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.mdyou 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.mdis 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.