What problem does it solve? Agent sessions repeatedly hit the same failed commands, wrong tool arguments, and user corrections, but those lessons disappear when the session ends. This Skill reviews session transcripts and history to convert mistakes into durable, reviewed instructions and new skills that future sessions will actually load. ## Core Features & Use Cases - Quick mode: Scans the current session transcript for failed-then-fixed commands, user corrections, and environment surprises, then proposes additions to AGENTS.md or CLAUDE.md with citations. - Full mode (/learn --full): Analyzes 30 days of session data, clusters recurring task patterns, proposes new skills for your skills repo, and flags friction in existing skills. - Review-first workflow: Every change is presented as a diff or a PROPOSED_SKILLS.md document and applied only after explicit user approval; nothing is written silently. - Use Case: After a session where the agent repeatedly fumbled shell quoting on macOS, run /learn to get a proposed global instruction entry covering BSD vs GNU flag differences, cited to the exact transcript moment. ## Quick Start Ask the agent to run /learn to review this session and propose instruction file additions for anything that went wrong.
