Self-Critique Prompts: AI as Its Own Reviewer

Why self-critique works

Generating text and evaluating text are slightly different cognitive operations. When a model is asked to critique an answer (especially without being told it's its own answer), it can spot flaws it missed when generating. This is the foundation behind Constitutional AI and many production pipelines.

Single-prompt self-critique

Task: Write a short explanation of why inflation happens, aimed at a 9th-grade audience. Step 1: Draft your answer in <draft> tags. Step 2: In <critique> tags, evaluate the draft: - Is it accurate? Flag any oversimplifications. - Is it level-appropriate? Point out any jargon. - Is it complete? Name one thing missing. Step 3: In <final> tags, rewrite the draft fixing the issues raised. Only the <final> block will be shown to the student.A single-prompt self-critique — all three stages in one call.

This works because Claude will genuinely evaluate its own draft when asked. The <final> block is almost always better than the <draft>, often markedly so. The cost is extra tokens (1.5-2x) for significantly higher quality.

Separate-turn self-critique

TURN 1: Write a resignation letter. (no mention of critique) TURN 2: You are a no-nonsense HR consultant. The letter below was written by an employee. Identify the five biggest problems and rewrite it. <letter> {RESPONSE_FROM_TURN_1} </letter>The AI has no idea it's critiquing itself. This sometimes produces harsher, better critique than self-aware critique.

Self-consistency via sampling

A related technique: generate N answers at temperature > 0, then use the model (or majority vote) to pick the best. On reasoning tasks, 5 samples with majority vote on the final answer dramatically outperforms a single sample.

# Pseudocode answers = [] for i in range(5): answer = model.complete(prompt, temperature=0.7) answers.append(extract_final_answer(answer)) best = majority_vote(answers) # or ask a judge modelSelf-consistency with majority voting.

Limits of self-critique

• If the model is wrong in a systematic way, its critique will be wrong too.
• Self-critique can INCREASE verbosity without improving substance — watch for that.
• On factual accuracy, self-critique helps less than retrieval. The model can't fact-check what it doesn't know.
• Sycophancy: models sometimes praise their own drafts. Neutral critique framings help.