What problem does it solve? When users correct an agent's implementation or design approach, that knowledge is often lost or dumped into ever-growing instruction files. This Skill routes each correction to the smallest authoritative destination so reusable knowledge is preserved without turning skills and instructions into append-only knowledge dumps. ## Core Features & Use Cases - Feedback classification: Routes corrections to the right destination — regression tests for concrete behavior, design documents for architectural invariants, instruction files for universal rules, or .github/learnings/<area>.md inboxes for immature guidance. - Duplicate prevention: Searches existing instructions, skills, design documents, and tests before writing, updating existing rules instead of adding duplicates. - Inbox maintenance: Enforces a ten-topic, 8 KB limit per learning inbox with compaction, promotion, and garbage-collection rules including a 90-day review cycle. - Use Case: A user rejects an agent's error-handling pattern in a subsystem. The skill generalizes the principle, checks for an existing rule, and records it once in the owning design document or a scoped learning inbox entry with scope, evidence, and disposition fields. ## Quick Start When I correct your approach or reject a pattern, classify my feedback and record it in the appropriate destination following the feedback-learning workflow.
