Critique Prompt Design
What you'll learn: How to construct prompts that guide language models to systematically identify specific flaws, errors, or weaknesses in generated content.
The Core Idea
When you ask a model to critique its own output (or another model's output), you're essentially requesting a meta-analysis. But vague instructions like "check if this is good" rarely produce useful results. Critique prompt design is the art of crafting explicit, structured prompts that direct the model's attention toward particular dimensions of quality, making flaws detectable and actionable.
Think of it like asking a professional editor to review an essay. Instead of "Is this good?", you ask: "Check for factual errors, logical inconsistencies, and unclear phrasing." The specificity transforms a fuzzy judgment into targeted feedback.
Key Elements of Effective Critique Prompts
1. Explicit Criteria
Specify exactly what to evaluate: accuracy, completeness, relevance, tone, logical coherence, formatting compliance, or adherence to constraints.
2. Granular Instructions
Break down the critique task: "List any factual claims that lack supporting evidence" beats "Is this accurate?"
3. Structured Output Format
Request critiques in a consistent format (JSON, bullet points, severity ratings) for easy parsing and downstream use.
4. Examples of Flaws
When appropriate, include examples of the types of problems to look for, priming the model's pattern recognition.
Example critique prompt:
Analyze the following answer for:
1. Factual errors (cite specific claims)
2. Logical inconsistencies
3. Missing key information
Provide your critique as a JSON object with 'factual_errors', 'logical_issues', and 'missing_info' arrays.
This transforms self-critique from a subjective guess into a structured, repeatable quality check.
Key Takeaway: Effective critique prompts break evaluation into explicit, measurable dimensions and request structured output, enabling models to systematically identify specific flaws rather than offering vague judgments.