skill-concierge

Manage skill-concierge's DISABLE list — skills the user ordered off, regardless of origin (personal, plugin, external catalog, or command-files surfaced as skills). Use this skill when the user wants to view, disable, block, or re-enable a skill, says "disable X", "block X", "stop using/offering X", "X is turned off", "re-enable/unblock X", "which skills are disabled", or asks why a Skill invocation was denied by the blocklist guard. Runs scripts/blocklist.py (list / add / remove); the blocklist lives under the canonical durable home (~/.claude/skill-concierge/blocklist.json) and is enforced by the enforcer (no offers/hints/routes), the engine (no search hits, get_skill refuses), and the PreToolUse guard (invocation denied) — ADR-0046.

Manage external skill-catalog roots — third-party skill collections indexed for retrieval WITHOUT being installed (ADR-0031). Use this skill when the user wants to register, list, or remove an external skills directory ("add my cloned skills repo to search", "index this skill collection without installing it", "which external catalogs are configured"), promote an external skill into ~/.claude/skills, or asks why a search hit is marked "[external]". Runs scripts/catalogs.py (list / add / remove / promote), which edits the operator-owned ~/.claude/skill-concierge/catalog-roots.json; the index follows at the next reindex.

Consult the best skill or chain of skills for a task — a deliberated, expert-style curation over the full catalogue (installed AND external), built on deep reading of skill bodies and capsule dossiers (ADR-0049). Trigger when the user asks to consult which skills fit a task, wants the best combo or chain of skills planned as a step BEFORE work starts, asks "which skills should I use for X" or "plan a skill strategy for this task", or wants the agent to think deeply and curate from the whole shelf instead of taking the quick per-turn offer. Runs a wide sieve (consult_candidates over the engine index), delegates body-level fit analysis to an analyst subagent, then composes a ranked RUN/⚠/ALSO verdict with promote-ready external picks. For the quick single-skill find mid-task, use skill-search instead.

Diagnose and repair a broken or degraded skill-concierge install. Use this skill when the skill-search MCP won't connect, search_skills returns nothing or stale results, skills have gone dark, the MCP seems to run old code after a plugin update, or anything about skill-concierge misbehaves after setup or a plugin update. Runs scripts/doctor.py to check the deployment layer (engine venv, engine freshness, Qdrant, MCP wiring, settings overrides, ledger) and delegates retrieval health to the engine; with --fix it applies safe repairs (start Qdrant, reindex, re-apply overrides).

See the retrieval-flywheel status and trigger an incremental utterance-generation run. Use this skill when the user asks about the flywheel, "how many skills have utterances / triggers", "flywheel coverage", "which skills are missing utterances", "is the LLM endpoint configured/reachable", or wants to "generate triggers", "run the flywheel", "refresh utterances", or "index the new skills' utterances". The flywheel is the utterance layer (ADR-0026) that teaches the retriever how users actually ask for a skill (EN+VN), lifting recall. Runs scripts/flywheel.py — status mode (default, read-only) prints endpoint config + reachability and per-skill utterance coverage (N/M covered, and the missing skills by name); --generate runs the incremental generator (only new/changed skills hit the LLM) then reindexes so the new points go live, printing before/after coverage. Generation fails loud if the LLM endpoint is unreachable.

Manage skill-concierge's always-ON allowlist — the skills kept fully described in every turn instead of name-only. Use this skill when the user wants to view, add, or remove always-on skills, asks "which skills are always on", "add X to always-on", "remove X from the keep-on list", "manage the always-on skills", or wants to curate what stays injected vs retrieved on demand. Runs scripts/keep-on.py (list / add / remove), which edits the always-on allowlist under the canonical durable home (~/.claude/skill-concierge/keep-on.json) and re-applies the settings.json overrides.

Bootstrap or repair the skill-concierge engine from scratch. Use this skill when installing skill-concierge on a new machine, right after a plugin update, or when skill-concierge:doctor reports the engine venv is missing. Runs setup.sh to build the stable engine venv, start the Qdrant container, build the multilingual index, and apply the curated skill-budget overrides, then verifies the result with doctor.

Find the right skills for a task before acting. Use at the start of any multi-step or unfamiliar request to retrieve relevant skills by meaning, not name. Triggers when the user asks to build, set up, design, deploy, fix, or automate something and the right skill isn't obvious.

Use when measuring whether a skill-concierge gate-threshold change helped real skill usage or adoption — "did the new gate values help", "skill usage impact after the change", "assess skill helpfulness", "audit skill usage post-deploy", "which telemetry is valid for skill-usage analysis". Stops the reflex of using the skill-invocation-ledger or treating offer→take as usage; routes to the transcript SKILL-FIRST trail instead.