Course contents
Write a Task Brief Claude Can Finish
Turn a vague request into a brief with an outcome, a format, inputs, boundaries and a definition of done, so Claude can finish without guessing.
A brief is the whole job description for a task. Claude can't see the meeting you came out of or know that "the numbers" means the Q3 tab and not Q2. Everything it doesn't know, it has to guess or ask about. A good brief removes the guesses that matter and leaves Claude free to handle the rest.
Anthropic's advice is to brief Claude the way you'd brief a colleague, and to include three things: the outcome you want, the format it should come back in, and the inputs Claude should work from. Two more parts make the difference between a draft you fix and a draft you send: boundaries, and a definition of done.
| Part | What it answers | Example |
|---|---|---|
| Outcome | What do you want to end up with? | A one-page summary of what changed in our Q3 numbers |
| Format | What shape, and what file? | q3-update.docx: five sections, and a table of the top ten numbers |
| Inputs | What should Claude work from? | kpis.xlsx (Q3 tab) and the notes in Board/Q3 |
| Boundaries | What must it not touch or do? | Don't edit kpis.xlsx. Don't send anything; save a draft. |
| Done when | How will you judge it? | Every number cites its cell, and anything unverified is listed at the end |
Your colleague wrote this brief for Cowork. Find the five lines that will make Claude guess.
Name the file
"Help me with the board update" has no finish line. "Create q3-update.docx" does. A named deliverable tells Claude when to stop and gives you one thing to open and check. For recurring work, put the date in the name, such as weekly-report-2026-11-13.md, so a new run never overwrites an old one.
Say what not to do
Boundaries are the cheapest safety feature you have. "Don't move or delete anything in Inputs", "Only use files in this folder" and "Save a draft; don't send it" each rule out a whole class of surprises. Anthropic's safety guidance makes the same point about agents in general: make your request specific so Claude doesn't do things you didn't intend.
Match each vague phrase to a precise version Claude can act on.
Pick one on the left, then its partner on the right.
Invite the right questions
End with what Claude should do when it's unsure. "If two sources disagree, ask me before choosing" turns a silent guess into a question. Don't ask it to check every step with you, though; that recreates the work you were trying to hand off. Say which decisions are yours (money, tone, anything sent to other people) and let Claude handle the mechanical rest.
Turn this request into a brief Claude can finish without guessing. Switch on what helps; leave out what hurts.
Summarize the customer interviews.
Keep a brief file for bigger jobs
For anything you'll run more than once, write the brief into a file in the folder, such as brief.md, and start the task with "Read brief.md, then…". The brief travels with the work, you stop retyping context, and you can improve it every time a result disappoints you. Later modules promote your best briefs into project instructions and skills.
A brief file doesn't need to be long. Five short lines cover most jobs, and you can copy this shape for your own:
- Outcome: what you want to end up with, in one sentence.
- Deliverable: the file name and format, plus a reference file to match if you have one.
- Inputs: which files, folders, links or apps to use, and which to ignore.
- Boundaries: what not to touch, and what not to send, post or delete.
- Done when: how you'll check it, and what Claude should flag or ask about.
Plain-English definitions from the AI Coding Dictionary.