Claude Skills Registry

Core repo: logic + index + site.
Main repo (merged artifact): https://github.com/majiayu000/claude-skill-registry
Data repo (skills archive): https://github.com/majiayu000/claude-skill-registry-data
Authority: core workflows are canonical; main is a publish mirror.

A searchable Claude Code skills registry, aggregated from GitHub and community sources.

What is this?

Search skill descriptions, inspect source links and registry signals, and find installation guidance. This core repository maintains the discovery, archive validation, search-index, and site publishing pipeline.

The same skill copied into multiple repositories counts once. Independent skills use a normalized name and an exact Markdown-body fingerprint; sources without a fingerprint count separately by install path and branch.

Read the choose and install one skill guide for source selection, licenses and troubleshooting.

Three ways to use:

1. Web Search - Fast browser-based search
2. skills CLI - Install into Claude Code, Codex, Cursor and other agents
3. API - Direct JSON access

Repo layout note: core owns workflows/pipeline logic, data stores skills/**, and main is generated from core + data. See SCHEME2_SPLIT.md.

Highlights

• Skill Index - Independent skills (see badge for live count), with matching source copies linked in each detail view
• Rich Categories - Development, Testing, DevOps, Design, and more
• Daily Updates - Automated crawling/validation by core scheduled workflows
• Quality Indexed - Metadata, descriptions, and star counts
• Lightweight Search - Gzip-compressed index for fast client-side search

Operational Ownership

• Core: source of truth for workflows, crawling, scanning, and index/site generation
• Data: canonical archived skill tree (skills/**)
• Main: publish artifact for merged browsing/compatibility consumers
• Publish contract: core dispatches main publish with pinned core_sha + data_sha
• Non-goal for main: main does not initiate canonical syncs, archive updates, or index generation

Resumable full discovery

The Sunday schedule starts a full discovery cycle. It snapshots the repository inventory and processes it in batches with a 100-minute discovery budget (120-minute runner timeout). Daily schedules resume an unfinished cycle before returning to ordinary bounded daily discovery. A new Sunday does not reset an unfinished inventory. GitHub search limits still apply; a cycle means all repositories in that snapshot, not every skill on GitHub.

sources/learning/full-discovery-progress.json records the inventory, the next repository, and the cycle’s start/completion timestamps. Budget exhaustion leaves an unfinished repository at the cursor; its already archived files are skipped on the next batch. Each batch still passes the full archive security checks. The cursor is committed to core only after the data repository push succeeds. A failed batch therefore retries from the last committed cursor rather than skipping data that never reached the archive. Network errors and incomplete search responses fail the batch; repositories confirmed unavailable are recorded in the candidate report.

The Actions summary distinguishes a successful batch from a completed cycle. The weekly alert stays open until the entire cycle finishes and the publish job succeeds. To retry a failed batch, start a new workflow dispatch with full_scan=true; GitHub’s Re-run remains reserved for replaying an existing immutable publish handoff. skip_discovery=true explicitly bypasses the cycle without advancing its cursor or closing its alert.

Release, Proof, And Support

• Release status: daily catalog refreshes are represented by live generated metadata, not GitHub Releases. GitHub Releases are reserved for intentional pipeline/API milestones.
• Live proof: use the Web Search UI, the badges above, and stats.json to verify current generated data.
• Routing: pipeline bugs, source intake, registry schema, Pages/search, and publish orchestration belong in this core repo. Archived skill body issues belong in claude-skill-registry-data. Generated mirror questions belong in claude-skill-registry only when the artifact itself is wrong.

curl https://majiayu000.github.io/claude-skill-registry/stats.json
curl https://majiayu000.github.io/claude-skill-registry/search-index-lite.json

Quick Start

Option 1: Web Search

For a complete first installation, follow the code review setup guide. It covers skill selection, project-local installation, verification, a first review prompt, and removal.

Visit https://majiayu000.github.io/claude-skill-registry/

For clone/update tips on large repositories, see docs/FAST_CLONE.md.

Option 2: CLI

Skill guides and search results show a ready-to-copy command for supported source paths, or a source link for manual installation. The skills CLI asks which coding agents to install into:

npx skills add https://github.com/anthropics/skills/tree/main/skills/docx

# Or install globally for one agent
npx skills add https://github.com/anthropics/skills/tree/main/skills/docx -g -a claude-code  # ~/.claude/skills/
npx skills add https://github.com/anthropics/skills/tree/main/skills/docx -g -a codex        # ~/.agents/skills/
npx skills add https://github.com/anthropics/skills/tree/main/skills/docx -g -a cursor       # ~/.cursor/skills/

The earlier sk CLI (claude-skill-manager) is no longer developed.

For skill authors

Each skill guide includes a “Listed on Claude Skills Registry” badge snippet that links readers to the registry home. Copy it from the “For the skill author” section of your skill’s guide page.

Option 3: Direct API

The public JSON surface is versioned in docs/artifact-api-contract.md. Use that contract when integrating with pointer files, manifests, shards, gzip variants, and deprecation markers.

