Find Agent Skills Before You Install: Search by Task and Inspect the Source

Find agent skills helps developers produce a source-review brief before installation. Start with a specific task, search for candidates, open each repository, inspect SKILL.md plus every bundled file, and send that evidence to evaluation instead of trusting a popularity list.

Skill Summary

Vercel’s official Skills CLI includes a find command for interactive or keyword search, while skills.sh offers a public directory surface. Both are discovery tools: they help you locate candidates, but they do not decide whether a package is trustworthy, compatible, or useful for your project.

The durable unit to inspect is the source package. Vercel documents a skill as a directory centered on SKILL.md, with optional scripts, templates, documentation, or resources; those bundled files can materially change what the skill asks an agent to do.

Where This Skill Fits

This guide owns one narrow route: task wording → candidate search → repository inspection → evaluation handoff → installation only after acceptance. It stops before the deeper Agent Skills evaluation workflow, which is where a reviewer records evidence and makes an accept or reject decision.

If a package contains executable helpers, unfamiliar dependencies, network behavior, or a large bundle, choose an appropriate check from the Agent Skills security scanner guide. A clean search result or a familiar repository name is never a security verdict.

Who It Helps

• Claude Code and Codex users: need a task-relevant candidate without browsing an unbounded catalog.
• Developers: want to see the actual instructions and bundled files before allowing a package into a project.
• Team leads: need a repeatable pre-install handoff that separates discovery from approval.
• Reviewers: want a source URL, fixed revision, and inventory before assessing risk or fit.

Access / Install Links

• Official CLI repository: vercel-labs/skills.
• Fixed README: reviewed commit.
• Fixed find-skills instructions: SKILL.md at the reviewed commit.
• Vercel reference: Agent Skills documentation.

These links are for discovery and source review. When a candidate has passed evaluation, use the separate platform installation guide for placement, activation, updates, and removal.

Setup Steps

1. Write the task as a short capability phrase, such as “review pull requests” or “create changelogs.” Keep product names only when they materially narrow the job.
2. Search with the official CLI or browse skills.sh. Treat every hit as a candidate, not a recommendation.
3. Open the candidate’s repository rather than relying on a directory card. Record the repository owner, URL, last reviewed revision, and license if present.
4. Read SKILL.md frontmatter and instructions. Then inventory adjacent scripts, templates, reference files, assets, configuration, and dependency manifests.
5. Note what the skill may read, write, execute, or send; flag unstated assumptions and missing files. Pass that packet into the evaluation workflow.
6. Install only after the reviewer accepts the candidate for the intended environment and scope.

Example Inputs

Translate the actual job into a compact query. The official find-skills instructions use task phrases instead of asking for an abstract “good skill.”

npx skills find [task phrase]

npx skills find [task phrase] --owner [GitHub owner]

Expected Outputs

The useful output is not “a skill to install.” It is a review packet: task statement, candidate repository, fixed revision, SKILL.md, bundled-file inventory, observed capabilities, permissions or network questions, compatibility assumptions, and unresolved concerns.

A strong packet lets another person reproduce the source review without trusting the search surface. It also makes the later accept/reject decision auditable.

Use Cases

• Repository automation: search for a PR-review or changelog capability, then inspect scripts and GitHub assumptions before evaluation.
• Design handoff: locate a design-to-code skill, then verify its input formats, output paths, templates, and supported agent environment.
• Research workflows: find a source-gathering skill and identify whether bundled helpers make network requests or store results.
• Team standardization: give reviewers a fixed source packet instead of a directory screenshot or copied install command.

Limitations

• Search ranking and directory placement do not prove quality, maintenance, compatibility, or safety.
• A readable SKILL.md may refer to bundled scripts or resources that need separate inspection.
• A repository can change after review; record an immutable commit or release when possible.
• Source inspection is triage, not a substitute for evaluation, scanner selection, sandboxing, or human judgment.

Common Mistakes & Fixes

• Searching for “best skills”: rewrite the query around the exact task and required environment.
• Stopping at the directory card: open the repository and inspect the source package.
• Reading only SKILL.md: inventory scripts, templates, references, assets, and dependency files beside it.
• Installing to test whether it works: complete the review and choose an isolated validation route before touching a production environment.

Related Tools / Prompts

For a concrete source-backed collection, see the ljg-skills Codex skill collection; use it as an example of inspecting named packages, not as a universal recommendation. The Hand-Drawn Styles Skill guide shows how a repository can expose multiple reusable recipes whose files and scope should be reviewed before use.

FAQ

• Q: Where can I find Agent Skills?
  A: Use an official search command such as npx skills find or browse skills.sh, then open each candidate’s repository before deciding what to evaluate.
• Q: What should I inspect before installing a skill?
  A: Review the repository owner, fixed revision, SKILL.md, bundled scripts, templates, references, assets, dependencies, and any behavior that may read, write, execute, or send data.
• Q: Does appearing in a directory mean a skill is safe?
  A: No. Discovery produces candidates only; safety and suitability require source-specific evaluation and, when appropriate, scanning or isolated validation.
• Q: Should I install a skill just to see what it does?
  A: No. Inspect the source first, document unknowns, and validate accepted candidates away from production systems.

Found a useful skill after a source-first review? Share the task and the evidence that helped you accept or reject it.

Explore more reusable capabilities in AI Skills and keep platform installation in the dedicated guide.

Follow @bigprompt for more skills, workflows, and prompt systems.

Continue the workflow:

Evaluate a candidate Agent Skill and record an accept or reject decision

Install an accepted skill in Codex, Claude, Cursor, or Gemini

Choose a scanner when the source review needs additional checks

Inspect a named Codex skill collection by capability

Review the files and recipes in a visual style skill

Big Prompt Hub Review

The strongest discovery habit is deliberately boring: describe the job, search narrowly, open the source, inventory every bundled file, and hand evidence to a reviewer. That sequence is slower than copying the first install command, but it preserves the distinction between “found,” “understood,” “accepted,” and “installed”—the distinction a reliable Agent Skills workflow needs.