which-skill

Intent match, not keyword match.

Recommend-only · quote-backed · live-catalog.

Ranks already installed skills and stops at the recommendation. Never silently load or run another skill.

Not find-skills (ecosystem install). Not a hardcoded domain→skill table.

Demos: examples.md · Harness: references/paraphrase-cases.md

Steps

1. Gate — negative space

Done when: you exited with a clear reason, or you proceed knowing a local skill may exist.

2. Extract intent

One sentence: job (outcome) · object (artifact) · constraints.

Pick altitude:

Done when: a restatement without the user's wording still picks the same family and altitude.

3. Load the live catalog

Single source of truth = this session's descriptions. No memorized tables.

1. Prefer the in-context skill catalog (name + description).
2. If missing / thin / stale: scan skill roots and read each SKILL.md frontmatter only. Paths: references/scan-roots.md. Follow symlinks; skip broken; optionally report reachable / broken.
3. Skip which-skill, skill-router, and non-actionable meta docs.

Done when: you have {name, description} for every reachable install — not a memory guess.

4. Semantic rank (quote-backed)

Score against the intent sentence (not surface keywords):

• Paraphrase test: phrasing would miss the skill's triggers, intent still matches → still a hit.
• Quote rule: every keep binds intent to a verbatim phrase from that skill's description. No "sounds related."

Return top 2–3. Combos count toward the cap. Specific beats generic; atomic beats flow when one skill suffices.

Done when: each candidate has intent ↔ quote; altitude is decided.

5. Present (recommend-only)

Reply in the user's language. Do not open or run a recommended skill until they confirm.

**{name}** — {one-line what it does}
- Quote: "{verbatim from description}"
- Why: {intent ↔ that quote}
- Altitude: atomic | combo step N | flow
- Invoke: {how to attach — wait for OK}

Weak but paraphrase-solid → still recommend, then step 6.

Done when: 2–3 picks with quotes, or an explicit no-match / no-skill-needed path. No catalog dump. No autoload.

6. Description feedback (weak / empty only)

When the best match is weak, or paraphrase would work but the description is too narrow:

Description feedback for `{name}`:
- Missed paraphrase: "{user phrasing}"
- Suggested trigger branch: "{one Use when… clause}"

At most one feedback block. For skill authors — optional for the user.

No local fit → offer bare help → find-skills only if they may want an uninstalled skill.

Done when: feedback emitted for a real gap, or skipped because quotes were strong.

Hard rules

1. Recommend-only — never autoload another skill
2. Quote-backed — no quote → no recommend
3. Live catalog — never hardcode domain→skill maps
4. Never invent a skill name absent from the loaded catalog
5. Prefer specific over generic; atomic over flow when one skill suffices

Paraphrase harness

When the user says "run paraphrase cases" / self-check / after editing this skill: load references/paraphrase-cases.md and score every case against the current catalog.

Done when: each case is hit | miss | skipped-not-installed + name + quote used.