AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people meet Cowork by typing the same loose request they would give a chat window, then sit watching it work on something they never defined. It feels fine for the first five minutes, because Claude produces something. It fails later, when the output is in the wrong format, touched files you did not mean it to touch, or burned a large chunk of your usage on a vague job.
The rule: treat Cowork as a delegate you brief, not a chat you steer. Write a short brief first, start with a task that cannot do damage, and widen access only after a task has gone right.
Anthropic describes Cowork as Claude applying agentic capabilities to knowledge work beyond coding: instead of answering one message at a time, it takes on multi-step tasks and carries them out for you, while you can see what it is doing. It can read and write files in folders you connect on the desktop app, produce Excel spreadsheets with formulas and PowerPoint decks, coordinate sub-agents on parallel strands of work, and run scheduled tasks. Cloud sessions run on Anthropic's servers, so work can continue when your computer is off.
What changed recently matters for beginners, because many tutorials are out of date:
So the difference from a normal chat is no longer a button. It is behaviour: a chat answers; Cowork plans, uses tools, touches files and connected apps, and keeps going. That is also why it consumes more of your usage than a quick question.
Fill this in before every task. It is five lines.
| Line | What goes in it | Example |
|---|---|---|
| Outcome | The finished thing, in one sentence | A one-page summary of the folder |
| Inputs | Exactly which files, apps or links | The three PDFs in the scratch folder |
| Format | File type, length, structure | A Word-style summary, under 400 words |
| Off limits | What it must not touch or do | No deleting, no moving, no emailing |
| Check | How you will know it is right | Every claim points to a file name |
Anthropic's own guidance says to describe the task the way you would brief a colleague: outcome, format and the inputs it needs. The Off limits line is ours, and it is the one people skip.
Here is the card as a reusable prompt.
You are my careful research assistant working in the folder I have connected.
Goal: [OUTCOME]. A good result is [WHAT_GOOD_LOOKS_LIKE].
Inputs: use only [FILES_OR_APPS]. If anything you need is missing or ambiguous, ask me before you start rather than guessing.
Steps: first read the inputs and tell me your plan in no more than five bullets. Wait for my go-ahead. Then do the work.
Output: [FORMAT, LENGTH, FILE_TYPE]. Save it as a new file called [FILE_NAME]; do not edit or overwrite my originals.
Constraints: do not delete, move or rename anything. Do not send messages or post anything. Do not browse beyond [SITES_ALLOWED]. Say so if a step would need something outside these limits.
Before you finish: check each claim against the source file, list anything you could not verify, and tell me what you did not do.
Fill in the capitalised square-bracket items; leave the structure alone.
Climb one rung at a time. Move up only when the rung below produced what you asked for.
Connect your scratch folder, paste the Brief Card above, and ask for a summary or an index of what is inside. Nothing can be harmed, so you learn the rhythm: plan, approve, work, report.
Give it a pile of notes, receipts or exported data and ask for a spreadsheet or deck as a new file. This is where Cowork earns its keep over chat, because it produces the file itself.
You are a careful analyst. In my connected folder, read [FILE_LIST]. Build a spreadsheet with one row per [ITEM_TYPE] and columns [COLUMN_NAMES]. Put anything you are unsure about in a column called "Needs checking" instead of guessing. Save it as [FILE_NAME] without touching the originals, then list any rows you could not complete.
Fill in the files, the row type and the columns.
Connectors let Claude reach apps such as your email or calendar. Add one:
Then ask for a read-only job, such as a summary of this week's calendar. Custom connectors via remote MCP also exist, but skip them until you are comfortable.
Cowork has three: Manual (asks before each action), Auto (read-only tools approved; for write or delete actions Claude judges, and blocks what looks unsafe) and Skip (no automatic checks). Stay on Manual through all three rungs.
/schedule command) for anything that sends, buys or deletes. Anthropic's safety page says to avoid sensitive data and hard-to-undo actions there.