs on a schedule you cannot guarantee at all.
One thing you got right, worth keeping: you did not claim the product would alert you. It will not, and a register that assumed it would is a worse artifact than one with a weak detection habit.
When this goes wrong
The break. Schedule the renewal risk report. Then reproduce the run that does not happen: check for the output where the register says it should be, on a day when your recorded pre-flight conditions are not met. If your pre-flight found local execution, that is any day your machine was asleep at the scheduled time. If it found remote execution, use the lab's file-based stand-in.
The tell. Nothing in the output folder, and nobody noticed. That is the entire tell. No error, no notification. The report simply did not appear, in a week where an unremarkable report also would not have appeared, on a Monday when you were busy.
The cost in the shipped scenario is concrete: the renewal risk pass is what would have surfaced ACC-03043, whose renewal date passed on 2026-07-28 with no renewal opportunity on the record. Two missed runs is two more weeks in which the only person who knows is the CSM who wrote it in a call note in April.
The guard. Two parts, and neither works without the other.
The positive run signal: a dated line in 50-runs/RUN-LOG.md on every run, including empty ones, and a dated output file even when the findings are empty.
The habit that reads it: your Monday routine checks that last week's line exists. Put it in the register as a named person and a named check, not an intention.
Worth knowing: the general lesson transfers to every scheduled thing you will ever build. Detecting absence requires a positive signal, because absence produces none. Every monitoring system you have ever admired is built on that sentence.
What you keep
ops/scheduled-tasks.md, the register, with one complete row: cadence, output location, execution mode as verified with the date, detection latency accepted, positive run signal, the four-question failure plan, and an owner. This is the module's signature artifact.
A cadence-plus-check design with its latency written as a sentence.
Check yourself
Four candidate workloads for a schedule: (a) the Monday pipeline pack; (b) the monthly hygiene audit; (c) the quarterly commission statement file that finance sends to payroll on the 3rd; (d) the weekly renewal risk list. Which one do you reject, and name the reason in the product's terms.
Your manager asks for an alert when any account's health score drops below the threshold. Design the substitute and state the latency you are accepting.
A scheduled report has produced no output for three weeks and nobody noticed. Name the likely cause classes, then name the thing that actually failed.
Somebody proposes removing the run-log line "because it is noise when there is nothing to report". Answer in two sentences.
Where this goes next
L8.4 covers the remote surface, writes the rollout brief a manager approves, and names the conditions under which a workflow should leave this tool entirely.
Module 8 · Retention, expansion, and work that repeats
M8 lesson / about 12 min
Concept
The remote surface, the rollout brief, and knowing when to leave
Do this nowName every approved workflow60s
Open ops/rollout-brief.md and start section 1 as a table.
Add one row per workflow you actually run: name, blast-radius class, named approver, reversal plan.
Write one line underneath: anything not in this table is not approved.
You know it worked when: Section 1 holds at least one row whose approver column contains a person's name rather than a role.
What you type
Run the brief in 20-briefs/renewal-risk.brief.md against 10-data/ as it stands
this morning, and the brief in 20-briefs/account-research.brief.md for ACC-02830.
Write both outputs to 30-output/ under today's date, append both to
50-runs/RUN-LOG.md, and message me the file paths with the headline counts.
Send nothing to anyone. Write nothing outside 30-output/ and 50-runs/.
What comes back
Both tasks finished.
30-output/retention/renewal-risk-2026-08-06.md
102 accounts scored. 47 renewal dates in the next 150 days or already past.
8 of those are past with no renewal opportunity dated to match, 413,600 USD.
30-output/demand/account-briefs/ACC-02830.md
Wexley Care Group GmbH, Enterprise, Germany, EMEA. Expansion OPP-007532 open
at 74,000 EUR, close 2026-09-20. Seat utilisation 0.924 in 2026-07 with the
payroll sync module not in use. Unknown section: 2 items.
50-runs/RUN-LOG.md appended, 2 entries. Nothing sent.
Notice: It had exactly the folder access you granted in Module 2 and you were not there to steer it, which is why the fence matters more here than the instruction does.
Why this matters
You have a workflow, the register you wrote in L8.3 and a failure plan. Now somebody senior wants three more people doing it, and they will ask questions you have not thought about: who approved this, what can it see, what happens when it is wrong, and what does our security team say.
