What it is
Draft, then critique is a second pass with a rulebook. The draft can be the model’s or your own. You list what it must do, such as name the new delivery date in the first two sentences, and the model reports pass or fail on each, quoting the line that fails. Then, and only then, it fixes the failures and leaves everything else alone.
Compare that with “make this better”. A model asked to improve a draft rewrites all of it, so you get new wording, new emphasis and sometimes new mistakes, and you have to read the whole thing again to find out what changed. A critique against criteria separates the verdict from the rewrite. You can read the verdict, disagree with one item, and ask for fixes to the rest.
When it helps
Use it on anything that goes out under your name: a customer reply, a proposal, a job ad, a policy. Use it when a draft has hard requirements, such as a word limit, facts that must appear or phrases your company does not use. And use it on your own writing when you are too close to it to see the gaps, or before a meeting where someone senior will pick it apart.
The criteria are the work. Good ones are checkable: “states the ask in the first paragraph” can be judged pass or fail, while “is compelling” cannot. If you have already written down what a good result looks like, those sentences are your criteria.
When it does not
Without criteria, a critique is generic. Ask for feedback with no standard and you tend to get advice that fits any draft, such as adding an example or tightening the conclusion.
In the same chat, the model shares the assumptions it wrote the draft with. If it misread your brief the first time, it will check the draft against the same misreading. For anything important, paste the draft and the criteria into a new chat, or run the critique on a second model in Model Compare.
A critique cannot check facts it does not have. It can tell you a figure has no source, not whether the figure is right. Check claims against the source yourself, with the fact-check list to help.
It costs a pass. The model reads the whole draft again, so the tokens and the wait roughly double. Some newer models already check their own work, and Anthropic warns that verification instructions carried over from older prompts can make Claude Opus 5 over-verify, adding tokens and time. One round of critique and revision is usually enough; ask for a second only on what still fails.
Before and after
Three drafts from different kinds of work. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Make this email better. """ [draft email] """
Below is my draft reply to a customer whose order arrived two weeks late. Check it before I send it. Draft: """ [draft email] """ Check the draft against these four criteria, one at a time: 1. It says what went wrong in the first two sentences. 2. It gives a specific new delivery date or refund, not a promise to look into it. 3. It does not blame the courier or the customer. 4. It is under 150 words. Answer in a numbered list: for each criterion, say pass or fail and quote the sentence at fault. Then rewrite only the failing parts and leave the rest unchanged. If a criterion cannot be judged from the draft, say so instead of guessing.
“Make this better” returns a new email in the model’s voice, often longer and warmer than yours. The critique returns four verdicts, perhaps one fail for a vague “we’re looking into it”, and a fix to that sentence only.
Can you review my business case? """ [business case] """
Business case for a second warehouse shift: """ [business case] """ The finance director will read this business case and decide whether to fund it. Critique it as that reader would, against three tests: - Every cost and saving has a number and a source. - The risks section says what happens if order volume falls 20%. - The ask is stated in the first paragraph. Return a table with three columns: the test, pass or fail, and the paragraph that fails it. Then list the three weakest claims in the document, weakest first. If a test does not apply, say so. Do not rewrite the document, and leave style and grammar alone.
A general review comments on everything from tone to commas. This one returns a three-row table and the three claims the finance director is most likely to question, which is what you need to fix before the meeting.
Is this job ad OK? """ [job ad] """
Job ad for a warehouse team lead: """ [job ad] """ Review this job ad before it goes on our careers page. Experienced shift leads who are not actively looking should read it and want to apply. Check it against our hiring rules: - No requirement that is not needed on day one, such as a degree. - Pay range and shift pattern stated. - No phrases that put off older or younger applicants, such as "digital native" or "recent graduate". Give a numbered list of problems, each with the line it appears in and a suggested fix, most serious first, at most eight. If the ad passes a rule, say so in one line. Do not rewrite the ad.
“Is this OK?” usually gets a yes with a few polishing tips. Against named rules, the answer can point to the line that asks for a degree the job does not need, which is exactly the kind of problem that keeps good applicants away.
The line to add
Paste the draft first, then the criteria. Three to five criteria is plenty:
Draft:
"""
{{draft}}
"""
Check the draft above against these criteria, one at a time:
1. {{criterion}}
2. {{criterion}}
3. {{criterion}}
For each, say pass or fail and quote the sentence at fault. Then rewrite only the parts that failed and leave the rest unchanged. If a criterion cannot be judged from the draft, say so instead of guessing.
When the model is writing the draft itself, you can fold the check into the same prompt, as Anthropic suggests: end with “Before you finish, check your draft against {{criteria}} and fix anything that fails.” For important work, keep the check as a separate step in a fresh chat.
In ChatGPT, Claude, Gemini and Copilot
All four vendors describe some version of this. Anthropic calls self-correction the most common way to chain prompts: generate a draft, have Claude review it against criteria, then refine it, each as a separate step (Prompting best practices). The same guide suggests ending a prompt with a request to verify the answer against test criteria, and notes the exception above: Claude Opus 5 checks its own work well without being told, so remove those instructions on that model (thinking section).
OpenAI’s advice for its reasoning models is to be very specific about the end goal and to encourage the model to keep iterating until it matches your success criteria (Reasoning best practices). Google’s example template for Gemini includes a Validate step that reviews the output against the user’s task (Prompt design strategies).
None of that replaces your own read. Microsoft’s guide to Copilot prompts says to review and verify Copilot’s responses and cross-check them with trusted sources (Get started writing prompts). The critique tells you where to look first.
Key takeaways
- Ask for a verdict on named criteria before any rewrite.
- Write criteria that can be judged pass or fail.
- Ask for the line at fault, and fixes to failing parts only.
- For important work, run the critique in a fresh chat or on a second model.
- A critique finds unsupported claims; it cannot tell you whether a figure is right.
Questions
Can ChatGPT critique its own writing?
Yes, and it does it better with criteria. Ask it to check the draft against three to five specific tests, say pass or fail on each, and quote the line that fails, before it changes anything. In the same chat it shares the assumptions it wrote with, so for important work paste the draft into a new chat for the critique.
How do I get honest feedback from AI on my writing?
Give it the standard to judge against and the reader, and ask for the weakest parts first. Tell it not to rewrite, so the answer is a verdict rather than a new version. “Is this good?” tends to get a polite yes; “List the three claims the finance director is most likely to question” gets something you can act on. The critique template is a ready version.
Should I use a different AI model to check another model’s work?
For anything that matters, it helps: a second model, or the same one in a fresh chat, reads the draft without the first conversation’s assumptions. Give it the same criteria. Model Compare runs one prompt on several models side by side. Neither replaces checking the facts against the source yourself.
Prompts to use with this
- Brainstorm names and taglines against your criteriaYou need options for a product, event or campaign name, checked against your rules and audience.
- Check a draft's claims against your sourcesBefore a report, article or proposal goes out, you want every factual claim matched to a source and anything unsupported pulled out for checking.