ablatelisted

# Maestro: Ablate Two phases. **Phase A** is a static audit: one reasoning pass per skill, no trial runs, covers everything. **Phase B** is an empirical loop: baseline, stub, measured add-back, real `claude -p` runs, expensive. **Run Phase A first, always.** Scripts do the mechanics, take `--help`, and are invoked by absolute path from this skill's directory. The judgement is yours: classification, task authoring, mapping failures to sections. ## Before anything writes `~/.claude/skills/` has no version control by default, and `apply.mjs` destructively rewrites `SKILL.md` files — it refuses to touch a skill whose own files are uncommitted. Git is the rollback path; the `.bak` is a convenience, not a safety net. ```sh cd ~/.claude/skills && git init && git add -A && git commit -m "chore: snapshot skills" ``` ## Cost — say this before starting Phase B Phase A costs one reasoning pass per skill. Phase B **spends real money against the user's own subscription, consumes their rate limits, and competes with the session they are in** — budget 30–60 nested runs per skill plus grading. Never begin one without saying so and getting a yes. ## Start with the inventory `node scripts/inventory.mjs --top 20` ranks targets by lines × recorded usage. - **Zero-usage skills are a roster-prune candidate, not an ablation target.** A description loads into every session whether the skill is invoked or not, so archiving an unused one helps every session; shrinking a skill nobody calls