- Skills
- Find Skills
The problem: a booming ecosystem, no real method to sort it
A few numbers as of 2026-05-17:
- The
anthropics/skillsrepo has 136k stars and hosts about a dozen official skills (frontend-design, document skills, etc.). vercel-labs/agent-skillsships skills with 100k+ installs each (react-best-practices, web-design-guidelines).microsoft/azure-skillspublishes half a dozen skills at 250k+ installs each.- Add to that the community lists in
awesome-claude-skills,awesome-agent-skills, and so on.
Concrete fallout: a developer wondering "how do I e2e-test a Next.js app with Claude Code?" has to know which organization to look at, in which repo, compare candidates, check stars, and read every SKILL.md. That's exactly the work find-skills automates.
What it is (and what it isn't)
Four concepts overlap in this ecosystem. The rest of this page assumes you can tell them apart:
| Concept | What it is |
|---|---|
The find-skills skill | A SKILL.md file (~135 lines) living in vercel-labs/skills/skills/find-skills/, which teaches Claude (or any agent that supports the SKILL.md standard) how to use the Skills CLI thoughtfully |
The Skills CLI (npx skills …) | The npm package skills (v1.5.7 as of 2026-05-17), published by vercel-labs, acting as the package manager for skills: find, add, check, update, init |
| skills.sh | The public website, both a directory and a leaderboard, that aggregates published skills and ranks them by installs |
| The SKILL.md standard | The format convention (YAML frontmatter name + description, then Markdown body) that every skill follows so it can be loaded by Claude Code, Cursor, Gemini CLI, etc. |
How find-skills works
When Claude detects discovery intent (phrases like "find me a skill for…", "how do I do X?", "is there a skill that…"), it loads the SKILL.md and runs the following workflow:
Understand the need
Claude identifies the domain (React, testing, deployment…), the specific task (write tests, optimize perf, generate a changelog…), and estimates whether a skill likely exists for it.
Check the skills.sh leaderboard
Before running a search, Claude scans the skills.sh (ouvre un nouvel onglet) home. The top skills there are battle-tested and cover most common needs (web, design, Azure, e2e…).
Run `npx skills find <query>`
If the leaderboard doesn't cover the need, Claude runs the CLI search with targeted keywords. The SKILL.md recommends trying alternative terms ("deploy" → "deployment" → "ci-cd") rather than one vague query.
Verify quality before recommending
This is the core of the skill: no blind recommendations. Three criteria with explicit thresholds (see next section).
Present the options
For every candidate skill: its name, what it does, its install count, the copy-paste install command, and the skills.sh link for deeper inspection.
Offer to install
If you confirm, Claude runs npx skills add <owner/repo@skill> -g -y directly, and the skill is available within seconds.
Install find-skills (one command)
npx skills add vercel-labs/skills@find-skills -g -y
The -g flag installs the skill at the user level (available in all your projects). The -y flag skips the interactive confirmation.
To verify the install worked:
npx skills check
Once installed, there is no slash command to invoke. The skill triggers automatically when Claude detects the patterns documented in its SKILL.md (triggers like "find me a skill…", "how do I…", etc.).
Three concrete examples
Example 1: optimize a Next.js app
"I need help optimizing the perf of my Next.js app"
Expected output from Claude once find-skills is loaded:
I found a skill that should help. The "react-best-practices" skill providesthe Next.js performance guidelines from Vercel Engineering.(~405k installs as of 2026-05-17, official source vercel-labs)To install:npx skills add vercel-labs/agent-skills@react-best-practices -g -yView the skill: https://skills.sh/vercel-labs/agent-skills/react-best-practices
Example 2: review my React PRs
"do you know a skill for code-reviewing my PRs?"
Claude first runs npx skills find pr review, then filters results against the quality criteria and proposes the best one or two candidates, with their install count, source, and install command.
Example 3: a niche domain with no existing skill
"is there a skill to generate Mermaid diagrams from a SQL dump?"
Expected output when no credible match is found:
I searched for skills related to "mermaid sql diagram" but didn't find acredible match (only two candidates with under 100 installs).I can still help you with this task directly using my native capabilities.If you do this often, you could also initialize your own skill:npx skills init mermaid-sql-diagram
That is exactly the right behavior: refusing to recommend a low-quality skill just to fill the silence.
The quality criteria the skill enforces
The find-skills SKILL.md sets explicit thresholds before any recommendation. These are not soft hints, they are the rules Claude follows:
| Criterion | "Prefer" threshold | "Be cautious" threshold |
|---|---|---|
| Install count of the skill | ≥ 1,000 | < 100 |
| Source reputation | vercel-labs, anthropics, microsoft (trusted officials) | Unknown authors |
| GitHub stars of the host repo | "More is better" | < 100 |
Install count is the most discriminating signal. A skill with 2 installs, however nicely written, may hide a bug, a hostile dependency, or simply fail to deliver what it claims. The SKILL.md formalizes this caution so the agent doesn't pollute your environment out of overzealousness.
The seven categories to know
The SKILL.md lists seven common search categories. Knowing these labels helps you phrase more effective queries:
| Category | Keywords to try |
|---|---|
| Web Development | react, nextjs, typescript, css, tailwind |
| Testing | testing, jest, playwright, e2e |
| DevOps | deploy, docker, kubernetes, ci-cd |
| Documentation | docs, readme, changelog, api-docs |
| Code Quality | review, lint, refactor, best-practices |
| Design | ui, ux, design-system, accessibility |
| Productivity | workflow, automation, git |
Pair it with the rest of the Codex
Once find-skills is installed, you can chain it with the other skills pages on the Codex:
- What is a Skill? to understand the core concept and how a skill differs from a
CLAUDE.mdfile. - Top Skills for 2026 for the Codex's own editorial picks, independent from the skills.sh leaderboard.
- Skills vs Agents vs Plugins to situate a skill against the other extension mechanisms.
- Create your own skill when a search returns nothing and you need to package your own know-how.
Limits and caveats
Going further
- Publish your own skill:
npx skills init <name>scaffolds a SKILL.md ready to push to GitHub, then to list on skills.sh. - The official repo: github.com/vercel-labs/skills (ouvre un nouvel onglet) (18.9k stars as of 2026-05-17, MIT license,
skillspackage v1.5.7) to track CLI changes. - The directory: skills.sh (ouvre un nouvel onglet) to browse the catalog by installs, by topic, or by 24h trending.
- The Anthropic repo: github.com/anthropics/skills (ouvre un nouvel onglet) (136k stars, Apache 2.0) which hosts Anthropic's official skills and the reference specification of the SKILL.md format.
Ultimately, find-skills is less a tool than a discipline: forcing Claude to apply an explicit reading grid before suggesting that you install third-party code into your environment. That discipline becomes more valuable as the ecosystem grows, which is probably why it tops the leaderboard.