Reflection and Self-Critique

Reflection

Part of: Advanced Prompting Patterns

A model writes a function, you ask "are there any bugs in what you just wrote?", and it finds the off-by-one error it created seconds ago. Nothing about the model changed. You simply gave it a second pass, a chance to read its own work with fresh eyes and fix it. That second pass is reflection , and it is one of the cheapest quality upgrades you can bolt onto any prompt. What it is Reflection (or self-critique ) is a pattern where the model produces a first answer, then critiques that answer against some standard, and finally revises it. Generation and evaluation become separate steps instead of being crammed into one shot. The model wears two hats in sequence: first the author , then the critic , then the author again with the critique in hand. It works for the same reason chain of thought works. A single forward pass that must be correct on the first try is fragile. Splitting it into draft, critique, and revise gives the model explicit room to catch what the draft missed, much like a writer who always edits before publishing. How it works The loop is generate, critique, revise, and optionally repeat: The critique step is doing the real work. A vague "make it better" yields little

Challenge: The Self-Critique Pass