Self-Critique Prompting

Ask the model to check its own output for errors before accepting it, as a lightweight reflection step inside a single generation or agent turn.

Self-critique (self-eval) prompting asks the model to verify its answer, nudging critical thinking the way chain of thought nudges stepwise reasoning. It is a prompt-only form of reflection and error correction — useful when a full multi-module Reflexion setup is unnecessary.

Like CoT, it costs tokens and latency. Use it when silent wrong answers are costly and when you can still afford the extra generation; do not treat self-declared correctness as proof the goal was met (see agent planning failure modes).

---

Sources

• AI Engineering (Chip Huyen), ch. 5