Self-critique prompting means asking a model to review its own draft against a clear checklist, then to revise only the issues that are real. In Self-Refine: Iterative Refinement with Self-Feedback, Aman Madaan and colleagues described a simple loop that uses the same model as generator, critic and reviser: produce an output, generate feedback, then refine, without extra training. They reported that, across seven tasks ranging from dialogue to mathematical reasoning, outputs from this loop were preferred over ordinary one-step generation by humans and automatic metrics, with about a 20 percentage point average improvement in task performance in their evaluation. Those results are from the paper’s tasks, not from eBook chapters, and they are not a guarantee about your prose.
The paper also found that vague feedback is weaker than specific feedback. In one ablation, telling the model only that “something is wrong” reduced preference rates compared with pointed comments that named what to change. That is the practical lesson for writers. “Make this better” is not a critique. “The draft invents an industry average I did not supply, and the ending is missing the fill-in line” is a critique.
For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so a draft can be written in one stage and reviewed in another. That generate-then-review split is the same habit, without exposing the template’s internal prompts. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article. They are not paid Builder prompts, and they are not recorded model outputs.
This is close to getting AI to edit without rewriting everything, but the focus here is the self-review loop: the same model, or a later pass, has to find problems against a checklist before it is allowed to change the text. It also sits inside prompt chaining: draft, critique, revise, each as a request you can stop.
What should the critique pass be allowed to do?
A good critique pass has a job, a list, and a stop condition.
- It should score or comment only on the items you named.
- It should be allowed to say “no issue” on a line, so it is not forced to invent problems.
- It should not rewrite the draft in the same breath, unless you have already accepted the findings.
- It should not be asked to certify facts it cannot see in your notes.
If you want a fact check, supply the notes and ask for a claim-by-claim comparison. If you want a tone or structure check, say that, and do not pretend the same pass has verified the research.
Some people prefer to start the critique in a new chat and paste only the draft plus the checklist, so the model is less attached to its first wording. That can help, but it is not magic. The checklist and your review of the findings matter more than the chat layout.
When should you use a self-critique loop, and when should you skip it?
Use it when the first draft is close but unreliable in a way you can name: missing the workbook close, drifting off the outline, adding a section you forbade, or slipping into generic pep talk. Skip it when the draft is already on a rejected plan, or when the real problem is missing sources. In those cases, go back to the brief, or use a recovery process like the one in What to Do When AI Gets It Wrong.
One or two rounds is usually enough. Madaan et al. iterate, but a book chapter can start to sound over-processed if you keep asking for another pass with no new criteria.
Example 1: draft, then critique, then revise
Start with an ordinary drafting brief, then stop.
Write the section "Name the floor before the call" for new freelance
designers. 280 to 320 words, calm tone. Use only these notes. End
with a fill-in line for the reader's floor. The only figure you may
use is a $90 floor for a one-page logo revision.
Turn 2: critique only.
Review the draft below against this checklist. Do not rewrite it yet.
Checklist:
1. Stays inside the supplied notes
2. Does not invent averages, quotes or studies
3. Uses the $90 floor only as the given example
4. Ends with a fill-in line for the reader's floor
5. Stays within 280 to 320 words
6. Does not add a package menu or discount policy
For each item, answer "Pass," "Fail," or "Unclear," and quote the
phrase that justified the call. If nothing failed, say "No issues."
Turn 3: revise only the accepted fails.
Revise the draft to fix only items 1, 4 and 5, which I have accepted
as fails. Do not otherwise rewrite. Keep the sentences that already
passed. Return the revised section and a two-line list of what you
changed.
The “only items I accepted” line is the difference between a review and a second, unsupervised draft.
Example 2: ask the critic to look for a specific failure, not a vibe
When drafts go wrong, they often go wrong in the same way. Build that into the checklist.
This model sometimes invents a compromise when my notes conflict.
Here are the notes and the draft.
Notes:
- One note says $75. Another says $120. I have not chosen.
Draft:
[paste]
Critique only this question: Did the draft pick a number, average the
two numbers, or flag the conflict as unresolved? Quote the sentence
that proves it. Then stop. Do not fix the draft until I say so.
If the critique misses the invented compromise, do not trust a later “full polish” from the same model. Fix the brief or fix the sentence yourself.
Example 3: keep the editor from rewriting the voice
Self-critique can still overreach if the revision prompt is loose. Borrow the constraint from a limited edit.
You already wrote a critique. I agree with only these two points:
- The close is missing the fill-in line.
- The phrase "industry standard rate" is unsupported.
Revise the draft to fix those two points. Rules:
- Do not change paragraph order.
- Do not add examples.
- Do not replace my calm tone with stronger marketing language.
- Track changes in a short list after the draft.
If the model still rewrites everything, the self-critique loop has failed as an editor, even if the critique itself was accurate. At that point, insert the fill-in line yourself and delete the unsupported phrase by hand.
Which kind of review is this, and which kind is it not?
Keep the pass types separate, or the critique will try to do everything and do none of it well.
| Pass | Checklist should cover | Do not ask it to |
|---|---|---|
| Structure | Missing close, extra section, off-outline heading | Verify a statistic |
| Instruction-following | Length, forbidden topics, required fill-in line | Improve the “energy” of the voice |
| Voice | Tone rules you can point to, such as “no hype” | Invent a new metaphor unless you asked for one |
| Facts | Claims compared with notes you pasted | Certify the draft against the live web |
Madaan et al. evaluated Self-Refine on tasks with clearer targets than “write a good chapter.” For a book, you have to supply those targets. If you cannot write a checklist, you are not ready for a self-critique pass. Fix the brief first.
How does this fit Drafter and Polisher?
Drafter can produce the first pass from an approved outline. Polisher, or a later review pass, can run the checklist. The important part is the human gate between “here are the problems” and “now change the text.” In the Prompt Engineering Lab, you can test a single-variable change: the same draft, two checklists, one of them vague (“improve this”) and one specific. Predict which critique will stay on the job, then look at the revision.
Do not ask the critique pass to do SEO BOT’s job, or to invent citations. Keep the loop narrow.
FAQ
How do I get ChatGPT to critique its own writing?
Ask for a review against a checklist, and tell it not to rewrite yet. Require a pass/fail or “no issue” on each item, with a short quotation as evidence. Then, in a new message, allow a revision of only the findings you accept. You can paste the draft into a new chat for the review if you want a little more distance, but the checklist is what makes the pass useful.
Does self-critique prompting actually work?
It can, especially on structure, missing sections and instruction-following. Madaan et al. (2023) reported that their Self-Refine loop was preferred over one-step generation on the tasks they tested, and that specific feedback beat vague feedback. It is much weaker as a fact-checker unless you supply sources and ask for a claim-level comparison. You still have to read the critique and the revision.
Should I review the draft in a new chat?
A new chat can reduce the model’s tendency to defend its first wording, but it is optional. What you should not skip is the separation of critique and revision, the specific checklist, and your approval of the findings. If a new chat is inconvenient, use the same thread and say “do not rewrite yet.”
Before you hit send
- Do you have a checklist that names real failure modes, including “no issue”?
- Is the critique separate from the rewrite?
- Have you accepted or rejected each finding yourself?
- Are you asking the revision to change only those findings?
- Is this a voice or structure pass, or a fact pass that needs your notes?
A second pass is useful when it is a review. It is noise when it is just another chance to rephrase the same chapter.