Agent Skills Directory

Feature Comparison

Keyword Evidence

agent skills directory is a P1 discovery page for the Claude Skills cluster. Latest backend data provided for this page shows US Volume 40, CPC $10.98, Global Volume 90, keyword ideas 11, and KD unavailable.

Directory Map

Different tools use different names for reusable agent behavior. Treat skills as a broad category that includes Claude skills, Codex instructions, Cursor rules, OpenClaw skills, shell skills, design engineer skills, DESIGN.md files, command packs, and MCP-backed workflows.

• Use official docs to confirm file locations and supported capabilities.
• Use awesome lists for discovery, not as proof of safety.
• Use marketplaces or registries only after checking publisher, version, and permissions.
• Use local skill managers when teams need ownership, review, and versioning.

Current Ecosystem Categories

Use this directory to classify skill assets before installation. The category matters because a Markdown-only rule file, a shell-capable skill, and a marketplace package have very different risk profiles.

• Claude Code: SKILL.md folders, hooks, slash commands, MCP configs, and CLAUDE.md references.
• Cursor: .cursor/rules files, editor-scoped guidance, MCP setup, and design-system rules close to code.
• Codex: AGENTS.md, task templates, repo review checklists, and validation commands.
• OpenClaw-style agents: marketplace or community skill folders that may touch local files, accounts, browsers, or shell commands.
• Shell skills: repeatable terminal workflows that need explicit command allowlists and rollback notes.
• Design engineer skills: DESIGN.md, token rules, component manifests, screenshots, accessibility checks, and product UI review habits.

Selection Criteria

A good agent skill should reduce repeated prompting without hiding risky behavior. Evaluate it by workflow value, compatibility, ownership, and blast radius.

• Does it solve a repeated workflow rather than a vague productivity wish?
• Can the same idea work in Claude Code, Codex, Cursor, or another tool?
• Does it name required files, commands, and validation steps?
• Does it avoid real keys, private endpoints, and destructive defaults?
• Can your team review and update it when the tool changes?

Installation Risk

Do not install third-party skills as if they were packages from a trusted internal registry. Read the files first, then test in a low-risk project.

Safe review path:
1. Read SKILL.md, rules, commands, scripts, and config files.
2. Search for tokens, env vars, curl commands, deploy commands, and delete commands.
3. Remove private examples and broad permissions.
4. Test in a disposable repo or branch.
5. Document owner, source URL, and last reviewed date.

Metadata Checklist For Each Entry

Every directory entry should be reviewable without opening five tabs. Capture the minimum operational metadata before recommending a skill to a team.

[ ] Source URL and publisher
[ ] Official, community, or internal status
[ ] Install path or copy procedure
[ ] Last reviewed date and upstream update date
[ ] License or usage terms
[ ] Supported tools and required dependencies
[ ] Commands, network calls, and file permissions
[ ] Sandbox, test repo, or rollback instructions

Awesome Lists, Marketplaces, And Skill Managers

Third-party directories are useful for pattern discovery, but they are not official endorsements unless the product owner says so. Label the source type clearly.

• Official docs: safest source for supported setup behavior.
• Official registries or marketplaces: useful, but still review permissions and publisher identity.
• Awesome lists: useful for examples and naming patterns.
• Team skill managers: production-friendly when ownership and review can be enforced.

Security Notes

Agent skills often sit close to file access, shell commands, network calls, and private business context. Security review should happen before a skill becomes part of a repo template.

• Never copy examples containing real account IDs, tokens, customer names, or private URLs.
• Prefer read-only tools and narrowly scoped commands.
• Separate public templates from internal policy references.
• Track source URL and last reviewed date for every imported skill.
• Require human approval for deploy, purchase, delete, reset, or credential operations.

Official And Community Starting Points

Start with official documentation where available, then use community lists only for discovery and inspiration.

FAQ