find-skills helps when you want to know whether an existing Agent Skill can support a task. It uses the Skills CLI and public directories to discover candidates, then presents their purposes, sources, and installation paths. Finding a package is different from installing it, and similar names do not establish equivalent capabilities.
Try it once, or install it for the project
To see whether this skill finds useful candidates, first make sure Claude Code runs on your machine. Then start a temporary session from your project directory:
npx skills@latest use vercel-labs/skills@find-skills --agent claude-code
skills use puts the skill material in a temporary directory and launches an interactive Claude Code session; it does not persistently install the skill in the project. Enter the sample task below in that session. If you decide to keep it for the project, run:
npx skills@latest add vercel-labs/skills --skill find-skills --agent claude-code
The installation command adds find-skills for Claude Code, not the packages it might later discover. Without --global, it uses project scope and still asks for installer confirmation. Use the following task in either the temporary session or a new session after installation:
Use find-skills to find an accessibility-review Skill for a React project.
I use Claude Code and want to inspect existing pages before changing code.
Compare two or three candidates by author, skill entry point, required tools, and output.
Read the original documentation before recommending. Do not install candidates yet.
Search for the actual work
Identify the domain, action, and desired output. “React accessibility audit” is more useful here than “design.” The CLI supports a query such as npx skills find react accessibility, with --owner when you need to narrow the source. If a search returns nothing useful, try synonyms or nearby task descriptions before concluding that no relevant skill exists.
Directories and rankings are discovery aids. Open the candidate's repository to confirm its SKILL.md name, directory, author, trigger conditions, and bundled resources. A repository can contain many skills, so its name may differ from the name selected during installation. Distinguish advice-only reviews from automatic repair workflows and from tools requiring browsers or external services.
Make the comparison actionable
For each candidate, explain which part of the task it covers, what it requires, and what it delivers. Installation counts and stars can suggest where to look, but do not replace reading the instructions. An unresolved requirement should remain explicit instead of being hidden behind a popularity claim.
After choosing, install the exact entry. The general form is npx skills add owner/repo --skill skill-name --agent claude-code; replace both placeholders with verified values. Use --list to inspect the skills in a repository before installation, and npx skills list to review installed skills. Global scope and skipped confirmations are separate options, not prerequisites for a first trial.
If nothing suitable exists, the agent can still help with the task directly, or a repeated workflow can become a custom skill. The useful result is a short comparison with concrete reasons and a clear next action, rather than a long collection of unread links.
Inspect one candidate before installing
Suppose the goal is accessibility review in an existing React project. After search, open the candidate repository’s SKILL.md: does it only provide review advice, edit code, or require a browser and external service? Check whether your agent can discover it, whether dependencies must be installed, and whether bundled scripts are part of the workflow. If two results share a name, distinguish them by author, repository and skill directory rather than display title.
To inspect a repository without installing, use npx skills add owner/repo --list; to actually try find-skills, use the temporary skills use route above. Install the exact item with --skill once you decide to keep it. Project scope is easy for a team to review; consider global scope when several projects need it. After installation, start a new agent session and try a small scoped task to confirm discovery and behavior.
What a recommendation should contain
For each candidate, state the task match, maintainer and original instructions, install command, prerequisites and unresolved questions in a few lines. If a result has only a slogan, unreadable instructions or unclear permissions, mark it as unverified. Install counts are discovery signals, not a quality score. The useful output helps you decide which one to try, what to prepare and when to choose something else.
Choose a candidate worth installing
Do not install a whole result list immediately. Open one or two candidates, read their triggers, dependencies, and full instructions, and check whether they fit the task at hand. One user reports finding two regularly used skills through this finder, while warning that install counts show popularity rather than quality.
Verify the installation record
If you install from a git@github.com: repository URL, do not assume a visible skill folder means future updates can track it. A user reproduced a missing .agents.skill-lock.json entry for find-skills with Skills CLI 1.3.8; an HTTPS source URL recorded the entry. Prefer the HTTPS install command, then inspect both the installed file and lockfile. Check your current CLI behavior if its version differs.