Answer those badly and the answer is no. Answer them in a document, before you are asked, and the answer is usually yes with conditions, which is the outcome you want.
What you will be able to do
Describe the remote surface accurately, including what it does not change.
Write a one to two page rollout brief a manager can approve or reject on its own terms.
State the audit position in both halves, honestly.
Name the conditions under which a workflow should leave this tool.
The remote surface, and the caution that comes with it
Dispatch lets you trigger work from your phone: you send a task from the Claude mobile app, the desktop agent picks it up, executes it with your local files and connectors, and messages you when the result is ready. Reported scope from a phone trigger: spinning up subagents, reading and writing local folders, connecting to MCP servers such as Gmail, Google Drive and Slack, and executing files.
Parallel tasks run genuinely in parallel, each with its own context, file access and connectors, and you can move between them. Any hard cap on concurrency is unverified: plan for "several", not a number.
Computer use lets Claude drive applications on your computer for tasks assigned through Dispatch. It is not available in the Linux beta, and the Linux beta's scope is itself unverified.
The revenue application, concretely: from a taxi you send "run the renewal risk pass and the account brief for the account I am seeing at 2", and you read the output when it lands. A real hour saved on a real day.
The caution, equally concretely: a task triggered remotely has exactly the folder access you granted in Module 2, and you are not there to steer it. The directory boundary being the real boundary applies harder when you are not watching. Fenced properly, nothing changes; fenced badly, this is the moment it matters. Re-read the SharedRoot disclosure and the mitigation posture in Module 2 before making remote triggering part of your routine.
The rollout brief
ops/rollout-brief.md, one to two pages, which you write now from the seed at templates/M8/rollout-brief.md in your practice pack, for a manager who will ask hard questions. Six sections, and the fifth is the one that gets it approved.
1. Which workflows are approved, at which blast-radius class, with which approver. A table: workflow, class (read-only, draft-only, writes-to-system), named approver, reversal plan if applicable. Anything not on the table is not approved, which is the point of a table.
2. Workspace layout and fenced folders. From Module 2: which folder is granted, what is deliberately not in it, who checks. Whatever is in the folder is in scope, so the folder is the control.
3. The never list. From Module 3, plus what Modules 5 to 8 added: the survivor decision on a duplicate merge, the forecast call, the QBR narrative and the ask, anything that touches money, anything customer-facing without a named reviewer.
4. Who reviews what before anything reaches a customer. From Module 7's review-gate.md, which you wrote in L7.4, referenced, not repeated.
5. The audit position, in both halves. Say both, in this order:
What exists. On Enterprise there is an admin dashboard and an Analytics API giving per-day aggregated Cowork usage, per-user activity, skill and connector invocations, and DAU/WAU/MAU, plus role-based access controls, spend limits and use controls, and OpenTelemetry events covering tool and connector calls, files read or modified, skills used, and whether each AI-initiated action was approved manually or automatically, compatible with SIEM pipelines. Real telemetry, genuinely useful.
What does not exist. The Compliance API and Audit Logs do not cover Cowork. Telemetry is not the compliance audit log a regulated team may be required to hold. For any workflow that touches a system of record, our audit trail is the one we build: the run log, the before-state export, and the named approver per run.
Say both halves. A manager who learns the second half from somebody else stops believing the first. A manager who reads both in your document concludes you should own this.
6. The usage position. One shared pool across chat, Claude Code and Cowork, with a 5-hour rolling window plus a weekly cap, and heavy multi-step work consuming materially more allowance than chat. Two consequences a manager needs: a team that moves reporting into this tool should expect chat allowance to be affected, and spend limits exist on the enterprise plan. Never teach a number for the allowance: the published figures are unverified and change, so teach the shape and link out.
Knowing when to leave
A course that never tells you to stop using the thing it is selling is selling rather than teaching. Keep this list. Any one of these means the workflow has outgrown this tool.
Exit criterion
What it looks like in practice
It needs an event trigger with low latency
"Alert the AE within the hour when a deal moves to Stage 3." No event triggers exist, and the cadence-plus-check latency is not acceptable. Move the trigger to the CRM or a workflow engine. Keep the analysis.
It must run at an exact time without fail
The commission file that finance needs on the 3rd. Scheduling is convenience automation. Move it.
It must be auditable in a way you cannot evidence
A regulated process needing a compliance audit log. The gap is real today. Move it, or keep it draft-only forever with a human applying every change through the system's own change process.
