Three Hats
For writing that matters: a researcher who only finds, a writer who only uses what was found, a reviewer who only flags. One chat, three prompts, you decide.
What it is
One Claude Project that plays three roles, one at a time, only when you name the hat. The researcher lists findings with sources and admits the gaps. The writer uses those findings and nothing else. The reviewer checks the draft against a list you wrote, says PASS or FAIL, and gives three fixes — without touching the text. It's a multi-agent workflow with no agents.
Why three
One AI asked to research, write and check its own work does all three badly at once — it invents a fact, writes around it, then approves it. Split the hats and each one has a rule it can't break.
Why the reviewer never rewrites
A reviewer that edits is a second writer, and now nobody checked anything. Flagging keeps the review honest. You apply the fixes, or you don't.
Fifteen minutes with it
- Research
- Findings, one line each, each with a source. Then "Gaps:" — what it couldn't find. Read the gaps first. That's where the piece will be weak.
- Write
- One draft from the findings only. No options, no preamble, nothing that wasn't found.
- Review
- Your checklist: PASS or FAIL, one line why, three fixes as instructions. Second draft applies those and nothing else. Then you read it as the person it's for.
Good for
A proposal
Where a wrong number costs you the deal. The writer can't invent one; the reviewer checks the ones that are there.
A newsletter people keep
The reviewer's list is your standard — under 400 words, one idea, one ask — and it gets applied every time.
A post that has to be true
Every claim traces to a finding. When someone asks "source?", you have one.
Anything with your name on it
The checklist is yours. Which means the standard is yours, not the model's.
The honest part: this is for writing that matters. Three hats on a two-line email is theatre — the tool this is adapted from says the same: split roles only when the task really benefits from it. And the reviewer is only as good as your checklist. If the drafts keep coming back wrong in the same way, the list is missing a line.
How long?
Where does the research come from?
What should it sound like?
Your hats. Run them in order.
What happens.
- First piece
- The researcher's gaps list will surprise you — it's the things you'd have asserted from memory. The reviewer fails something you thought was fine. Both are the point.
- Third piece
- You've edited the checklist twice. It's now your actual standard, written down, applied every time. That's the asset — the drafts are just the output.
- After a month
- You run "just the reviewer" on things you wrote yourself. It's the same list, and it doesn't care who wrote the draft.
What it won't do
- Have taste. The checklist is the standard, and you wrote it. A vague list gets a vague review.
- Be worth it for small things. A two-line reply doesn't need three hats. Use one.
- Replace reading it yourself. The last step is you, as the reader it's for. Always.
Prefer to read it? The whole kit in plain text, and what we changed in translation.
The kit, in plain sight
This is the general version. The tool above builds yours, with your checklist in it.
Three Hats is one Claude Project that plays three roles in order, one at a time: a researcher who only lists findings with sources, a writer who only uses those findings, and a reviewer who checks the draft against your list and flags fixes but never rewrites. You run the three prompts in one chat and decide at the end. It's a multi-agent workflow with no agents — and it's for writing that matters, not for an email.
# Three Hats — how to run it I'm writing: [what, for whom] Length: [one paragraph / one page / a long piece]. Tone: [plain and direct / warm / formal / my voice]. You wear three hats, one at a time, only when I name the hat. Never two at once. ## Hat one — Researcher (when I say "research") [Search the web, three to six sources, primary over secondary — OR — use only what I paste in — OR — interview me, up to five questions, one at a time.] Output: a list of findings, each one line, each with where it came from. No prose, no opinions. End with "Gaps:" — what you couldn't find. ## Hat two — Writer (when I say "write") Write the piece from the researcher's findings and nothing else. Every factual claim traces to a finding; if it isn't in the findings, it doesn't go in. One draft, no options, no preamble. ## Hat three — Reviewer (when I say "review") Check the draft against this list. For each item: PASS or FAIL, one line why. 1. [e.g. under 400 words] 2. [e.g. one clear ask, in the first third] 3. [e.g. no claim without a source] 4. [e.g. a reader who skims only headings still gets the point] Then: the three most important fixes, as instructions to the writer — not rewritten text. You never rewrite. You flag. ## Rules - The writer never invents facts. The reviewer never rewrites. The researcher never has opinions. - If I say "second draft", the writer applies the reviewer's fixes and nothing else. - Short beats thorough at every stage.
Research. The piece is: [what]. Go. Write. Use the findings above and nothing else. Review. Against the list. PASS/FAIL each, then the three fixes. Second draft. Apply the reviewer's three fixes and nothing else. Show me what changed in one line at the end.
Just the reviewer hat. Check this against my list — PASS/FAIL each, three fixes, no rewriting. --- [paste anything]
What changed in translation
| In the source | Here | Why |
|---|---|---|
| A Python framework: agents with roles, goals and backstories; tasks; tools; a crew that runs them | Three hats in one Project, run by you in order | The roles were the idea. The framework was the plumbing. |
| Agents hand off to each other automatically | You say the next hat's name | You see every handoff, so you catch the researcher's gaps before the writer builds on them. |
| Roles: researcher, writer, reviewer | Kept, with one rule each: no opinions, no invented facts, no rewriting | The rules are what make three hats better than one. |
| A reviewer that can edit | A reviewer that only flags | A reviewer who rewrites is a second writer. Flagging keeps the check honest. |
| "Multi-agent only for tasks that benefit from a division of roles" | Kept — for writing that matters, not for an email | The source's own caveat. Three hats on a two-line reply is theatre. |
Adapted from the researcher → writer → reviewer pattern in CrewAI (crewAIInc/crewAI, MIT, ~58,000 stars, checked 6 September 2026).