distill-skilllisted

# Distill Skill Turn **project memory** into curated, auto-loading **repo skills**. Durable patterns and non-obvious gotchas become a `SKILL.md` the agent loads on matching future tasks. **Topic** = skill name or free text from the user. Empty topic = **batch** (every dense cluster). Named topic = **single** skill. ## Hard principle Memory is raw. Skills are curated. - Never dump memory verbatim. - Keep: repeated procedures, stable file/command facts, multi-hit gotchas, lasting decisions. - Drop: fixed one-offs, dead ends, timestamps, PR numbers, superseded claims. - **Verify every load-bearing claim against live files.** Unverified → hedge, not assertion. ## Procedure ### 1. Memory surface (fail closed) Detect first available backend and scope to **this repo**. Details: [references/memory-backends.md](references/memory-backends.md). Order: mem0 (or similar) → claude-mem / local store → notes/transcripts the user provides. No hits → stop; say what you looked for. Never invent memories. ### 2. Cluster → candidates Search deploy, build, core domain, recurring bugs. Good target = multiple memories across sessions about a **repeatable procedure with non-obvious traps**. Each candidate: kebab-case name, one-line scope, rough count. ### 2a. Dedup and label Inventory project skills (`.agents/skills/*/SKILL.md`, `.claude/skills/*/SKILL.md`). User-global skill dirs are **read-only**. | Label | Meaning | |-------|---------| | **NEW** | Create | | **UPDATE** | Compact re