reflect

name	reflect
description	Extract reusable, evidence-supported improvements after a completed engineering session, repeated failure, or explicit retrospective request. Use when the goal is to improve future agent behavior or an existing workflow. Do not activate automatically after every task or create new rules from a single preference.

Reflect

Turn observed behavior into the smallest justified improvement.

Gather evidence

Review the objective, decisions, corrections, failures, verification results, and user feedback. Identify where the process saved effort, created rework, missed evidence, or required repeated instruction.

Classify the lesson

Distinguish a one-off circumstance, project convention, tool limitation, skill-routing problem, missing procedure, and enforceable mechanical rule. Put project-specific facts in project guidance. Prefer a script, lint, or check for deterministic constraints. Reserve skill changes for reusable judgment or workflow.

Test the proposal

State the observed failure and the new behavior that should result. Check whether an existing skill already owns it. Propose a narrow edit and a realistic scenario that would demonstrate improvement. Do not broaden a skill based on speculative edge cases.

Boundaries

Do not edit skills or project rules unless the request authorizes changes. Preserve unrelated existing instructions. Report when evidence is too weak to justify a change.

Output

Return the observations, lessons, destination for each lesson, proposed minimal change, and validation scenario.