self-reflection

• SKILL.mdcore · 6,429 bytes
• SKILL.mdcore · 6,429 bytes

## Overview

Self-reflection gives a language model a second pass over its own work: generate,
critique, revise. The critique can target correctness ("does this code pass the
tests?"), quality ("is this summary faithful?"), or strategy ("is this plan
actually going to work?"). Unlike gradient-based learning, reflection improves the
output within a single episode using only inference — the model reads its own
draft the way an editor reads a manuscript, spots problems, and fixes them.

The research lineage runs from simple "review your answer" prompts through
Reflexion — which stores reflections as episodic memory across trials, so the
agent learns from past failures without weight updates — to verbal reinforcement
methods where the model maintains textual "lessons learned." The pattern is most
powerful when the critique is grounded: test results, tool outputs, or retrieved