The Role Switch That Makes Claude Actually Critique Its Own Work

Ask it to improve its own draft and you get polish. Ask it to review someone else's and it takes the thing apart.

Ask Claude to improve a document it just wrote, and you'll get a slightly different document. Tidier sentences. A heading moved. Nothing that would embarrass the original.

Ask it to review a document written by someone else, and it will take that document apart.

Same model, same text, completely different output. The only variable is who it thinks wrote the thing.

I found this by digging into it about a year ago, drafting a project charter under time pressure. The self-review pass gave me polish. The reply had a few useful tips in it, so I ran a second check — fresh session, same charter pasted in cold, asked for a critical review of a junior PM's work. It came back with six problems, three of which were real, and one of which would have cost me a difficult conversation two weeks later.

That's this week's workflow.

Why it happens

A model asked to improve its own output is doing continuation. The draft is already in context as something it produced and committed to, and the natural next move is refinement — the same document, slightly better.

A model asked to review someone else's document is doing evaluation. Different task, different behaviour. Nothing is being defended, so gaps get named rather than smoothed over.

You're not tricking it. You're telling it which job to do, and the two jobs genuinely produce different work.

The prompt

Fresh session, and not inside the project that holds the draft.

``` <role> You are a senior editor and PM quality reviewer. You are reviewing a document produced by a junior PM. Find every weakness — do not be polite about it. </role>

<task> Review the document below and identify what is wrong, weak, missing, or unclear. </task>

<context> Document type: [CHARTER / BRD / STATUS REPORT / PLAN / OTHER] Audience: [WHO WILL READ THIS AND WHAT THEY NEED FROM IT] Document: [PASTE] </context>

<format> CRITICAL ISSUES — things that would cause a problem if this shipped WEAK SECTIONS — present but not doing their job, with what is missing UNSUPPORTED CLAIMS — statements presented as fact with nothing behind them MISSING — what a document of this type should contain and this doesn't </format>

<constraints>

• Do not rewrite the document
• Do not open with anything positive
• If a section is fine, skip it entirely

</constraints> ```

Three details are doing most of the work.

Fresh session, outside the project. Two things break this. If the draft is still in the conversation that produced it, the model treats it as its own work and reverts to continuation. And if it sits in a project folder, the model can pull it back in along with the context that shaped it — the same problem arriving by another route. Paste it in cold, somewhere with no history.

"Do not rewrite." Without this you get a polished replacement and no diagnosis. You want the list, not the fix — you'll decide which problems are real.

"Do not open with anything positive." The opening paragraph of strengths is filler, and worse, it sets a tone the rest of the review then has to live up to.

Watch for

It will over-generate. Told to find every weakness, it finds weaknesses, including invented ones. Expect roughly half the list to be noise. That's an acceptable trade — you're scanning for the two or three items that make you stop, and a shorter list would miss them.

It doesn't know your context. It will flag a missing section that you deliberately left out, or call a claim unsupported when the support is in a document it can't see. Your judgment is the filter. This produces candidates, not verdicts.

Don't run it on the same draft twice. The second pass finds new problems because it's been told to find problems, not because they're there. One pass, act on it, move on.

Where it fits

This is a second-pass tool. It's not for improving a rough draft — a rough draft has obvious problems and you don't need a model to list them.

It's for the document you think is finished. The charter you're about to circulate, the BRD going to sign-off, the status report the steering committee will read. The point where you've stopped seeing your own work clearly, which is exactly the point where a stranger's eyes are worth something.

Thirty seconds to run. The failure it catches is the one where a document reads well and is missing something structural, which is the failure your own review is worst at finding.

Try it this week

Take the last document you sent to a stakeholder — one you were happy with. Run it through cold in a fresh session.

If it comes back with nothing, you write better documents than most of us. If it comes back with one thing that makes you wince, that's the pass paying for itself.

This piece went out first to newsletter subscribers. Subscribe on Substack to get them as they publish.