Self-Critique / Refine
Reasoning
Level 2 · Advanced reasoningGenerate a first answer, critique it against explicit criteria, then produce an improved version.
Self-critique prompting asks the model to draft an initial answer, then review that same answer against explicit criteria — factual accuracy, completeness, adherence to constraints, clarity — and only then produce a revised final answer that addresses what the critique found. Separating 'produce' from 'judge' tends to catch errors and omissions that don't surface when a model simply commits to its first answer.
When to use it
Use it for tasks where correctness or completeness genuinely matters and a first draft is likely to have gaps — factual writing, code, anything with explicit requirements that are easy to half-satisfy.
When to avoid it
Avoid it for simple, low-stakes tasks where a first answer is already reliable enough — the extra critique-and-revise pass adds tokens and latency without a meaningful quality gain.
Advantages
- Catches errors and omissions that don't surface in a single-pass answer
- Makes the criteria for 'good' explicit and checkable, rather than implicit
- Works well combined with almost any other technique — it's a pass you add at the end, not a replacement for anything
Disadvantages
- At least doubles the tokens and turns compared to a single answer
- The critique step can itself miss real problems or flag non-issues
- Doesn't guarantee the revision is actually better — occasionally a 'fix' introduces a new problem
Examples
Write a product description for a running shoe. Input: lightweight trail running shoe, waterproof, 280g, max 80 words. First, draft the description. Then critically review your own draft against these criteria: does it mention the weight? Does it respect the 80-word limit? Is the tone consistent with an athletic product? Then produce a revised, final version that addresses anything your review found.
Advanced variant — repeated refinement (4 iterations): draft an answer. Critique it against: {{CRITERIA}} Revise. Repeat the critique-and-revise cycle 3 more times, using the previous revision as the new draft each time. On the 4th and final pass, stop and present the result even if further improvement seems possible — diminishing returns set in well before iteration 4 for most tasks. (Running this as 4 separate, sequential prompts — pasting each revision's output back in as the next draft — is exactly what a Pipeline template does, docs/prompts/16.2-pipeline-editor.md.)Pairs well with
Try it
Fill in the placeholders below and see the compiled prompt update live.
What you want to achieve.
The material to work from.
How should the alternatives be compared? (e.g. cost, impact, time to ship)
- 29
- Words
- 215
- Characters
- 0/3
- Placeholders filled
- 1 min
- Reading time
- Completeness0
- Clarity0
- Context0
- Specificity0
- Constraints0
- Format0
- Output0
- Constraints note0
- Success criteria0
- Scope focus100
- Verification step0
- Negative example0
- Confidence0
- Placeholder syntax100
- Markdown structure0
- Fill in the remaining placeholders.
- Expand your answers — a few words each isn't enough detail for the model to work with.
- Add background context so the model understands the situation.