feedback-learning

What problem does it solve? When users correct an AI agent's implementation or design choices, that knowledge is often lost or dumped into ever-growing instruction files. This Skill ensures each correction is classified and stored once in the smallest authoritative artifact, preventing duplicated or contradictory guidance. ## Core Features & Use Cases - Feedback Classification: Routes each correction to the right destination, such as a regression test, design document, instruction file, or learning inbox entry. - Learning Inbox Management: Maintains scoped .github/learnings/<area>.md files with strict size limits, compaction, and 90-day review cycles. - Deduplication Discipline: Searches existing rules before writing and removes inbox entries once promoted to authoritative documents. - Use Case: A reviewer rejects an agent's error-handling pattern in a pull request. The Skill determines the feedback is a concrete behavior that must not regress, encodes it as a regression test, and avoids adding a redundant rule to the instructions. ## Quick Start Record the user's correction about the retry logic in the API client using the feedback-learning workflow.