Find Skills

Discover published Agent Skills before writing one, and install them only after their provenance survives inspection. The npx skills CLI resolves a name against the public skills.sh registry, which is a different surface from the local jeo-skills catalog — mixing the two is the most common routing mistake here.

Third-party skills run with full agent permissions. Treat every search result as untrusted input until its source is verified: an install is a code-execution decision, not a bookmark.

When to use this skill

Check whether a published skill already solves the user's problem.
Search the registry by keyword, or scope a search to a GitHub owner.
Triage candidates on install count, source reputation, and repo stars.
Install, update, or remove a skill from the public ecosystem.
Decide between installing an existing skill and authoring a new one.

Do not use this skill for neighboring jobs:

Browse, filter, or install from the local jeo-skills catalog: use jeo-skill.
Rank and load an already-installed skill for a task: use openspace.
Author a new skill from scratch: use write-a-skill.
Validate a SKILL.md against the Agent Skills spec: use skill-standardization.
Improve a weak skill against a benchmark: use skill-autoresearch or upskill.

Instructions

Step 0: Confirm the user means the public ecosystem

find-skills searches skills.sh. jeo-skill searches this repository. If the user says "our catalog", "the jeo skills", or names a skill already in .agent-skills/, hand off rather than searching the public registry.

Step 1: Turn the request into a search

Identify the domain, the specific task, and whether it is common enough that a published skill plausibly exists. Then check the skills.sh leaderboard before running a CLI search — it ranks by total installs and usually surfaces the battle-tested option first.

npx skills find <query>
npx skills find <query> --owner <owner>

Use specific keywords (react testing beats testing) and try synonyms (deploy → deployment → ci-cd) before concluding nothing exists.

Step 2: Triage before recommending — never recommend from search rank alone

Signal	Green	Yellow	Red
Install count	1K+	100–1K	<100
Source	`vercel-labs`, `anthropics`, other official	known community author	unknown author
Repo stars	100+	20–100	<20
Recency	pushed recently	stale months	abandoned

A skill is executable instruction text that your agent will follow. Before recommending, open its SKILL.md and check that the described behavior matches its name, that it does not request credentials without cause, and that any bundled script is inspectable. Report a mismatch instead of installing.

Step 3: Present options, do not auto-install

Give the user: what the skill does, install count and source, the exact install command, and the skills.sh link. Then stop and let them choose.

Step 4: Install only on explicit confirmation

npx skills add <owner/repo@skill>              # project-level
npx skills add <owner/repo> --skill <name>     # explicit skill selection
npx skills add <owner/repo@skill> -g -y        # global, no prompts

Flag semantics that change blast radius:

-g / --global writes to the user-level skill directory, affecting every project. Confirm separately from the install itself.
-y / --yes skips confirmation prompts, including the CLI's own security summary. Do not pass it on a first install from an unfamiliar source.
--agent <runtime> targets one runtime instead of every detected one.
--copy vendors files instead of linking — the copy will not track upstream.

The CLI prints a security assessment before installing. Read it; do not suppress it with -y and then report the install as verified.

Step 5: Resolve names by frontmatter, not folder

The registry resolves --skill <name> against the name: in a skill's YAML frontmatter, which need not match its directory. A name absent from the repository tree can still install correctly. Verify against the installed file rather than concluding the command is wrong.

Step 6: Maintain and remove

npx skills update          # update installed skills
npx skills init <name>     # scaffold your own

Removal is a filesystem operation on the target skill directory. Know whether the install was project-level, global, or --copy before deleting.

Step 7: When nothing fits

Say so plainly, offer to do the task directly, and only then suggest authoring. Route authoring to write-a-skill; do not improvise a SKILL.md here.

Examples

Example 1: Ambiguous catalog

Request: "Find me a skill for React performance."

Ask nothing if context is clear, but pick the right surface: if the user is working inside jeo-skills, react-best-practices already exists locally — route to jeo-skill. Otherwise search skills.sh.

Example 2: Low-reputation result

A search returns a skill with 40 installs from an unknown author. Report the counts, state that it falls below the recommendation threshold, and offer either a higher-install alternative or doing the work directly. Do not install it because it was the top hit.

Example 3: Global install request

Request: "Just install it everywhere, skip the prompts."

Confirm once, explicitly, that -g -y writes user-level and bypasses the security summary — then run it. Do not treat an earlier "install it" as consent for both flags.

Best practices

Separate the two catalogs — skills.sh is public; jeo-skills is local.
Leaderboard before CLI search — it answers most queries in one step.
Never recommend on rank alone — installs, source, and stars first.
Read the skill before running it — it executes with full permissions.
Treat -g and -y as separate approvals — scope and prompt-bypass are different risks.
Verify by frontmatter name — a missing folder is not a missing skill.
Prefer an existing skill over a new one — author only after searching.

find-skills