Claude Cowork 101: Delegation Without The Terminal
A field manual for workers who live in files, folders, notes, meetings, PDFs, inboxes, and client messages. The goal is not to make everyone a developer. The goal is to make messy office work delegable without surrendering judgment.
- Day 1 - Read-only summaries. Give one folder or small document set. Ask for a summary and source list. No edits.
- Day 2 - Briefs from bounded sources. Ask for a one-page brief with risks, open questions, and next actions.
- Day 3 - Comparisons. Compare two documents or versions. Ask for material differences only.
- Day 4 - Drafting. Prepare a client update, memo, checklist, or meeting brief. Do not send.
- Day 5 - Folder map. Ask for a proposed organization plan. Still no moving files.
- Day 6 - Local artifact creation. Allow creation of a new draft file in an approved folder.
- Day 7 - Review the workflow. Keep the tasks that saved time. Remove the ones that created confusion.
Claude Code makes sense to developers because developers already live in repositories, terminals, diffs, tests, commits, and pull requests. Most workers live somewhere else.
They live in folders, PDFs, spreadsheets, inboxes, notes, meetings, screenshots, client messages, and half-finished documents. Cowork matters because useful AI cannot stay trapped inside developer workflows.
Cowork is not "AI for non-technical people." It is delegated office execution with boundaries.
The worker still owns judgment. Cowork helps carry the search, comparison, drafting, formatting, and packaging that normally sit invisibly inside office work.
| Mode | What it does | Review surface | Risk |
|---|---|---|---|
| Chat | Answers, drafts, explains, summarizes pasted content | Conversation | Context may be incomplete or stale |
| Cowork | Reads selected files, compares, drafts, creates artifacts, proposes changes | File, table, brief, source list, change log | May act on wrong source or hidden assumption |
| Claude Code | Works in codebases, tests, commits, branches, PRs | Diff, test output, commit | Code changes may exceed scope |
| OpenClaw | Persistent assistant shell reachable through channels and webhooks | Skill output, channel message, completion report | Permission creep if not governed |
The point is not which one is better. The point is which body fits the work. A worker who understands the difference rarely hands the wrong task to the wrong tool.
The artifact is the review surface
When Cowork works on files, you can see exactly what it read, what it produced, and what it assumed. The artifact is inspectable in a way that a chat response is not.
Still needs source discipline
Cowork is only as reliable as the source set you give it. Vague access (“use my files”) produces vague work. Bounded access produces inspectable work.
Cowork Delegation Contract
Task:
[What you want done]
Source material:
[Files, folders, notes, links, or context]
Output:
[Exact format you want]
Boundaries:
[What not to touch, what requires approval]
Check:
[How you want it to verify the result]
This is not bureaucracy. It is how the worker stops the agent from guessing where judgment is required. Anthropic ↗
If the task has consequences, write the boundaries first. Most bad Cowork outcomes start because the assistant was asked to solve an undefined problem with undefined permissions.
Cowork is only as reliable as the material it is allowed to inspect. "Use the project files" is weaker than "Use only the files in /Project Notes/June and ignore archive folders."
| Source type | Good instruction | Bad instruction |
|---|---|---|
| Folder | Use only this folder and list every file used. | Look around and find what matters. |
| PDFs | Summarize these three PDFs and cite page or section names where possible. | Summarize the documents. |
| Notes | Use notes dated after May 1 and flag contradictions with older notes. | Use my notes. |
| Email export | Extract tasks only. Do not draft or send replies. | Handle this thread. |
Source Discipline Prompt
Before drafting, list:
- every source you will use
- every source you will ignore
- anything missing
- anything that looks outdated or contradictory
Cowork should produce something the worker can review: a brief, table, checklist, draft, folder map, source list, task list, or change proposal.
One-page project brief
Summary, timeline, risks, open questions, next actions, source list.
"Here is what I found"
Too vague to approve, repeat, or hand off.
Comparison table
Difference, impact, recommendation, confidence, source.
Long narrative summary
Hard to scan, hard to verify, hard to turn into action.
