Getting AI to Really Critique Your Text

The treasurer of a hiking club has written the statement he'll read at the annual general meeting to announce that dues are going from 45 to 55 euros. He pastes his text into an AI and asks: "What do you think?" Answer: "Excellent text, clear and well structured. Here are a few minor suggestions for improvement." He feels reassured. On the day of the meeting, the room only remembers the increase, because the text apologized for three paragraphs before giving the figure.

The AI didn't lie to please him. It did what a language model does when the question is open-ended: it took the most probable tone for someone showing their work, the tone of encouragement. Our article on AI bias calls this sycophancy. To get a real critique, you have to change the question.

Give it a reader and a goal

"What do you think?" doesn't say by what criterion to judge. A text is never good or bad in itself: it's good if it does what you expect of it, with those who read it. So write down both.

• The goal: what the reader should understand, feel, or do after reading. Here, accept the increase while knowing the reasons.
• The reader: who they are, what they already know, what state of mind they arrive in. Here, forty members, some of whom are already grumbling, listening to the text read aloud.

You can go further and ask the AI to read in place of a specific reader, as in the article on roles, but a skeptical reader is more useful than a friendly one. Your friends and family forgive a lot, whereas the skeptics are the ones who will vote against.

Ban compliments and demand quotes

Two instructions change almost everything. First, remove the flattering part, which teaches you nothing. Second, force the AI to quote the passage it's targeting, because a critique without a quote stays vague and can't be checked.

You are a member of my hiking club, who came to the annual general meeting with doubts about the dues increase. You're listening to this text read aloud. You're not mean, but you don't go easy on me.

My goal: for the members to accept the increase from €45 to €55 while understanding what the money will be used for.

Text:
"""
[PASTE YOUR TEXT]
"""

Give me the three problems most likely to make me fail, from most serious to least serious. For each one:
- quote the exact passage concerned;
- explain in one sentence what you understand or feel at that point;
- say what I could change, in one line, without rewriting the paragraph.

No compliments, no introduction. If a passage works very well, say so in a single line at the end, only if it needs to be protected during the edits.

The result will be harder to read than an "excellent text," and that's the point. The treasurer would have received something like: "The figure doesn't arrive until the fourth paragraph, I get the feeling you're justifying yourself before saying it." He could have fixed that in two minutes.

Limit the number of problems. With no cap, the AI finds fifteen, several of them unimportant, and you no longer know where to start. Three is enough for a first pass. You'll rerun it on the corrected version.

Separate critique from rewriting

If you ask "critique and then improve," you get a brand-new text whose voice you no longer recognize, and you no longer know what moved. Do the critique first, fix yourself what seems right to you, and only then, if you want help, use a second prompt to rewrite without distorting. You stay the author of every decision, and that's what makes the final text sound like you.

Another case: testing a text that has to convince someone else

A baker is writing a job posting for an apprentice. She doesn't want a style critique, she wants to know whether a sixteen-year-old would want to apply. The prompt changes target:

You're 16, you're looking for an apprenticeship, and you're reading this posting on your phone between classes.

[PASTE THE POSTING]

Answer these four questions, quoting the posting when possible:
1. What did you understand about the job, concretely?
2. What makes you want to apply, and what makes you hesitate?
3. Which words or abbreviations do you not understand?
4. At what point do you stop reading?

The prompt no longer asks for a judgment but for a reading. That's often more reliable: the AI describes what it understands, and you compare with what you meant to say.

What this critique can't see

The AI reads your text, nothing else. It doesn't know that the president once promised never to raise the dues, or that your club had a difficult year. It doesn't check your figures. It can also invent a flaw when asked for too much: "find what's wrong" calls for problems, even if there are almost none. Keep your judgment, set aside the remarks that don't speak to you and, if you disagree, ask it to defend the passage before deciding. If you'd rather have one of your ideas challenged than a text, you can also have it play devil's advocate.

Finally, think about what you paste. A text that contains member names or internal figures goes through an outside service: reread what you're better off keeping to yourself before sending it.

To try this week

Take a text you have to send or read soon, and write the goal and the reader in two sentences. Run the critique prompt, fix what seems right to you, then rerun it on the new version in the same conversation. If the AI finds the same problems again, they were real. If it finds others, you have a second layer to deal with, and you stop when you no longer agree with it.