What problem does it solve? Teams running an AI-assisted engine accumulate failures, unserved user intents, and feedback that get lost or never converted into actual system improvements. This Skill provides a memory-backed improvement loop with an append-only lesson store, a demand ledger, and a human-gated promotion pipeline so recurring observations reliably become system changes. ## Core Features & Use Cases - Capture mode: Append a lesson (L-entry) or query-log row (Q-entry) at the moment a failure or unserved intent is noticed, with deduplication by owning surface and failure pattern. - Scan mode: Run one periodic retro over the capture buffer, feedback/ items, open lessons, and recent run reports, emitting PROMOTE flags and writing retro-latest.md. - Promote mode: Turn an at-threshold lesson into a minimal diff via draft, shadow-sandbox validation, fresh-context assessor grading, and a per-proposal human gate before any live surface is edited. - Status mode: Report store counts, PROMOTE flags, and last-scan date via store_check.py. - Use Case: After noticing the same documentation drift failure three times, run capture to log the lesson, then promote to draft and gate a minimal fix to the owning doc surface. ## Quick Start Ask the assistant to run a self-improve scan to review the capture buffer, feedback items, and open lessons and write the retro report.
