retrolisted

# retro A lesson that stays in the transcript is lost when the session ends; the compounding move is writing it back into the surface the *next* session loads. This skill closes that loop deliberately — after the work, not during it. Adapter: `.claude/ai-dev-kit.config.json` (`docs` block); a missing field → derive it from the repo and say so. ## 1. Collect the signals Sweep the just-finished work for the four signal types — from the session itself, the diff, and any review feedback: - **Corrections** — places the user redirected you, or you redirected yourself after wasted work. - **Surprises** — behavior that contradicted an assumption (an API, a build quirk, an environment fact) and cost time to discover. - **Re-derivations** — knowledge rebuilt from scratch that some session had already learned (the strongest memory-candidate signal). - **Repetitions** — any manual step performed 3+ times this program (automation candidate). ## 2. Route each lesson to its cheapest durable home One lesson, one home — pick the *lightest* surface that reaches the next session that needs it: - **Project memory** — non-derivable project facts and recipes (ports, seeds, gotchas). Cheapest; no repo change. - **Instruction file** (`AGENTS.md`/`CLAUDE.md`) — only rules every future session must hold; respect the budget (the context-economy skill polices it). - **Adapter field** — a mechanical per-project parameter a skill should read instead of re-deriving. - **Skill edit**