Self-Critique
Built-in quality check.
The Pattern
Section titled “The Pattern”After generating output, prompt the same or a follow-up call to critique the output against specific criteria, then revise based on findings.
When to Use
Section titled “When to Use”- Quality-sensitive outputs
- When you can articulate what “good” looks like
- As a lighter alternative to adversarial review
- When external verification is expensive
Implementation
Section titled “Implementation”Two-phase approach:
- Generate initial output
- “Review this output for [criteria]. List any issues. Then provide a revised version.”
Or single-pass:
- “After your response, critique it for [criteria] and revise if needed.”
Criteria Examples
Section titled “Criteria Examples”- Accuracy: “Are all claims verifiable?”
- Completeness: “Is anything missing?”
- Clarity: “Would this confuse the reader?”
- Tone: “Does this match the requested voice?”
Limitations
Section titled “Limitations”- Models have blind spots about their own errors
- Self-critique is weaker than external review
- Can add tokens without adding value on simple tasks
Related Patterns
Section titled “Related Patterns”- Adversarial Review: External skeptic instead of self-review
- Iterative Refinement: Multiple passes with different focuses
