Find Skill

Help the user discover installable agent skills that match a task. Default to a quick, scannable list of candidates with summaries — no installation, no long explanations, no padding.

When to use this

Phrasings that should activate this skill:

• "Find me a skill for X" / "Is there a skill that does Y?"
• "What skills exist for [domain]?"
• "I need a skill to handle [task]"
• "Show me skills for [tool/format]"
• Any moment the user describes a recurring or specialized task and is wondering whether a pre-packaged skill exists for it.

If the request is genuinely ambiguous ("find me a coding skill" — coding is enormous), ask ONE narrowing question before searching. Otherwise, just go.

Where to search

Search all of these sources for every query — don't stop at the first hit. Each marketplace has a different catalog; comprehensive results require hitting several of them.

Anthropic-controlled (gold standard)

1. anthropics/skills — https://github.com/anthropics/skills
Anthropic's official skill examples (docx, pdf, pptx, xlsx, skill-creator, frontend-design, mcp-builder, etc.). 126k+ stars. The catalog lives at /skills/ at the repo root. Install via Claude Code with /plugin install document-skills@anthropic-agent-skills or via Claude.ai Settings → Features.

2. anthropics/claude-plugins-official — https://github.com/anthropics/claude-plugins-official
Anthropic-managed directory of high-quality, partner-vetted Claude Code plugins. Includes third-party integrations (SonarQube, Upstash Context7, CrowdStrike, GitLab, etc.) alongside skill bundles. Good for enterprise/integration skills beyond document processing.

3. anthropics/claude-code — https://github.com/anthropics/claude-code
The Claude Code repo itself has a .claude-plugin/marketplace.json listing bundled plugins (PR review agents, frontend design, hooks toolkit, etc.). Worth checking for coding/workflow skills.

High-volume community marketplaces

4. claudemarketplaces.com — https://claudemarketplaces.com/
Curated directory of 2,500+ Claude Code plugin marketplaces — acts as a meta-directory. Browse by category (Frontend Development, Backend & APIs, Document & Office, Security, DevOps, etc.). Best for discovering whole plugin collections rather than single skills.

5. skillsmp.com — https://skillsmp.com
Largest raw aggregator — 900,000+ skills auto-synced from GitHub. Cross-platform (Claude Code, Codex CLI, ChatGPT), but Claude Code is a first-class filter. Good breadth; use as a sweep net after checking more curated sources.

6. skillhub.club — https://www.skillhub.club
7,000+ skills with AI quality ratings (S/A/B ranks). Rated on Practicality, Clarity, Automation, Quality, and Impact. Includes a live playground to test skills before installing. Most useful when you want quality assurance alongside discovery.

7. lobehub.com/skills — https://lobehub.com/skills
Broad aggregator spanning Claude, Codex, OpenCode, Cursor, and others. Large catalog but requires filtering for Claude-compatible skills (see Claude-only filter below). Good for surfacing niche skills that don't appear elsewhere.

8. mcpmarket.com/tools/skills — https://mcpmarket.com/tools/skills
Active directory focused on agent skills across Claude.ai, Claude Code, Codex, and ChatGPT. Well-maintained; categorized and searchable.

9. claudeskills.info/skills/ — https://claudeskills.info/skills/
Curated Claude-first marketplace (~140+ handpicked skills). Includes the complete official Anthropic collection and notable community contributions. Smaller catalog but more signal per entry.

Aggregator / open standard reference

10. agentskills.io — https://agentskills.io
The home of the open Agent Skills standard (cross-platform SKILL.md format). Points to conformant skill repositories across all platforms. Use this to understand what cross-platform skills exist and which can be adapted for Claude.

If a source goes stale or a new trusted source emerges, update this list rather than letting it drift.

How to search each source

• GitHub repos (anthropics/skills, claude-plugins-official, claude-code): Use the GitHub Contents API at api.github.com/repos/{owner}/{repo}/contents/{path} for a clean file listing. If the API is unreachable, web_search for site:github.com/{owner}/{repo} {query} works as a fallback. Don't rely on the plain GitHub HTML directory page — it returns navigation chrome, not the file list.
• Web marketplaces (claudemarketplaces.com, skillhub.club, etc.): Use web_search with the query + site filter, or web_fetch the search/category page directly. Most sites have search URLs like https://skillsmp.com/?q=pdf that web_fetch handles cleanly.
• Parallel is better than sequential: Fetch multiple sources simultaneously if your tooling allows. Aim to cover at least sources 1–5 for every query; add 6–10 for broad or hard-to-match requests.