# Startup search index and bounded full-search shards
curl https://majiayu000.github.io/claude-skill-registry/search-index.json
curl https://majiayu000.github.io/claude-skill-registry/search-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry/search-shards/part-000.json

# Deduplicated catalog index with quality/security/install signals
curl https://majiayu000.github.io/claude-skill-registry/search-index-lite.json
curl https://majiayu000.github.io/claude-skill-registry/quality-index.json
curl https://majiayu000.github.io/claude-skill-registry/quality-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry/quality-shards/part-000.json
curl https://majiayu000.github.io/claude-skill-registry/security-index.json
curl https://majiayu000.github.io/claude-skill-registry/security-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry/security-shards/part-000.json
curl https://majiayu000.github.io/claude-skill-registry/ranking-index.json
curl https://majiayu000.github.io/claude-skill-registry/ranking-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry/ranking-shards/part-000.json

# Lightweight registry summary (counts only)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry-core/main/registry_summary.json

# Full registry manifest and shards (merged publish artifact)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry-manifest.json
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry-shards/00.json

# Compatibility registry pointer (merged publish artifact)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry.json

# Category manifest and bounded parts
curl https://majiayu000.github.io/claude-skill-registry/categories/index.json
curl https://majiayu000.github.io/claude-skill-registry/categories/development/manifest.json
curl https://majiayu000.github.io/claude-skill-registry/categories/development/part-000.json

# Legacy category pointer
curl https://majiayu000.github.io/claude-skill-registry/categories/development.json

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Layer 1: Data Collection                                       │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐          │
│  │ GitHub Crawl │→ │ Download     │→ │ Security     │          │
│  │ (discover)   │  │ (sync)       │  │ (scanner)    │          │
│  └──────────────┘  └──────────────┘  └──────────────┘          │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Layer 2: Index Generation                                      │
│  ┌────────────────┐  ┌────────────────┐  ┌────────────────┐    │
│  │ search-index   │  │ categories/    │  │ featured.json  │    │
│  │ .json          │  │ *.json         │  │ (featured set) │    │
│  └────────────────┘  └────────────────┘  └────────────────┘    │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Layer 3: Consumption                                           │
│  ┌────────────────┐  ┌────────────────┐  ┌────────────────┐    │
│  │ Web UI         │  │ sk CLI         │  │ API            │    │
│  │ (GitHub Pages) │  │ (Go)           │  │ (JSON)         │    │
│  └────────────────┘  └────────────────┘  └────────────────┘    │
└─────────────────────────────────────────────────────────────────┘

Search Index Schema

// Lightweight index for fast search
interface SearchIndex {
  v: string;           // Version (date)
  t: number;           // Indexed skill count (always s.length)
  s: SkillMini[];      // Skills array
}

interface SkillMini {
  n: string;           // name
  d: string;           // description (truncated 80 chars)
  c: string;           // category code (dev, ops, sec, etc.)
  g: string[];         // tags (max 5)
  r: number;           // stars
  i: string;           // install path
}

---

Directory Structure (Core)

claude-skill-registry-core/
├── registry.json           # Lightweight core registry metadata
├── docs/                   # GitHub Pages
│   ├── index.html          # Web search UI
│   ├── search-index-lite.json
│   ├── search-index.json   # Compatibility pointer to bounded shards
│   ├── search-index-manifest.json
│   ├── search-shards/      # Bounded full search index parts
│   ├── featured.json       # Featured skills snapshot
│   └── categories/         # Category manifests and bounded parts
├── sources/                # Data sources
│   ├── anthropic.json
│   ├── community.json
│   └── skillsmp.json
├── scripts/                # Build scripts
│   ├── build_search_index.py
│   ├── discover_by_topic.py
│   ├── security_scanner.py
│   └── ...
└── (no committed skills/)  # skills/** lives in registry-data; mounted in CI when needed

The merged claude-skill-registry publish artifact additionally contains registry-manifest.json and registry-shards/**.

---

Categories

Canonical category slugs, short codes, governance status, and heuristic keywords live in taxonomy/categories.yaml. Pipeline scripts read that file instead of keeping their own category lists. Legacy category names are diagnostic inputs only; they are routed to review and are not accepted as publish categories. The canonical category proposal is documented in docs/plan/canonical-taxonomy-proposal.md, and the current target count report is docs/plan/canonical-category-target-counts.json. Validate the taxonomy itself with python scripts/check_taxonomy_governance.py. To validate publish output against canonical categories, pass repeated --publish-category values. To review category quality across the archive before moving anything, run python scripts/audit_category_quality.py --skills-dir skills. The default audit uses metadata and paths for a fast full-archive pass; add --include-frontmatter when checking frontmatter/category drift. The audit also reports non-standard nested skill paths such as category/category/skill/SKILL.md. For those layout issues, use python scripts/normalize_skill_depth.py --skills-dir skills --json to review the exact move plan before applying it. For semantic reclassification, generate a review-only migration plan with python scripts/plan_category_migration.py --skills-dir skills --output category-migration-plan.json. The plan records action, confidence, source categories, target category, keyword signals, and reason for every proposed change; it does not move files. For a bounded second-pass model review, set MIMO_API_KEY and run python scripts/review_category_plan_with_llm.py --plan category-migration-plan.json --output category-llm-review.json --checkpoint-jsonl category-llm-review.checkpoint.jsonl --resume. The default endpoint is https://token-plan-sgp.xiaomimimo.com/v1 with mimo-v2.6-pro. The default request uses --thinking disabled and --max-completion-tokens 1024 so the model output budget is reserved for the required JSON; raise the token value for larger reviews, or pass --thinking default when using a non-MiMo endpoint that does not accept the provider-specific thinking field. The report remains review-only: it records model category, confidence, decision, parse status, token/thinking policy, and evidence without modifying the archive. The checkpoint file is append-only JSONL, so interrupted long reviews can resume without re-calling the model for completed candidates.

