In this article
Problem: First-pass LLM outputs frequently contain errors, gaps, or suboptimal structure. Without review, these defects reach users.
Solution: After generating an initial response, make a second LLM call instructed to critique the first output — identify errors, missing information, and improvements. Then make a third call to revise based on the critique. Optionally loop until a quality threshold is met.
Implementation:
- Generate: First LLM call produces initial response
- Critique: Second call with prompt: "Review this response for errors, gaps, and improvements: [response]"
- Revise: Third call: "Rewrite the response addressing these critiques: [critique]"
- Gate: Check revision quality with a short evaluator call; loop if below threshold (max 3 iterations)
Trade-Offs:
- ✔ Pro: Significant quality improvement on complex tasks (20-40% reduction in errors)
- ✔ Pro: Self-contained — no external data required
- ✖ Con: 2-3x cost and latency over single-pass generation
- ✖ Con: Risk of "sycophantic revision" — model agrees with its own critique without meaningful changes
When To Use: Code generation, long-form content, technical explanations, and any high-stakes output where quality trumps speed. When to avoid: Real-time chat, simple factual lookups, latency-sensitive applications.