Summary
Reflection prompting: ask the model to critique its own output and revise. Cheap, sometimes powerful, occasionally counterproductive. Works best when the critique step has fresh context and a specific rubric.
Objectives
- 01Structure a reflection prompt.
- 02State when reflection improves and when it degrades output.
- 03Distinguish reflection from self-consistency.
The Lesson
The pattern
1. Generate. 2. Critique against a specific rubric ('check for factual errors, unclear phrasing, missing citations'). 3. Revise. The critique must be specific — 'is this good?' produces meaningless self-approval.
Failure modes
Models often defend their own output rather than critique it. Fresh context (a separate call, or explicit rubric) helps. Compounding drafts can wander; cap at one or two reflection rounds.
Key Ideas
- Reflection needs a specific rubric to add value.
- Two rounds max; more compounds drift.