Critique Loops
Separate generation, review, and revision so outputs improve against a rubric.
Learning Objectives
- Use a rubric to evaluate generated output.
- Separate draft creation from critique.
- Rewrite outputs based on specific failure points.
Separate generation from review
A critique loop asks the model to draft, inspect, and revise. This works better than one-shot prompting when quality depends on judgment.
The key is to use a rubric. A rubric can include specificity, evidence, tone, completeness, risk, and actionability. The model first evaluates the draft against those criteria, then revises only after the weaknesses are named.
A strong loop
- Generate or provide a draft.
- Score it against clear criteria.
- Identify the weakest points.
- Rewrite the draft using the critique.
This pattern is useful for customer emails, executive summaries, strategy memos, support replies, and product launch notes.
Rubric: specific, evidence-backed, concise, action-oriented.
Task: Score the draft, name the weakest points, and rewrite it.Examples
Rubric-led revision
Ask the model to score a customer email on specificity, evidence, tone, and next action before rewriting it.
Practice Exercise
Create a critique loop
Paste a draft from your workflow and build a rubric with four criteria. Ask the model to score and revise it.
- The first pass generates or reviews a draft.
- The critique names concrete weaknesses.
- The revision responds to the critique.
Mini Prompt Templates
Critique and Revise
Draft: [PASTE_DRAFT] Rubric: [QUALITY_CRITERIA] Task: Score the draft, identify the three weakest points, then rewrite it. Format: Scorecard followed by Revised Draft.