| name | skill-finder |
| argument-hint | [capability or task] |
| description | Finds existing agent skills before creating a new one. Searches a curated list of 1184+ official skills from Anthropic, Vercel, Stripe, Supabase, Cloudflare, Sentry, MongoDB, Figma, and 40+ other teams. Use when the user asks "find me a skill for X", "is there an existing skill for Y", "how do I do Z", or expresses a capability gap — even if they do not explicitly mention "skill". Run BEFORE skill-creator to avoid duplicating work.
|
| license | Apache-2.0 |
Skill Finder
Discover existing official skills before creating a new one.
Quick start
- Capture intent — domain (PDF, auth, PostgreSQL) + task (extract, validate, migrate)
- Search
references/official-skills.md using grep-style keyword lookup
- Rank matches by publisher reputation and relevance
- Present 3-5 candidates with owner/name + one-line description + link
- Fallback to skills.sh /
npx skills find / skill-creator if nothing fits
Workflow
Step 1: Understand intent
Extract two dimensions from the user's request:
- Domain — the subject area (e.g.,
PDF, auth, postgres, stripe, react)
- Task — the action verb (e.g.,
extract, validate, migrate, test, deploy)
If the user is vague ("help me with auth stuff"), ask one clarifying question
before searching. A narrower query returns better matches.
Step 2: Search the local list
Grep references/official-skills.md with the domain keyword first, then
refine with task keywords if too many hits.
grep -i "pdf" .agents/skills/skill-finder/references/official-skills.md
grep -i "auth\|oauth\|jwt" .agents/skills/skill-finder/references/official-skills.md
Expand queries for common synonyms:
database → also match postgres, sql, mongo, mysql
auth → also match oauth, jwt, session, login
UI → also match frontend, react, design, component
Step 3: Rank matches
When multiple candidates surface, prefer publishers in this order:
- Official publishers —
anthropics/, vercel-labs/, stripe/, supabase/,
microsoft/, google-gemini/, cloudflare/, openai/
- Well-known teams —
huggingface/, sentry/, mongodb/, auth0/,
figma/, notion/, datadog/
- Community — any remaining matches
Within a single publisher, prefer the skill whose name most directly matches
the user's task verb.
Step 4: Present candidates
Format each recommendation as:
owner/skill-name — one-line description
https://officialskills.sh/owner/skills/skill-name
Present 3-5 candidates. Lead with the single best match; list alternatives
under a "See also" line if useful.
Step 5: Fallback when nothing matches
If no skill fits the query:
- Point to skills.sh — browse the leaderboard by
install count
- Suggest
npx skills find <query> — Vercel Labs' CLI search
- As a last resort, invoke
skill-creator to build one from scratch
Never invent a skill name. If it's not in references/official-skills.md
and not on skills.sh, the user needs to create one.
Gotchas
references/official-skills.md is a static snapshot — regenerate it when
the upstream source changes significantly. The header records the last
update date.
- Trust the publisher prefix.
anthropics/ ≠ anthropic-community/ — the
former is the official Anthropic repo, the latter is community-maintained.
- The list contains both skills and reference implementations. If a
match looks like a template or tutorial rather than a working skill,
flag that to the user.
- Do not recommend a skill based solely on name match. Verify the
description aligns with the user's actual task.
Refresh
references/official-skills.md is a static snapshot. To refresh, replace
its content with an updated curated list from a source of your choice,
then update the header (date, skill count).