Find Skills

tl;dr: a meta-skill — instead of doing the task, it finds an existing skill that does. Lives in the same repo as the npx skills CLI (vercel-labs/skills). Triggers on intent phrases like “how do I do X” and “find a skill for X”, then walks skills.sh leaderboard → npx skills find → verify install count and source → present install command. Never installs without asking.

What it actually does

SKILL.md (from vercel-labs/skills, 2026-09-09) walks the agent through six steps when the user wants capability X and a skill might exist for it:

1. Understand the need. Identify the domain (React, testing, design, deployment) and the specific task.
2. Check the skills.sh leaderboard first. It ranks skills by total installs and surfaces the popular, battle-tested options — running find blind misses these.
3. Search. Run npx skills find <query> [--owner <owner>]. The skill explicitly says do not recommend a skill based on search results alone.
4. Verify quality. Install count (1K+ preferred, cautious under 100), source reputation (vercel-labs, anthropics, microsoft > unknown authors), GitHub stars on the source repo (treat <100 stars as suspect).
5. Present options. Skill name, what it does, install count and source, the install command, a skills.sh link.
6. Offer to install. With explicit confirmation, run npx skills add <owner/repo@skill> -g -y.

The skill is opinionated about not jumping to step 3. “Find a skill for React performance” → leaderboard first, because vercel-labs/agent-skills will already have react-best-practices ranked there.

What the CLI is

vercel-labs/skills is both the npx skills package manager and the home of this discovery skill. Key commands from the README:

Sources can be GitHub shorthand (vercel-labs/agent-skills), full URL, a tree path, GitLab, any git URL, or a local path. Private repos work with whatever credentials the host already has configured (Git credential helper, gh auth, SSH).

The CLI claims support for 75+ coding agents (Claude Code, OpenCode, Codex, Cursor, Windsurf, Zed, Aider, Cline, Copilot, Kiro CLI, etc.). Each agent maps to a per-agent skills directory; installation defaults to project scope, -g for global.

Usage contract

When the user signals a capability gap (rather than a concrete task):

1. Identify domain + task from the question.
2. Check the skills.sh leaderboard before running CLI search.
3. Run npx skills find [query] (scope with --owner <owner> if the org is known).
4. Verify install count ≥ 1K and source reputation before recommending.
5. Present skill name, install count, install command, skills.sh link.
6. Ask before installing. Run with npx skills add <owner/repo@skill> -g -y only on confirmation.

If no relevant skill exists, the skill explicitly tells the agent to say so and offer to do the task directly — or help the user create one with npx skills init my-xyz-skill.

Install

npx skills add vercel-labs/skills

Per the README, npx skills add accepts GitHub shorthand as the source format. The same repo also exposes the npx skills CLI itself, so installing the find-skills procedure does not separately require installing the CLI.

To install just the discovery skill without using the CLI installer:

git clone https://github.com/vercel-labs/skills
# Claude Code / Cursor pick up SKILL.md from:
#   ~/.claude/skills/find-skills/SKILL.md
#   .claude/skills/find-skills/SKILL.md

Requires Node and npx on PATH. The CLI has a --copy flag (independent copies per agent) for hosts where symlinks break; default is symlink.

Gotchas

• Meta-skill, not a task skill. Loading find-skills does not produce code or prose — it produces a recommendation. Agents that load it expecting a procedure to follow against user files will be confused. The trigger is the user asking for capability, not asking for help with files.
• Verify before recommending. The skill is explicit: “Do not recommend a skill based solely on search results.” Agents that paste the first npx skills find hit without checking install count and source reputation are not following the contract. 1K+ installs and a known publisher are the floor.
• Leaderboard first, CLI second. Skills that already rank high on skills.sh (e.g., vercel-labs/agent-skills, anthropics/skills) will appear at the top of find results anyway, but going to the leaderboard directly is faster and surfaces cross-repo install totals.
• CLI telemetry is on by default. The npx skills CLI collects anonymous install / discovery telemetry. Disable with DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1. GitHub repo and skill identifiers are sent only for repos GitHub positively confirms are public.
• Symlink vs copy. Default install is a symlink from each agent’s skills directory to a canonical copy — one source of truth, easy updates. On hosts where symlinks break (some sandboxed filesystems), pass --copy to get independent per-agent copies that drift.
• Project scope vs global scope. Default is project (.claude/skills/ etc.). -g makes it global (~/.claude/skills/). Project-scoped installs get committed with the repo and shared with teammates — fine for a team skill pack, wrong for personal preferences.
• Private repos need creds configured. HTTPS shorthand uses normal Git credentials → gh repo clone → SSH fallback. The CLI does not call gh auth token or copy the stored credential into the Node process. For tree lookups, GITHUB_TOKEN or GH_TOKEN can be set explicitly.
• sandbox_test.verdict: untested — we read SKILL.md and the README; we did not run npx skills find against a real query.

Related: Web Design Guidelines — what find-skills is most likely to recommend for UI audit tasks; Vercel React Best Practices — what find-skills is most likely to recommend for React perf tasks; Refactoring UI Audit — another high-signal skill this one will route to for design-review asks; UI Skills catalog — same shelf, broader routing index for UI jobs.