It is now stable, deterministic and high volume
The same transformation on 40,000 rows every night, unchanged for six months. A purpose-built pipeline is cheaper and faster, and the judgment that justified an agent has been designed out.
The verification cost has grown past the work
You spend 40 minutes checking output that takes 30 minutes to produce by hand. The Module 4 rule reasserting itself. The honest answer is to stop.
Two things about that table. Leaving is a success state: a workflow that graduated to a pipeline was specified by you, and the brief, rules file and run log are the handover document. And most of what you build will not leave. The list exists so the two or three that should, do, rather than being defended for another year because somebody built them.
Try this now
Draft section 5 of your own rollout brief. Here is the version that decides the meeting, written for a real team:
## 5. Audit position
We can evidence, per run: the brief that ran, the inputs, the reconciliation
result, the sample checked, the exceptions raised, and the named person who
reviewed the output. That evidence lives in 50-runs/RUN-LOG.md and in the dated
output files, both inside our own workspace, both under our normal file
retention.
On Enterprise, the platform gives per-day aggregated usage, per-user activity,
skill and connector invocations, RBAC, spend limits, and OpenTelemetry events
covering tool and connector calls, files read or modified and whether each
AI-initiated action was approved manually or automatically, SIEM compatible.
The Compliance API and Audit Logs do not cover Cowork.
Platform telemetry is not the same thing as a compliance audit log.
What this means for the proposal in front of you: every workflow in section 1 is
read-only or draft-only, so no system of record is changed by an agent and the
audit gap does not bite. The moment we propose promoting one to
writes-to-system, this section has to be re-approved by <security owner>, and
the promotion conditions in 40-reference/field-policy.md apply in full.
The last paragraph makes the document approvable: no blanket yes, just what is in scope now and the event that would require asking again.
What you keep
ops/rollout-brief.md, one to two pages, six sections, ready to send.
The exit criteria list, kept alongside the register, re-read once a quarter.
Check yourself
A rollout brief omits the audit position entirely. Name what a manager cannot decide without it, and what they will probably decide instead.
Three workflows: (a) a nightly transformation of 40,000 rows, unchanged for six months, verified clean every time; (b) a weekly renewal risk list; (c) a Monday pipeline pack. Which should graduate off this tool, and what is the deciding criterion?
Your VP says "so it can just watch the CRM and tell us when something changes". Answer in two sentences, then say what you can offer instead and what it costs in latency.
A colleague wants to trigger the hygiene pass from their phone while travelling. Name the one thing you check first, and why the answer is different from when they run it at their desk.
Where this goes next
LAB-08, the exercise sheet in labs/LAB-08 of your practice pack (start with its README), runs the retention pack and the register on the shipped dataset, then on your own book behind the redaction gate, including the execution-mode pre-flight you start early and read at the next session.
Then the capstone, on your track. Whatever you build, it requires a register entry and a failure plan: the part that proves you understood what you were operating.
Two things to do in the week after you finish, worth more than another module:
Run one workflow twice, for real, with the run log filled in both times. Everything here is theory until the second run happens on a week you did not feel like it.
Re-read the exit criteria against what you built. If one of your workflows already meets one, say so out loud to whoever you built it for. That sentence is the most credible thing you can say about this work.
Module 8 · Retention, expansion, and work that repeats
Lab / ? min / draft-only
Lab
Renewal risk, and the report that did not run
LAB-08: Renewal risk, and the report that did not run
The situation
Two problems, and they are the same problem.
You have a renewal book with a contract anniversary on every account, and you currently discover that an
account is at risk when the customer tells you. By then the decision is mostly made. The signals were in
the data: the shipped dataset is built so that accounts heading for a difficult renewal show six
consecutive months of falling seat utilisation before the renewal date (DATA-DICTIONARY.md,
product_usage.csv). Six months of warning, unread.
Meanwhile the weekly executive report is a Sunday night job, every week, forever, and it is the same
report.
Both are the same shape: work that repeats on a cadence, over data you already have. This module builds
the retention pack and then answers the question the whole course has been deferring: how do you make
it happen again next week without you, and what do you do about the fact that the scheduling is not
guaranteed.
