Prompt to Self-Critique and Improve Its Own Answer
An HR consultant was reviewing a ninety-day probation policy for a 180-person software company, and the draft had a problem she could feel but not name. Section 4.2 let a manager extend probation once without telling anyone. Section 6.1 required HR sign-off for any extension within fourteen days. The two rules could not both be true, and somewhere in the gap a manager would quietly keep someone on probation for a second quarter with no paperwork, no appeal route, and a criteria list that leaned on the phrase “cultural fit.” She had a version of the policy that read smoothly and one that was actually sound, and they were not the same document.
The prompt below makes ChatGPT criticize its own work before you accept it: it writes a revision, reviews that revision as a skeptical reader, names its own weak points, and rewrites again with the fixes marked. Built for that probation policy, it works on any document where the first answer looks fine and hides real problems.
The Draft and Its Hidden Risk
The probation policy governs how the company confirms or parts ways with new hires in their first ninety days. Its stated goal is a fair, consistent process, but its actual behavior depends on dozens of small choices: the scoring scale, who signs off, what happens to benefits if probation is extended, how a disagreement is appealed, and whether the manager has to give feedback before a final decision. A policy fails not on the big idea but on those details, and a model asked for “an improved version” will happily polish the prose while leaving a contradiction intact, because smoothing the surface is easier than auditing the logic.
Self-critique is the fix. You ask the model to judge its own output against explicit criteria, and the act of committing to a score forces it to look for the cracks it would otherwise paper over with better sentences.
The Copy-Paste Self-Critique Prompt
Paste the block into ChatGPT with the draft at the end. It runs in four stages and refuses to let a clean-but-weak answer pass.
You are an HR policy editor helping me strengthen a document I will paste below.
Document to improve: {{DOCUMENT}}
Company context: {{CONTEXT}}
Stated goal of the policy: {{GOAL}}
Hard constraints: {{CONSTRAINTS}}
Work in four stages and label each stage.
Stage 1. Produce your best revision of the document. Do not critique yet.
Stage 2. Now review THAT revision as a skeptical reviewer who did not write it. Do not defend it. Score it 1 to 10 on each criterion and justify each score in one sentence:
- Clarity for a first-time manager
- Internal consistency, meaning no rule contradicts another
- Fairness and legal exposure, meaning subjective or ambiguous criteria
- Actionability, meaning a manager can actually follow each step
- Completeness, meaning missing steps, edge cases, and an appeal route
Stage 3. Quote the three weakest passages from your Stage 1 revision and say why each is weak. A critique with no real findings is a failed critique.
Stage 4. Rewrite the document, fixing every weakness you named. Mark each materially changed section with the tag [CHANGED] and add one line explaining the fix.
Rules:
1. In Stage 2, be harsh. Do not give a high score to satisfy me.
2. Do not invent legal requirements or cite laws you are unsure of. Flag legal questions as "escalate to counsel" instead.
3. Write in plain business English. Define any term a manager might misread.
4. Where the draft is silent on an edge case, say so in a note rather than quietly deciding it.
The Stage 2 rule against defending the revision is the load-bearing one. A model asked to “double-check” its own answer tends to agree with itself. Forcing a numeric score against named criteria, and demanding three quoted weaknesses, makes agreement expensive and turns the review into real work.
Fields to Set Before You Run It
| Field | What to enter | Example here |
|---|---|---|
| {{DOCUMENT}} | The full draft, pasted in | 14-page probation review policy, sections 1 to 8 |
| {{CONTEXT}} | Company size and situation | 180-person software firm, first HR policy refresh in 4 years |
| {{GOAL}} | What the policy must achieve | a fair, consistent, documented 90-day process |
| {{CONSTRAINTS}} | Rules the model cannot break | no new headcount, keep the 90-day window, no legal claims |
Fill {{CONSTRAINTS}} carefully, because a model left unchecked will redesign the org chart to fix a wording problem. Telling it to keep the ninety-day window and the existing headcount keeps the rewrite inside what the company can actually ship. If you know the specific faults, such as the contradictory extension rules, put them in {{GOAL}} or a note, and the critique will find them faster.
Reading the Critique and the Revision
Expect four labeled stages, with the value concentrated in Stage 2 and Stage 3. A trimmed slice of the expected shape:
Stage 2 - self-review scores
Clarity for a first-time manager: 7/10. The steps are ordered, but "satisfactory progress" is never defined.
Internal consistency: 4/10. Section 4.2 allows a single extension by manager discretion; Section 6.1 requires HR sign-off within 14 days. These conflict.
Fairness and legal exposure: 5/10. "Cultural fit" is subjective and hard to defend if challenged.
Actionability: 6/10. Managers have no template for the written feedback.
Completeness: 4/10. No appeal route and no rule for benefits during an extension.
Stage 3 - weakest passages
"extension is at the manager's discretion" - leaves the decision unaccountable and untimed.
"cultural fit" - a criterion with no observable definition.
[third quote and reason]
Stage 4 - rewrite notes
Section 4.2 [CHANGED] - extension now requires HR sign-off within 5 business days and one written feedback session beforehand.
Section 2.3 [CHANGED] - "culture fit" replaced with three observable behaviours.
New Section 7 [ADDED] - appeal window of 10 days and the benefits rule during extension.
Judge the output by the critique, not the rewrite. If Stage 2 gives everything an eight or nine and Stage 3 names no real weakness, the model is flattering itself, and you should re-run with “Assume the document is flawed. Find the flaws.” The rewrite is only as good as the faults the review was willing to admit.
Variations: Adversarial, Rubric, and Looped
- Adversarial reviewer. Add “Argue the strongest case that this policy would fail a tribunal or an employee complaint, then fix that case.” This pushes the model past cosmetic edits into real risk.
- Rubric mode. Replace the five criteria with your own scoring rubric so the critique grades against standards your company already uses.
- Loop it twice. Run Stages 2 to 4 a second time on the revised document and stop only when a stage reports no new weaknesses. Two passes usually surface what one pass misses.
- Side-by-side. Ask for the original and final version of each changed section printed next to each other, so a reviewer can see exactly what moved.
- Multi-persona. Ask for three separate reviews, one as a new manager, one as an employee’s representative, and one as a compliance officer, then reconcile them.
Save the Stage 3 weaknesses as a checklist and confirm each one is actually fixed in Stage 4, because a rewrite that misses a fault it just named is a common failure. To build stronger review instructions, see the chain-of-thought prompt. To give the reviewer a clearer role to play, read the role-prompting guide, and to control exactly how the stages are formatted, use the output format guide.
Where Self-Critique Goes Wrong
Three traps are worth naming. First, a model that only flatters itself produces a confident, useless review, so insist on harsh scoring and quoted faults. Second, a critique can drift into inventing law; without the escalate-to-counsel rule it may cite statutes that do not exist. Third, self-review cannot catch everything, since the model shares its own blind spots across both passes, so a policy about people still needs a human HR reviewer and, where the stakes are legal, a lawyer. The technique narrows the gap between a smooth draft and a sound one. It does not close it on its own.
Sources & Notes
This prompt is written for ChatGPT (GPT-4o and later) and works in Claude and Gemini. The behavior it relies on, that asking a model to score and critique its own output surfaces issues a plain request misses, is a widely observed prompting pattern rather than a formal guarantee, so treat the critique as a first pass. No external data or legal sources were fetched, and the company, policy, and section numbers here are illustrative. Nothing in this article is legal advice: a probation or performance policy touches employment law, which varies by jurisdiction, so have any final version reviewed by qualified counsel before it is used with real employees.