What problem does it solve? Owner corrections — the moments a user says "no, that's wrong" — are the highest-value learning signals in an agent system, yet they survive only when someone remembers to write them down. This Skill scans session transcripts so those corrections are captured systematically instead of being lost when the session ends. ## Core Features & Use Cases - Transcript scanning: Parses Claude Code .jsonl transcripts and classifies user turns with a two-tier signal system (strong signals like "u didnt" or "thats not" stand alone; weak signals like "not" or "instead" require a second match). - Machine-turn filtering: Strips compaction summaries, injected skill bodies, tool results, and system reminders that arrive as user turns but were never typed by the owner. - Candidate reporting: Outputs scored correction candidates with the assistant text they corrected, leaving the final judgment to a human or the learner agent, which appends confirmed entries to memory/corrections.md. - Use Case: At the end of a build milestone, run the scanner over the session transcript to surface every moment the owner pushed back, then record the real ones so the next session does not repeat the mistake. ## Quick Start Ask the agent to run the correction-capture scanner on the latest session transcript and review the candidates before appending confirmed corrections to memory/corrections.md.
