Ablate

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.

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 saves nothing.
• Vendor-maintained skills get re-pinned upstream, not ablated. Ablating forks them.

Phase A — static audit

node scripts/sections.mjs <skill>
node scripts/apply.mjs <skill> --decisions d.json --dry-run
node scripts/apply.mjs <skill> --decisions d.json
node scripts/report.mjs <skill>

Read references/static-rubric.md before deciding and references/classify.md for the gotcha inventory. Four rules, none negotiable:

• Extract before drop. A template or lookup table is misplaced, not dead: in prose it costs tokens every invocation, in a referenced file it costs nothing until read.
• Phase A never drops a gotcha. A gotcha may only be cut if a task built to trigger the failure it guards passes without it, which a static pass cannot establish.
• kind, prose ratio and section size order the search. They never justify a drop. The best content in most skills, verification steps and irreducible gotchas, is pure prose.
• Every drop names a rubric category. If you cannot name one, keep it.

Hand over the diff, the decision table with low-confidence drops on top, and the gotcha inventory.

Phase B — empirical loop

Follow references/protocol.md. The commands, in order:

node scripts/labinit.mjs <skill> --model <id>
node scripts/run.mjs <lab> --variant control --task <id> --reps 2
node scripts/stub.mjs <lab>
node scripts/run.mjs <lab> --variant stub --task <id> --reps 2
node scripts/run.mjs <lab> --gate                     # BLOCKING
node scripts/stub.mjs <lab> --keep <ids> --label <l>
node scripts/run.mjs <lab> --variant iter --iter <l> --task <id> --reps 2
node scripts/run.mjs <lab> --grade --a <dir> --b <dir>

• run.mjs is the only way to start a trial or a grading call. Improvised setup drifts, and drift invalidates every comparison in the lab.
• --gate is blocking; exit 2 means halt. A stub that passes everything means the harness is broken, not that the skill body is worthless.
• A split result across reps is a keep. Two reps disagreeing means you do not know.
• Delete the lab when done: <lab>/config/ holds a copy of the user's credentials.

Validate the rig first per fixtures/README.md. Do not self-ablate this skill.

The two gates

Grading in between is autonomous. Gate 1, before the runs: operator reviews bar.md, tasks.json, the gotcha inventory — a bad bar wastes every run after it. Gate 2, after the rebuild: rebuilt versus control per task, dropped and extracted with iteration evidence, thin-margin drops flagged for individual veto. Then spot-check the rebuild in real work: the clean room cannot see the context pressure that motivates the method.

References

references/: static-rubric.md Phase A · classify.md deterministic-vs-taste and gotchas · protocol.md the loop · bar-template.md success criteria. Also SPIKE.md (why isolation works) and fixtures/README.md (rig validation).