Dashboard

How to Prompt AI to Critique Your Writing Honestly

Stop getting flattery. A prompt pattern that makes AI critique your writing with quoted evidence, ranked problems, and no rewrite until you ask for one.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
8 October 20261 min read

To get an honest critique from AI, remove the invitation to be nice. Tell it to act as a skeptical editor, ask for the three biggest problems ranked by damage, require a quoted line as evidence for each, and forbid rewrites until you ask. Because models still lean toward agreement, add a calibration step: paste in a paragraph you know is weak and check that the model says so. A critique that only agrees with you is a known failure mode, and what sycophancy in AI is explains why it happens.

Why AI critique defaults to praise

Ask what the model thinks of your draft and you will usually get a compliment, three mild suggestions, and an offer to polish it. That is not laziness. Assistants are trained on human ratings, and people tend to rate agreeable answers higher. Anthropic researchers found that human preference data and the models trained on it sometimes favor responses that match a user's views over correct ones, which they describe in Towards Understanding Sycophancy in Language Models.

For feedback, that bias works against you. The more your message shows that you are proud of the piece, the softer the critique gets. If you want to see the same effect outside of writing, the post on how to prompt AI to disagree with you instead of agreeing covers it. The training process behind the bias is explained in the entry on the reward model.

The three-pass prompt

One request for feedback blends structure, argument, and sentence-level edits, and the model picks whichever is easiest, which is usually the sentences. Split them into three passes and run each in its own message.

text
You are a skeptical editor who reads for a living and is not my friend.
Below is a draft for [audience], meant to make them [do or believe X].
Pass 1, structure only. Ignore wording and typos.
- State the main point of the draft in one sentence. If you cannot, say so.
- List the three biggest structural problems, ranked by how much damage each does to a reader who has never heard of me.
- For each problem, quote the exact line or section that shows it.
- Do not rewrite anything. Do not praise anything.

Once pass 1 is addressed, run the second.

text
Pass 2, argument. List every claim a skeptical reader would not accept without evidence.
Quote each claim, say what evidence is missing, and label it unsupported, overstated, or fine.

And only then the third.

text
Pass 3, sentences. Quote the five weakest sentences and name the fault in each:
vague, redundant, passive, jargon, or buried point. Then stop.

Why the quote requirement matters

Requiring a quoted line is the most useful single constraint. A critique that must point at the text cannot be generic. If the quoted phrase is not in your draft, the model invented the problem, and you can see that in two seconds by searching for it.

It also changes the quality of the advice. A comment such as the intro is slow is easy to produce and hard to act on. A comment that quotes your second paragraph and says the point arrives in paragraph five gives you something to move.

Calibrate before you trust it

Take a paragraph you know is bad, one that rambles and buries its point, and run pass 1 on it. If the model calls it clear and well organized, the framing is not strong enough, or that model is too agreeable for this job. Tighten the persona, ban praise outright, and test again.

Two more tricks help. Present the draft as written by someone else, such as a junior colleague or a competitor, because people and models are blunter about work that is not yours. And never tell the model how long you worked on it or how much you like it. Models differ in how far they bend, so repeat the calibration whenever you switch tools.

What to do with the critique

  1. Fix structure first. Sentence edits on a paragraph you are about to cut are wasted effort.

  2. Decide for yourself which points to accept. The model is a reader, not a judge, and a reader can be wrong.

  3. Rewrite in your own words. If you let the model rewrite, you inherit its voice, and the fix is a separate job described in the guide to prompting AI to match your brand voice.

  4. Run pass 1 again on the revised draft and compare the problem lists. If the same problem comes back, the fix did not work.

The pattern is one case of a wider habit, covered in the prompt engineering guide: say what role the model plays, what it must not do, and how you will check its answer.

FAQ

Can AI replace a human editor?

No, but it is a useful first reader. It catches structural gaps, unsupported claims, and repetitive sentences quickly. It does not know your audience's real objections, and it cannot tell you whether the piece will land with the person you most need to persuade.

Which AI model gives the most honest feedback?

It varies by model and changes with updates, so do not rely on a ranking. Run the calibration test with your own weak paragraph on each tool you use and keep the one that names the problem plainly.

Should I paste my whole draft or one section at a time?

For long pieces, work section by section after a first structural pass on the whole. A model reading a very long draft gives thinner, more generic comments than one reading a few pages.

Is it safe to paste a confidential draft?

Check your tool's data settings before pasting unpublished, client, or personal work. If you cannot confirm how the text is stored or used, remove names and figures first.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.