What you will produce
File
What it is
40-reference/health-rules.md
Signal families, thresholds, and the evidence line each flag emits
30-output/retention/renewal-risk-2026-08-02.md
The risk list, every flag carrying the rule and the value that fired it
30-output/retention/qbr/<account-id>.md
One full QBR pack, including a required "what we do not know" section
ops/scheduled-tasks.md
The register. This module's signature artifact. One row per scheduled task with cadence, output location, execution mode as verified with the date, detection latency accepted, failure plan and owner
ops/rollout-brief.md
The manager-facing approval document, one to two pages
Outcome accounting
Two baselines, because two different things repeat. MODULES/08 requires both, and the weekly one is
the honest one because it recurs.
BASELINE 1: QBR preparation, one account
Manual: 41 minutes, measured. The lab author assembled one QBR pack by hand: account row,
opportunity history, 17 months of usage, ticket history, activity history, then read
the business review document and wrote the sections. Timed once, on one account.
With Cowork: 11 minutes of running plus 6 minutes of verification = 17 minutes.
Net: 24 minutes saved per account, checking included.
BASELINE 2: the weekly renewal risk report
Manual: 34 minutes, measured. Join usage to accounts, apply the thresholds by hand, pull
tickets, write the list.
With Cowork: Run one 24 minutes (of which about 14 is writing health-rules.md).
Run two onward 9 minutes, including the 5-minute verification.
Net, run one: 34 - 24 = 10 minutes saved.
Net, run two: 34 - 9 = 25 minutes saved, every week.
Over a quarter, 13 weeks, with verification inside every number:
By hand: 34 x 13 = 442 minutes
With method: 24 + (12 x 9) = 132 minutes
Saved: 310 minutes, about 5 hours 10 minutes
Verification cost: QBR pack 6 min: reconcile the usage series against source, check the two named
tickets, confirm every section that is agent-assembled is marked as such.
Risk list 5 min: profile the health field by owner, sample 5 flagged accounts
against source, reconcile the account count.
Counted separately and inside every net above.
Your actual: _____ QBR by hand _____ QBR with the method
_____ weekly by hand _____ run one _____ run two
Elapsed time is not work time
One step in this lab cannot be compressed: the execution-mode pre-flight requires scheduling a
trivial task, closing the app, and coming back later to see whether the output exists.
That wait does not count against the 50 minutes.part-a.md step 1 starts it at the very beginning,
you do the rest of the lab while it sits there, and you read the result at the start of your next
session. Time budgets in this curriculum are work minutes (PEDAGOGY.md §2.5).
Nothing in Part B Lite requires waiting at all. Lite transcribes the finding from Part A into the
register rather than re-deriving it.
Blast radius
Draft-only. The risk list and the QBR pack are read by a human and acted on by a human. Nothing here
writes to a CRM, and the scheduled task writes only into 30-output/ and 50-runs/.
Obligation
This lab
Review gate
The CSM reads every flag before it reaches an account conversation; the QBR pack's agent-assembled sections are marked as such
Approver
The CS lead for the risk list; the account owner for anything in a QBR pack
Reversal plan
Do not send. Delete the output
Two rubric rules are absolute in this lab (MODULES/08): a register entry with no failure plan fails
Safety at 1, and a scheduled task with no positive run signal fails Safety at 1.
scale-mismatch is the discriminating catch and it is the subject of failure lane 1. It is invisible
in any single row and only appears when you profile the field by owner.
Hard-class coverage requires an assertive identification and accumulates across the four scored core
labs toward the 3-of-6 bar (SCORING.md §2, §7).
How you verify your own output
Profile the field before you threshold it. Minimum, maximum and distribution by owner, on
every field a rule thresholds. Stop if one owner's range does not look like everyone else's. This is
a permanent step in the brief, not a one-off.
Reconcile the account count. Customers considered, flagged, not flagged, excluded, summing to the
customer count.
Sample five flagged accounts against product_usage.csv and support_tickets.csv by hand.
Every flag carries the rule name, the record it fired on, and the value. A single composite health
score with no components is worse than no score, because a CS lead cannot act on a number.
Check the register against reality: for each scheduled task, go and look for last week's output
where the register says it should be. If it is not there, that is the finding.
Files you need
accounts.csv, product_usage.csv, support_tickets.csv, activities.csv, opportunities.csv,
notes/qbr/, notes/call-notes/ and DATA-DICTIONARY.md.