Never begin folder work by moving files. Begin with a map. Anthropic ↗ frames this as read-before-write: let the system observe before it acts.
Folder Map Prompt
Read this folder and propose an organization plan.
Do not rename, move, delete, upload, or edit anything yet.
Return:
- current structure summary
- file categories
- duplicate or outdated files
- proposed folder structure
- files you are unsure about
- risks before changing anything
Document work becomes valuable when the assistant can extract decisions, risks, contradictions, and next actions instead of merely compressing text.
Document Brief Prompt
Read the selected documents.
Output:
- executive summary
- key facts
- decisions already made
- unresolved questions
- contradictions
- risks
- recommended next action
- source list
Do not invent missing facts. Mark uncertainty clearly.
| Document type | Best output | Extra check |
|---|---|---|
| Proposal | Scope, cost, assumptions, exclusions | Find vague obligations |
| Contract | Non-legal summary and risk flags | Require human/legal review |
| Report | Findings, evidence, recommendation | Separate data from interpretation |
| SOP | Step list, owner, trigger, failure mode | Flag missing decision points |
The useful output of a meeting is not a transcript. It is a decision record, task list, and risk register.
Meeting Debrief Prompt
Turn these meeting notes into:
- decisions
- tasks
- owners
- deadlines
- blockers
- open questions
- follow-up message draft
If owner or deadline is missing, mark it as missing.
Do not invent names or dates.
Spreadsheet work needs extra caution because a small formula mistake can create large business consequences. Use Cowork to inspect, summarize, and flag issues before allowing edits.
| Spreadsheet type | Safe Cowork role | Requires human review before |
|---|---|---|
| Data export / list | Classify, count, summarize, find duplicates | Any structural edit |
| Project tracker | Extract tasks, owners, deadlines, blockers | Changing status fields |
| Budget or forecast | Read totals, flag anomalies, note missing rows | Any formula or value edit |
| Payroll / invoicing | Verify structure and flag missing data only | All of it — escalate to owner |
| Client reporting | Check numbers against source data, note discrepancies | Any change to client-facing figures |
Spreadsheet Inspection Prompt
Inspect this spreadsheet.
Return:
- sheet purpose
- columns and meanings
- obvious data quality issues
- missing values
- suspicious formulas or hardcoded numbers where a formula is expected
- summary statistics
- questions before editing
Do not modify the file.
Do not recalculate or reinterpret formulas — report what is there.
The inbox contains commitments, private context, money, legal issues, client expectations, and emotional landmines. Do not begin by letting an agent "handle email." Begin with triage.
Inbox Triage Prompt
Review the selected messages.
Extract only:
- urgent items
- needs reply
- waiting on someone else
- FYI/archive
- deadlines
- commitments I made
Do not draft, send, archive, delete, unsubscribe, or mark anything read unless I approve.
| Stage | Allowed | Requires approval |
|---|---|---|
| Triage | Classify, extract tasks, summarize | Changing message state |
| Draft | Prepare reply drafts | Sending or reply-all |
| Cleanup | Recommend filters or folders | Unsubscribe, delete, archive |
| Level | Permission | Examples | Default for beginners? |
|---|---|---|---|
| 1 | Read-only | Summarize, classify, compare, extract | Yes |
| 2 | Drafting | Create briefs, memos, checklists, reply drafts | Yes |
| 3 | Local edits | Create or edit files in approved folders | After review |
| 4 | External actions | Send, publish, create tickets, update systems | No |
| 5 | Sensitive actions | Money, credentials, deletion, legal, production | No |
Approval Gate
Before any Level 4 or Level 5 action, stop and ask:
- What action will be taken?
- What source supports it?
- What could go wrong?
- How can it be reversed?
- Do you approve?
Every Cowork output should include a review trail. The worker should not need to guess whether the assistant used the right files or invented missing details.
Cowork Completion Report
Completed:
- [artifact created]
Sources used:
- [file/link/note]
Not used:
- [ignored source and why]
Assumptions:
- [inferences]
Uncertainties:
- [unknowns]
Needs approval:
- [decision/action]
Ask for uncertainty explicitly. If you do not create a place for uncertainty, many assistants will bury it under fluent prose.
