find-skills: The Skill That Teaches Your Agent to Find Other Skills

The agent-skills ecosystem has a discovery problem. There are now thousands of installable skills — instruction sets that extend a coding agent with specialized knowledge — scattered across GitHub. The hard part is no longer writing a skill; it’s knowing one exists, finding the good one, and trusting it.

find-skills is a small, clever answer: a skill whose whole job is to help your agent discover and install other skills. Ask “is there a skill for X?” and instead of a shrug, the agent searches the open ecosystem, vets the results, and offers to install the right one. This post covers what it is, how it works on top of the skills CLI and the skills.sh registry, how to use it, and why “let the agent install its own capabilities” deserves a careful read.

What it is

find-skills is a single SKILL.md file from vercel-labs/skills (the home of the skills CLI, MIT-licensed). It’s a meta-skill: it doesn’t do design or testing itself — it teaches the agent the procedure for finding the skill that does.

It sits on top of two things:

The skills CLI (npx skills) — the package manager for the open agent-skills ecosystem, which installs skills into 50+ agents (Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Copilot, and many more).
skills.sh — the registry and leaderboard, ranking skills by total installs so the popular, battle-tested ones surface first.

Why it matters

Closes the discovery gap. Your agent stops saying “I can’t do that” and starts saying “there’s a skill for that — want it?”
Builds in a trust check. find-skills doesn’t just grab the first search hit; it bakes vetting into the procedure (install count, source reputation, repo stars) before recommending anything.
Portable by design. Skills follow the shared Agent Skills spec, so the same skill installs across dozens of agents — low lock-in.
Self-extending workflow. It turns capability gaps into a one-line install instead of a research project, and if nothing exists it offers to scaffold a new skill (npx skills init).

How it works

find-skills is just instructions the agent loads. When you ask “how do I do X,” “find a skill for X,” or “can you do X,” the skill activates and runs a defined playbook:

Identify the need — domain (React, testing, deploy…) and the specific task.
Check the leaderboard first — look at skills.sh, where skills are ranked by installs (e.g. vercel-labs/agent-skills and anthropics/skills, 100K+ installs each).
Search — if the leaderboard doesn’t cover it, run npx skills find [query].
Verify before recommending — prefer skills with 1K+ installs, trusted sources (vercel-labs, anthropics, microsoft), and a repo with real stars; be skeptical of anything under 100 installs or from an unknown author.
Present options — name, what it does, install count, the install command, and a skills.sh link.
Offer to install — and if nothing fits, do the task directly or scaffold a new skill.

That step 4 is the interesting part: discovery without vetting is how you end up trusting a random repo’s instructions, so the skill makes “check the install count and the source” part of the procedure.

Getting started

You can use the CLI directly, without any skill installed:

# search the ecosystem
npx skills find react performance

# install a specific skill, globally, to Claude Code
npx skills add vercel-labs/agent-skills -s react-best-practices -g -a claude-code

To get the conversational version — where your agent does the finding for you — install find-skills as a skill:

npx skills add vercel-labs/skills -s find-skills -g

After that, asking your agent “is there a skill for writing changelogs?” triggers the discovery playbook instead of a guess.

In practice

Fill a gap mid-task. You’re deep in a PR and want a consistent review pass: ask “is there a skill for PR reviews?” — the agent runs npx skills find pr review, checks installs and source, and proposes one to install.
Bootstrap a new project’s tooling. Starting a Next.js app, you ask for the relevant skills; find-skills points you at the high-install web-development skills from official sources, and installs the ones you pick to your agent of choice.

The shift is subtle but real: you stop maintaining a mental list of “good skills I should remember” and let discovery happen at the moment of need.

How it compares

Approach	What you get	Tradeoff
find-skills + skills.sh	the agent finds, vets, and installs skills for you, ranked by installs	the vetting is heuristic, not a guarantee
Hunting GitHub / awesome-lists	full visibility, browse everything	manual — you search, judge, and install yourself
MCP registries	runtime tools (servers) the agent can call	a different layer — tools vs instruction-set skills

Be fair to the alternatives: an awesome-list gives you the raw directory and total control, and MCP solves a different problem (live tools, not instructions). find-skills’ edge is that it lives inside the agent loop and turns “I wish it could do X” into an action, with a sanity check attached.

Tradeoffs

Popular ≠ correct, and ≠ safe. Install count and stars are proxies, not proof. A skill is a set of instructions your agent will follow, so a wrong or malicious one is a real risk — the vetting heuristics reduce it, they don’t remove it. Review what you install, especially anything that runs commands.
Auto-install is a supply-chain decision. Letting the agent install with -g -y means it’s adding third-party instructions to itself unattended. Convenient, but treat it like npm install from a stranger: fine for trusted sources, worth a look otherwise.
It only helps where skills are loaded. find-skills is itself a skill; it works in agents that support the spec and actually invoke it, and its usefulness depends on the registry’s coverage for your domain.
Telemetry on by default. The CLI collects anonymous usage data; disable it with DISABLE_TELEMETRY or DO_NOT_TRACK if that matters to you.

Takeaway

find-skills is the missing discovery layer for agent skills: a tiny meta-skill that makes your agent search the ecosystem, weigh the options by popularity and source, and install the right one on request — across 50+ agents, via the skills CLI and the skills.sh registry. Add it when you live in a coding agent and keep hitting “I wish it could do X.” Just keep the human judgment in the loop on the one thing the heuristics can’t fully cover: whether a popular skill is actually a good, safe one to let your agent follow.