What comes next
This is the last core module. The register and the failure plan are required in every capstone variant
regardless of which workflow you build, and ops/scheduled-tasks.md is one of the three governance files
every track produces.
Module 8 · Retention, expansion, and work that repeats
7 questions / pass at 70%
Checkpoint
Module 8 checkpoint
One question at a time. Pick an answer to see why each option is right or wrong. Nothing leaves this browser.
Question 1 of 7
Four workloads are proposed for a schedule.
Monday 07:00: assemble the weekly pipeline movement pack into the output folder.
Daily 06:00: refresh the renewal risk list from usage, tickets and activity.
Last working day of the month, 16:00: produce the commission calculation file that payroll processes
the same evening. Payroll cannot run without it.
Friday 17:00: summarise the week's support escalations by account.
Which one do you reject, and what is the reason in the product's terms?
Hint
Ask the four failure-plan questions of each workload. Three of them have an acceptable answer to "what
breaks downstream". One of them has an answer that ends the conversation.
A. Incorrect. Allowance consumption is a real consideration and it is a reason to size or stagger a workload, not to reject it. A daily renewal risk refresh has a manual fallback (run it yourself), a tolerable delay (a day) and nothing downstream that stops if it is late. Rejecting on cost while accepting a payroll dependency gets the risk ordering backwards.
B. Correct. You applied the failure-plan question rather than a feeling about importance, and it gives the same answer every time. The four questions are: how would I know it did not run, what breaks downstream if it does not, what is the manual fallback and how long does it take, and what is the maximum acceptable delay. When the second answer is a business process with a hard dependency, the workload does not belong on a schedule that is documented as convenience automation (FACTS §7.2).
C. Incorrect. You identified a genuine weakness in workload 4 and drew the wrong conclusion. The problem with a Friday evening report is detection, and detection is solvable: have the run write a dated line to the run log and check for last week's line as part of a Monday routine. A fixable detection gap is not a reason to reject a workload, and it applies to all four rather than only this one.
D. Incorrect. "Schedule it earlier to leave margin" treats an unreliable mechanism as a slow one. Margin helps when the failure mode is lateness. The documented limit is that the mechanism is not guaranteed to run at all, and no amount of margin addresses that. This is the reasoning that puts a revenue-critical dependency on a best-effort schedule.
Question 2 of 7
A scheduled renewal risk report has produced no output for three weeks. Nobody noticed until a CS lead
went looking for last month's file and found the folder empty after 2026-07-06.
Name the likely cause classes and, more importantly, the missing mechanism.
Hint
Separate two questions. Why did it stop, and why did three weeks pass before anybody knew. Only the
second question has a fix that works regardless of the answer to the first.
A. Incorrect. Waiting for someone else to produce a history you did not record is the position this module is designed to keep you out of. The enterprise Compliance API and Audit Logs do not cover Cowork yet (FACTS §5.10, §7.7), so the run history that would answer this question is the one you build, and the reason it is missing is the defect the item is asking you to name.
B. Incorrect. The first half is a real limit and the conclusion throws away the workflow. A monthly renewal risk report with a one-week tolerable delay is a good candidate for a schedule; what it needs is a detection mechanism, not manual execution forever. Abandoning automation every time it fails silently means never automating anything that repeats.
C. Incorrect. You picked one of the two readings of an unresolved conflict and treated it as established. Our fact base carries both readings and marks the question UNVERIFIED, and the honest move is to run the pre-flight (schedule a trivial task, close the app, come back and see whether the output exists) and record the result with the date in the register. Prescribing "leave the machine awake" as a fix asserts an answer we do not have.
D. Correct. You listed plausible causes without asserting one, which is correct because the execution-mode question is genuinely unresolved in public sources and the only way to know is the pre-flight the learner did not run (FACTS §5.3). Then you named the thing that turns a three-week outage into a three-day one: failure here is silent, there is no error in anybody's inbox, and the absence of a file produces nothing. Detecting absence requires a positive signal, which is the general form of this lesson and the reason the run log line is mandatory rather than nice to have.
Question 3 of 7
A CS director asks: "I want to be alerted the moment an account's seat utilisation drops below 50%."
Design the substitute and state its cost.
Hint
Rewrite the director's request as a question about time rather than a question about events. Once it says
"within how long", one of these four options is obviously the design and the others are obviously not.
