retro-agent-instructions

To stop recurring AI agent errors, you need structured instruction refinement that formalizes feedback from session friction into durable project rules. This skill analyzes session logs to identify repeated corrections and generates atomic updates for instruction files like CLAUDE.md to prevent future occurrences.

The best way to update project instructions is through systematic evaluation of session friction signals against defined criteria. This skill identifies contradictions and missing context, then proposes evidence-based, concise improvements to formalize agent behavior rules and close instruction gaps.

To add permanent rules to CLAUDE.md or AGENTS.md, analyze recent session logs for friction signals like repeated corrections. This skill generates concise, atomic proposals for structural improvements, embedding the formatting rule directly into your instruction files to apply across all future sessions.

Yes, you can use session logs to generate documentation updates for AI workflows. This skill systematically evaluates session logs against friction criteria to identify instruction gaps, producing actionable and atomic documentation updates that refine agent behavior over time without manual rule writing.

You should refine agent instructions when you observe recurring errors, repeated corrections, or instruction contradictions during a session. If friction signals indicate missing context that affects long-term workflow optimization, formalizing these improvements prevents the agent from repeating the same mistakes in future interactions.

Yes, this approach works for both project-specific and user-scope instruction files. The skill targets the iterative refinement of instruction files at either scope, systematically evaluating session logs to generate atomic updates that prevent recurring errors across your specific configuration.