extract

/extract — propose a skill, evaluate it, stage it

Goal: capture a repeatable workflow as a candidate skill a human can review and promote. This closes the loop the Stop detector opens — it never auto-creates a skill (the v0.1 staging rule). The output is a proposal, not a live skill.

Opt-in. No default-pipeline skill or hook invokes /extract. Running it is always explicit and changes no baseline behavior. (v0.3 AC-E4; preserves the staging rule.)

When to run

• A Stop candidate landed in .claude/skills-staging/candidates.md (the detector found a tool/skill subsequence that recurred this session), or
• you just did something repeatable/non-obvious worth keeping.

Do this

1. Read the evidence. Open .claude/skills-staging/candidates.md for the detected sequence + recurrence count. Cross-check what actually fired in .claude/traces/<date>.jsonl (skills, subagents, MCPs). Skim the relevant transcript span to recover the purpose — the sequence names the steps, not why.
2. Draft the skill. Write a candidate SKILL.md with valid frontmatter (name, description) and numbered steps that capture the workflow. Pick a name that does not collide with an existing skills/*/SKILL.md.
   • Write it ONLY to .claude/skills-staging/<slug>/SKILL.md. Never create or edit anything under skills/ (or any live skill dir).
3. Evaluate it (deterministic gate):

   node scripts/eval/extract-rubric.js .claude/skills-staging/<slug>/SKILL.md
   

   It scores five criteria and exits 0 (all PASS, none manual — not reachable until the AC-E3 benchmark lands) · 1 (a check FAILED) · 2 (no FAIL, MANUAL criteria remain).
   • R1–R3 are deterministic (frontmatter present · ≥2 steps · name not a duplicate).
   • R4 (genuinely reusable?) is always MANUAL — the script never auto-approves a skill.
   • R5 (empirical value) is result-driven: if /benchmark produced .claude/eval/benchmarks/<name>.json for this draft, R5 reports PASS/FAIL from it; otherwise it degrades to MANUAL. Run /benchmark --component <draft-name> to close it.
4. On FAIL (exit 1): fix the draft (add frontmatter, add steps, or rename to avoid the collision) and re-run. Do not stage a draft that FAILs.
5. Adjudicate MANUAL (exit 2 — the expected happy path): judge R4/R5 yourself with evidence — is this a workflow you'd actually reach for again, distinct from existing skills? Record the verdict + reasoning in .claude/skills-staging/<slug>/eval.md alongside the draft.
6. Stage for approval. Stop there. Tell the user the proposal is staged at .claude/skills-staging/<slug>/ (draft + eval note) and that promoting it into skills/ is their explicit call. Do not promote it yourself.

Security & trust

• extract-rubric.js runs no shell and never executes the draft's body — an ```eval fence inside a draft is inert text to it, so there is no check output to leak.
• --list is the safe inspect path: node scripts/eval/extract-rubric.js --list <draft> prints the parsed name/description + which criteria are deterministic vs. MANUAL.
• The Stop detector and this skill record only tool/skill names (from the trace, which stores no args) — never session content, commands, or secrets.

Notes

• Self-contained: rubric + detector are pure Node, no companions (mgrep/graph/context7).
• The AC-E3 benchmark (/benchmark, fork + worktree, pass@k/pass^k) feeds the R5 seam in extract-rubric.js: it reads the benchmark artifact if present, else R5 stays MANUAL and never blocks.

Exit criterion

A proposal (SKILL.md draft + eval.md) is staged under .claude/skills-staging/<slug>/, the rubric shows no FAIL, and every MANUAL criterion is adjudicated with evidence. Promotion into skills/ remains a human decision — never automatic.