A. Correct. This is the whole pattern in one answer: a cadence, a saved previous state, a comparison, and a stated latency the requester agrees to. The last clause is what makes it honest rather than a workaround, because "alert me when this changes" and "tell me within a day of it changing" are different commitments, and the second one is the one you can keep. Naming the exit (a workflow engine or a CRM automation) when the latency is unacceptable is part of the answer rather than an admission.
B. Incorrect. There is no such capability. There are no event triggers and no inbound email address in Cowork, so nothing inside it watches a file or a system and fires on a change (FACTS §7.3). This is the sharpest architectural boundary between Cowork and a workflow engine, and telling a director it exists commits you to building something that cannot be built.
C. Incorrect. The cadence is a real dial and this answer turns it without noticing the two costs. An hourly cadence consumes materially more allowance from a pool shared with your chat and Claude Code use (FACTS §4, §6), and it still is not a trigger, so it still needs the saved-state comparison and it still has a latency you have to state. "Close enough that the distinction does not matter" is the sentence that puts a revenue commitment on a best-effort schedule.
D. Incorrect. The first half is nearly right and the second half gives away the value. The literal request cannot be met, and a substitute exists that gets the director most of what they want. Falling back to a manual weekly check delivers worse latency than the automated daily one and costs somebody an hour a week.
Question 4 of 7
Three workflows, all currently running in Cowork.
Renewal risk list, weekly, 300 accounts, rules change every few weeks as the CS team learns.
Lead routing, on every inbound form fill, about 400 a week, the rule set has not changed in
eight months, and a lead that is routed late loses the SDR their five-minute callback window.
QBR pack assembly, on demand, roughly six a month, each one different.
Which should graduate off Cowork, and what is the deciding criterion?
Hint
Read the five exit criteria and check each workflow against all of them. One workflow trips two criteria.
One of the others looks like it trips a criterion and is actually the strongest case for staying.
A. Correct. You named both criteria and you noticed what makes the case decisive: the five-minute callback window is a latency requirement no cadence can meet, and eight months of unchanged rules means you are paying for adaptability you are not using. That combination is the textbook graduation case. Saying so is also the credibility move the course is built on: a course that never tells you when to leave the tool is selling rather than teaching.
B. Incorrect. You have the criterion inverted. Rules that change every few weeks are the reason to keep a workflow here, because a rules file you edit and rerun is cheaper to change than a configured pipeline. Version control for a rules file is a file in the workspace with its history, which is exactly what the CS track lab asks for.
C. Incorrect. Low volume is a reason to question whether a workflow is worth building, not a reason to move it somewhere else. Workflow 3 is also the one where each instance is different, which is precisely what a purpose-built pipeline handles worst and an agent handles well.
D. Incorrect. "It is working" is doing a lot of work in that sentence. Workflow 2 is not working: a five-minute callback window against a cadence-based substitute means routing is late by design, and nobody has said so out loud. Defending a tool past the point where it is the right one is the specific failure this outcome exists to prevent.
Question 5 of 7
A rollout brief submitted to a VP for approval.
# Cowork rollout, revenue team, 12 seats
## Approved workflows
Weekly pipeline pack (read-only), monthly hygiene audit (draft-only),
renewal risk list (read-only). Approver for draft-only output: RevOps manager.
## Workspace standard
Everyone uses the standard folder layout. Fenced: comp, personnel, legal, board.
## Never list
Anything reaching a customer unread. Anything that changes the CRM without a
tested reversal. Forecast calls.
## Review gate
Seller reviews factual claims. Marketing reviews messaging. Legal reviews
product claims and regulated segments.
## Usage
One shared pool across chat, Claude Code and Cowork, rolling window plus weekly
cap. Expect chat allowance to be affected.
What can the VP not decide without the missing section?
Hint
List the areas a manager has to decide on before more than one person does this work. Tick off what the
brief covers. One area is missing, and it is the one where the honest answer has a good half and a bad
half.
A. Incorrect. Cost is a real approval input and it is not what this brief owes. Plan pricing is published and changes, so a rollout brief links to it rather than printing it, and the usage section already carries the operational cost that a VP cannot look up: that Cowork draws on the same allowance as chat and Claude Code.
B. Correct. Audit is the section a VP cannot supply for themselves and the one their own reviewers will ask about first. The answer has to be both halves or it is misleading in one direction or the other: leaving out the telemetry understates what is available and makes the rollout look unmanageable, and leaving out the gap overstates it and sets the team up to discover in an audit that the compliance log they assumed exists does not cover this product yet (FACTS §5.10, §7.7).
