Critique-and-revise loop prompt builder
A model’s first answer is a draft, not a final. The critique-and-revise pattern — also known as self-refine or reflexion — makes the model grade its own work against a rubric and then fix it in the same conversation. This builder turns your task and a set of criteria into a multi-stage prompt: generate, critique, revise. You control the rubric, how many rounds run, and the final output format.
How it works
You enter the task and a rubric (one criterion per line). The builder writes a prompt with explicit stages. Stage one produces a draft answer. Stage two critiques that draft against each rubric line, scoring it and naming concrete weaknesses — not “looks good” but “claim X has no source; tone is too casual.” Stage three rewrites the draft to address every weakness raised. If you choose multiple rounds, the critique and revision stages repeat on the latest version. The final output is formatted as you specified. Everything is assembled locally; you copy the prompt and run it anywhere.
Why this works: separating generation from evaluation
A single-pass prompt asks the model to generate and evaluate simultaneously, which means both tasks compete for attention. Separating them into explicit stages lets the model bring full focus to each: generation produces the richest draft it can, critique brings evaluative attention without distraction from composition, and revision integrates the feedback. Research on self-refine techniques consistently shows that grounded critique — against concrete criteria — catches errors that a single pass misses.
Writing an effective rubric
The rubric is the most important input. Each line should be a binary checkable statement:
| Weak criterion | Strong criterion |
|---|---|
| ”Be accurate" | "Every statistic cites a specific source" |
| "Good tone" | "Tone is formal; no contractions; no colloquialisms" |
| "Clear structure" | "Output starts with a one-sentence summary, then uses numbered steps" |
| "Complete" | "All five requested sections are present” |
Vague criteria produce vague critiques. The model needs to be able to say “this fails criterion 3 because …” — that requires the criterion to be specific enough to fail or pass.
When to use multiple rounds
One critique-revise pass captures most of the quality gain for typical tasks like email drafting, code explanation, or summary writing. A second round helps when the first revision still has structured weaknesses — for example, a technical document where the first revision fixes factual errors but the tone still needs calibrating. Beyond two rounds the improvements tend to be marginal, and the risk of the model drifting from the original task grows. A clearer rubric usually beats adding a third round.
Practical notes
- Keep the original task in the rubric. If the revision stage can reinterpret the task, add an explicit criterion like “output must answer the original question, not a variation of it.”
- Specify the final format. A format requirement in the rubric (“output must be valid JSON”) lets the critique catch malformed output before it reaches your code.
- No API call is made. The builder assembles the multi-stage prompt text locally and you run it wherever you like — ChatGPT, Claude, the API. No key needed.