Which tool to choose
It works on any AI assistant, because it relies on dialogue, not on a special feature. The only practical difference: if you work in a single conversation, the AI remembers the answer to review and you just ask it to criticize it. If you want a more ruthless review, open a new chat, paste the answer as if it were by another author and ask for the critique: without the memory of "having written it itself," the AI is less indulgent. If you don't know which is better, try first in the same chat; switch to a new chat if the self-critiques are too soft.
How to do it
- Get the first answer normally. Don't aim for perfection right away: you need a basis to work on.
- Ask for a critique with a change of role. This is the heart of the technique. The operational syntax:
Now change role: you are a strict, independent reviewer who did not write the text above.
List the three most serious problems with the answer, starting from the most important.
For each, explain why it's a problem. Don't rewrite the text yet, only criticize.
- Have it rewrite, correcting the flaws found. Separating critique from rewriting improves both:
Now rewrite the answer correcting the three problems you just listed.
Keep what was working, change only where needed.
- Add your own criteria to the review. If you know what matters, say so: "also check that the numbers are consistent," "check that the tone stays formal," "flag every claim without a source."
- Repeat only if needed. One review is usually enough. A second round makes sense only for important texts, and beyond the second the improvements drop off.
A concrete example
You ask the AI for an email to request a raise. The first version is correct but generic: it talks about "commitment and dedication" without proof. You ask it to review it as a strict editor, and it itself identifies the flaws: no concrete result cited, too submissive a tone, a vague request. You ask it to rewrite, correcting, and the second version inserts a space for measurable results, stiffens the request into a figure and cuts the apologetic sentences. You got in two steps what a single request didn't give, and not because the AI got better: because you put it in the shoes of someone judging instead of someone writing.
When it does NOT work (and how to fix it)
If the self-critique is generic and gentle
Phrases like "clarity could be improved" are useless. Fix: ask for specific, numbered flaws, and impose a strict tone: "be ruthless, quote the exact sentences that don't work and explain why." A new chat with the text presented as someone else's helps a lot here.
If the text gets worse the more you review it
Too many reviews can sand away the character or add useless caution. Fix: stop after one or two passes; if the latest version convinces you less, go back to the previous one ("take the version from two messages ago and apply only the first correction").
If the AI says it corrected but the flaw remains
Sometimes it declares the change without actually making it. Fix: check yourself, and if it's missing, insist by pointing the finger: "the submissive-tone problem is still present in the second sentence, rewrite only that one."
A tip from someone who actually uses it
Always separate the critique from the rewriting, in two distinct messages. If you ask "improve this text" in one go, the AI makes cosmetic touch-ups. If you first force it to say what's wrong, and only then to rewrite, the second version tackles the real problems instead of polishing the surface. The change of role between the two phases is what switches off the instinct to defend its own work.
Frequently asked questions
Why doesn't the AI write well the first time?
Because the first draft aims to cover the request, not to be perfect, and the AI has no judgment criteria unless you give them to it. Asking it to review activates a different mode, that of checking: it's the same reason a human author rereads what they wrote with a cool head.
Is a new chat or the same conversation better?
The same chat is more convenient and often enough. A new chat, where you paste the text as someone else's work, gives harsher critiques because the AI doesn't "recognize" it as its own to defend. Use the new chat when the self-critiques in the same conversation seem too soft.
If the AI reviews itself, can I trust the final result?
More than the first draft, but not blindly. Self-review corrects inconsistencies and weaknesses of form; it doesn't guarantee the facts are true, because the AI can confidently confirm a wrong piece of data. On the content you publish, the AI's review is one more filter, not the last: you verify the facts yourself against real sources.