C. Incorrect. Workflow selection is upstream of this document, in the delegation policy, and the brief correctly presents the shortlist with its blast-radius classes and approver. A VP who wants to challenge the selection asks for the policy. A VP who wants to know whether the team can evidence what happened has nothing to ask for, because that section does not exist anywhere.
D. Incorrect. Five of the six required areas are covered and covered well, which is what makes this the tempting answer. Count them against the list: approved workflows with classes and approvers, workspace standard and fenced folders, the never list, the review gate, the usage position, and the audit position. The last one is absent, and it is the only one with a legal consequence attached.
Question 6 of 7
A renewal risk list arrives.
Account Health Renewal Action
Grantham Foods 31 2026-11-04 Critical
Ravensworth Care 38 2026-09-19 Critical
Tarnfield Distribution 9 2026-10-02 Critical
Selby Facilities 44 2027-01-11 Watch
The health figure is a composite with no components shown. Why can a CS lead not act on it?
Hint
Pick one row and write the next thing you would do about it. Then notice what you would have to find out
before you could write anything at all, and look at the row whose number does not sit comfortably on the
same scale as the others.
A. Incorrect. "Component detail can be requested" moves the work back to the CS lead and undoes the point of the list. If understanding a flag requires a second round trip, the list is a prompt to do the analysis rather than the analysis, and the time saving evaporates on the first account.
B. Incorrect. ARR is a real omission and it is a prioritisation problem rather than an actionability one. Adding an ARR column tells the CS lead which account to open first, and they still do not know what to do when they open it. Both matter, and the item asks why they cannot act.
C. Incorrect. Two action bands are a reasonable design and more bands would not help. The problem is not the granularity of the verdict, it is that there is no evidence behind any of the verdicts, and a five-band scale of unexplained verdicts is five times as much unusable output.
D. Correct. The trap is that a single composite score with no components is worse than no score, because it looks actionable and is not: "Critical" tells a CS lead to do something and nothing tells them what. Every health flag should emit the rule name, the record it fired on and the value, exactly as in the pipeline module. You also spotted the second-order problem, which is the one that costs people a quarter: a threshold rule is only as good as the assumption that everybody fills the field the same way, and a value far outside the usual range is the signal that the assumption may be false for one owner's book. The check is to profile the field by owner before any threshold runs.
Question 7 of 7
A QBR pack comes back containing, among the retrieval sections, these two paragraphs.
The story of the year. This has been a year of partnership. After a difficult onboarding, the team
has built genuine momentum, and the relationship is now one of mutual trust.
Our ask. We propose expanding to 240 seats and adding the payroll sync module, at a 12% uplift, to
be agreed before the renewal date.
Which boundary has been crossed, and which fit-test factor is behind it?
Hint
Go through the pack section by section and ask, for each one, what record you would open to check it. Two
sections have no answer to that question, and they are the two the human writes.
A. Incorrect. Nothing forbids a number appearing in agent output, and a blanket ban on pricing would remove the commercial position section, which is legitimate retrieval from the contract and opportunity records. What is wrong is not that a price appears, it is that a proposal was made. Retrieving "current contract is 180 seats at this rate" is fine. Deciding to ask for 240 is not.
B. Incorrect. Draft-only is the correct class and the review gate is not the issue. A CSM reading this pack is being handed a plausible narrative and a specific number, and the practical effect is that they edit them rather than write them, which is how an agent ends up authoring the customer story with a human's signature on it. The control for a boundary violation is not a stronger review, it is not producing the section.
C. Correct. You separated the halves the module separates: what happened since the last review, usage and adoption, support, commercial position and open risks with evidence are retrieval, and the narrative and the ask are not. The reason is not squeamishness. A claim about mutual trust has no source of truth to check it against, and a commercial ask depends on precedent, relationship and appetite that no file records, so both fail the Checkability veto. The fix is in the brief: state the split explicitly so the output does not drift into writing the story.
D. Incorrect. Showing the working would improve the second paragraph and would not fix it, and it does nothing at all for the first. An ask with a calculation behind it is still a commercial decision the CSM has to own, and "mutual trust" has no calculation that could be shown. Correctness is the wrong frame: these paragraphs are not incorrect, they are not the agent's to write.