Quality and popularity signals

Before surfacing a candidate, check whether real people have actually used it. Here's how to read the signals per source:

Reject skills with no visible adoption signal — if a skill has 0 stars, no install count, no rating, and comes from an account with no other activity, skip it. Don't surface something untested just because it matches the query.

When nothing with strong signals matches, say so explicitly rather than padding with low-signal results: "I found candidates but they have very little usage data — do you want me to list them anyway, or would you prefer to build a custom skill?"

Anthropic-controlled sources are always trusted regardless of star counts — they're officially maintained and pre-tested.

Claude-only filter

The Agent Skills format is an open cross-platform standard, but skills written for other ecosystems often depend on platform-specific tools and won't run in Claude as-is. Surface only Claude-compatible skills.

Reject any candidate prefixed by a competing platform:

• openclaw-*, voltagent-*, manus-*, goose-*, block/agent-skills/* (Goose), anything obviously Codex-only or Gemini-only.

Reject any candidate whose SKILL.md depends on competing-platform tooling:

• Hard requirements like pdauth CLI, the OpenClaw browser tool, the Goose goose CLI, Manus's VM, etc.
• Install instructions that route only through a non-Claude path with no Claude-native option.

Prefer candidates that:

• Live in an anthropics/* namespace.
• Explicitly mention "Claude Code" or "Claude.ai" in the description.
• Install via /plugin install foo@anthropic-agent-skills, drop into ~/.claude/skills/, or upload through Claude.ai Settings → Features.

When in doubt, peek at the SKILL.md and check what CLI/tool calls it actually expects.

Workflow

1. Restate the task in one sentence. Make sure you understand what the user wants the skill to do. If it's truly unclear, ask one question; otherwise proceed.
2. Search broadly — hit multiple sources. Don't stop at the first match. Start depth-first on the Anthropic-controlled sources (sources 1–3), then sweep the high-volume marketplaces (sources 4–9). For rare or niche requests, check all 10.
3. Read each candidate's SKILL.md (or at least the frontmatter description). That description field is the gold — it tells you exactly what the skill does and when it triggers. Don't guess from the folder name; folder names lie.
4. Filter ruthlessly: task fit AND Claude-compatible. Apply the Claude-only filter before ranking by task fit. Three strong matches beats ten loose ones. If nothing fits after filtering, say so plainly.
5. Format the output as a scannable list (see below).

Output format

Default to this shape — terse, no preamble, no closing pep talk:

Here's what I found:

**[skill-name]** — [one line: what it does and when to use it]
Source: [source label] · [direct link] · [adoption signal, e.g. "2.1k ⭐", "A-rank", "official"]

**[another-skill-name]** — [one line]
Source: [source label] · [link] · [signal]

[Optional brief note if matches are weak or if writing a custom skill might be the right move instead.]

Conventions:

• Cap at 5 candidates. If there are more, say how many and offer to expand.
• Rank by fit, not by source priority. A perfect match in a community marketplace beats a weak match in the official repo.
• Link directly to the skill folder (e.g., the GitHub tree URL), not the repo root — the user wants to land on something useful.
• If a candidate comes from a less-vetted source, note it briefly (e.g., "Source: skillsmp (community aggregator, less vetted)").
• If you found nothing, say so plainly: "Nothing in the curated sources matches. Closest adjacent skill is X, or you may want to author your own — the skill-creator skill can help." Don't fabricate.

What this skill does NOT do

• Install skills. Just point at them. The user wants candidates and summaries, not setup steps.
• Editorialize. Don't critique code quality, predict whether a skill is "good," or rank by anything other than task fit, unless the user asks.
• Search uncurated terrain. Random GitHub repos, forum threads, blog posts, gists — skip them. The whole point is trusted sources only.

Example

User: "Is there a skill that helps with reading messy PDFs and pulling tables out?"

Output:

Here's what I found:

**pdf** — Read, extract, and manipulate PDF files. Covers text/table extraction, OCR for scans, form filling, merging, splitting.
Source: anthropics/skills · https://github.com/anthropics/skills/tree/main/skills/pdf · official

**pdf-reading** — Focused on inspecting PDF contents and choosing the right extraction strategy for text-heavy vs. scanned vs. data-heavy docs.
Source: anthropics/skills · https://github.com/anthropics/skills/tree/main/skills/pdf-reading · official

The first is probably what you want unless you're specifically dealing with scanned or data-heavy PDFs.