Category counts are published in categories/index.json; full category payloads are available through categories//manifest.json and bounded part-*.json files. The legacy categories/.json URL is now a small compatibility pointer. The publish path runs python scripts/check_category_artifacts.py --categories-dir docs/categories so legacy category files cannot silently grow back into large full-payload JSON. Common category codes include:

---

Roadmap

Current Status

• [x] Index count tracked by the badge (core registry.json)
• [x] Archive size: tracked by badge (raw SKILL.md count from stats.json)
• [x] Daily auto-update via GitHub Actions
• [x] Security scanning for all skills
• [x] sk CLI for installation

In Progress

• [x] Lightweight search index (gzip-compressed; see stats.json)
• [x] Web search UI (GitHub Pages)
• [x] GitHub Pages deployment (https://majiayu000.github.io/claude-skill-registry/)

Planned

• [ ] AI semantic search (vector similarity)
• [ ] Skill recommendations (based on usage)
• [ ] Version tracking for skills
• [ ] Skill quality scoring
• [ ] API rate limiting and caching

---

Contributing

Add Your Skill

Option 1: Submit via Issue

1. Open an issue
2. Use the “Add Skill” template
3. Provide: repo URL, name, description, category

Option 2: Submit via PR

1. Fork majiayu000/claude-skill-registry-core
2. Add your skill to sources/community.json:

{
  "name": "your-skill-name",
  "repo": "your-username/your-repo",
  "path": "optional/path/to/skill",
  "description": "What your skill does",
  "category": "development",
  "tags": ["testing", "automation"]
}

3. Submit a PR to majiayu000/claude-skill-registry-core (not claude-skill-registry, which is a generated publish artifact)

Validate a submitted skill locally

Before opening an issue or PR, confirm that the source repository is public and that its SKILL.md starts with YAML frontmatter containing non-empty name and description values:

---
name: your-skill-name
description: A concise description of what the skill does.
---

Remove secrets, tokens, and private endpoints from the skill and any bundled files. From a core checkout with the repository requirements installed, run:

# If the PR updates sources/community.json, validate the source entry
python scripts/validate_sources.py community.json

# Scan the checked-out skill directory, including bundled files
python scripts/security_scanner.py /path/to/skill-directory

These are the normal community-submission checks. The python scripts/rebuild_registry.py --help and python scripts/build_search_index.py --help interfaces are archive-wide maintainer references; rebuilding the registry or indexes is not a normal submission prerequisite.

Report Issues

We welcome feedback! Please open an issue for:

• Bugs - Search not working, incorrect data
• Feature requests - New categories, better search
• UX improvements - Web UI, CLI enhancements
• Data quality - Duplicate skills, wrong categories

👉 Open an Issue

Contribute Code

# Clone the core repo (authoritative pipeline repo)
git clone --filter=blob:none --sparse https://github.com/majiayu000/claude-skill-registry-core.git
cd claude-skill-registry-core

# Pull only what you need (add more paths later as needed)
git sparse-checkout set --cone docs scripts sources schema

# Install dependencies
uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

# Build search index locally
python scripts/build_search_index.py --registry registry.json --output docs

# Test web UI
cd docs && python -m http.server 8000
# Visit http://localhost:8000

See docs/FAST_CLONE.md for more options (existing clones, getting full checkout, Windows notes).

---

Related Projects

---

License

MIT License applies to the registry code/pipeline only - see LICENSE for details.

Third-Party License & Attribution

Third-party skills under skills/** keep their original licenses and copyright ownership.

• Repository-level MIT does not relicense third-party skill content.
• Every imported skill metadata file should include:
  • author
  • source_url
  • license
  • copyright
  • permission_note
  • distribution (compatible or restricted)
• restricted entries are not MIT-compatible by default and require explicit upstream permission before redistribution/use.
• Notices are generated by compliance checks into THIRD_PARTY_NOTICES.md.
• Metadata compliance runs in advisory mode by default to avoid blocking ingestion; strict blocking can be enabled when needed.

---

Made with ❤️ for the Claude Code community

Author requests

Third-party content retains its original ownership and license. Authors can request removal or attribution corrections. Public skill guides include metadata and source links; they do not republish instruction bodies.