Project Brief
Selected notes in, one-page brief out: summary, timeline, risks, questions, next actions.
Folder Map
Approved folder in, proposed structure out. No movement until approved.
Proposal Comparison
Two proposals in, material differences table out: scope, cost, risk, recommendation.
Meeting Action List
Notes in, decisions, owners, deadlines, blockers, and follow-up draft out.
Client Update Draft
Source notes in, client-ready draft out. Do not send.
Inbox Task Extraction
Selected emails in, commitments and deadlines out. No sending or archiving.
SOP Builder
Messy process notes in, trigger, owner, steps, checks, and failure modes out.
Blueprint prompt pack
BP1 - Project Brief
Create a project brief from the selected files.
Use only the selected files.
Output:
- summary
- timeline
- decisions
- risks
- open questions
- next actions
- source list
Do not move, rename, delete, send, or upload anything.
BP3 - Proposal Comparison
Compare these two proposals.
Return a table:
- item
- Proposal A
- Proposal B
- material difference
- operational impact
- question to ask before choosing
BP7 - SOP Builder
Turn these rough notes into an SOP.
Include:
- trigger
- owner
- inputs
- steps
- checks
- failure modes
- escalation point
- example completed output
1. What exact artifact should Cowork produce?
2. Which files, folders, or notes are allowed sources?
3. Which sources should be ignored?
4. What actions are forbidden?
5. What requires approval?
6. How will the output be checked?
7. What uncertainty should be flagged?
8. What would make this task unsafe to automate?
9. Where should the artifact be saved?
10. What is the next human decision?
Pass/fail standard
Pass if the task has bounded sources, visible output, explicit forbidden actions, and a review method. Fail if the prompt says "handle," "clean up," "fix," or "organize" without defining scope and approval.
This library points to the major prompt blocks throughout the manual and fills in the missing field prompts so the page works as a reference desk.
| Prompt | Use it when | Jump |
|---|---|---|
| Cowork Delegation Contract | You need to define task, sources, output, boundaries, and checks. | M2 - Contract |
| Source Discipline Prompt | You need the assistant to declare what it will use before drafting. | M3 - Sources |
| Artifact Types | You need to choose the right output shape before delegating. | M4 - Artifacts |
| Folder Map Prompt | You need a folder organization proposal without moving files. | M5 - Folder |
| Document Brief Prompt | You need decisions, risks, contradictions, and next actions from documents. | M6 - Documents |
| Meeting Brief Prompt | You need meeting notes turned into decisions, owners, deadlines, and blockers. | M7 - Meetings |
| Inbox Triage Prompt | You need tasks and commitments extracted from selected messages. | M9 - Inbox |
| Approval Prompt | You need to stop before Level 4 or Level 5 actions. | M10 - Approval |
| QA Review Prompt | You need source trail, assumptions, uncertainties, and review points. | M11 - QA |
| Blueprint Prompt Pack | You need reusable prompts for project briefs, comparisons, and SOPs. | M12 - Blueprints |
Approval Prompt
Approval Prompt
Before taking this action, stop and prepare an approval request.
Include:
- exact action
- destination or system affected
- source evidence
- risk if wrong
- rollback path
- what you need me to approve
Do not act until approval is explicit.
QA Review Prompt
QA Review Prompt
Review this Cowork output before I use it.
Check:
- sources used
- missing sources
- unsupported claims
- assumptions
- uncertainty
- formatting issues
- actions that require approval
- next human decision
| Term | Meaning |
|---|---|
| Cowork | An AI delegation pattern for office/file work through a friendly interface rather than a terminal. |
| Artifact | The visible output of work: brief, table, checklist, draft, file map, or report. |
| Source set | The bounded material the assistant is allowed to use. |
| Approval gate | A required human decision before consequential action. |
| Visible work | Output plus source trail, assumptions, uncertainties, and next decision. |