claude-skills-library

Claude Skills Library

114 Agent Skills for every skills-aware runtime, and 7 enforcement packs for Claude Code

Skills teach an agent a domain. Packs make the agent go through the gate. This is the only public library that ships both, with provenance pinned to commits, counts that check themselves in CI, and hooks that block "done" until the evidence exists.

Install · Packs · Catalog · Strategy · Contribute · FAQ

Install

Three ways in. Pick one.

# 1. Claude Code plugin — auto-updating, skills discovered on their own
/plugin marketplace add frankxai/claude-skills-library
/plugin install claude-skills-library@claude-skills-library

# 2. skills.sh — the top-level skills, or a curated lane
#    (the skills nested under anthropic/, creative/ and technical/ come via the plugin or a clone)
npx skills add frankxai/claude-skills-library
npx skills add frankxai/skills            # AI architects: MCP, orchestration, model routing
npx skills add frankxai/creator-skills    # creators: video, music, images, brand voice

# 3. Clone and copy only what you want
git clone https://github.com/frankxai/claude-skills-library.git ~/claude-skills-library
mkdir -p ~/.claude/skills
cp -r ~/claude-skills-library/free-skills/mcp-architecture ~/.claude/skills/   # one
cp -r ~/claude-skills-library/free-skills/* ~/.claude/skills/                  # all

Some skills are namespaced (free-skills/anthropic/pdf). Copy the folder that directly contains the SKILL.md. Restart Claude Code and the skills are discovered.

Then work. Skills fire on their own from their description, or you name one:

"Design an MCP server with the mcp-architecture skill."
"Review this PR with the github-code-review skill."
"Run the verification-quality skill before you call this done."

What is here

Two layers. The first is what every skills repo has. The second is what none of the large ones ship.

A skill is loaded when the task matches its description. That makes it available, not mandatory. A pack adds what turns a suggestion into a default, and the mechanism differs by pack. web-excellence installs three hooks: a SessionStart note, a PreToolUse reminder on the first UI edit, and a Stop hook that blocks once if UI changed with no audit, each shipped with the test that proves it cannot be satisfied by its own reminder text. estate-guard installs a PreToolUse gate that denies the hard stops and a taint hook for fetched content. routine-contract ships a CI workflow and a digest primitive that fails the run when nothing durable was written, for each routine you wire to it. risk-classifier ships a fail-closed classifier script for you to wire into CI. agent-infrastructure installs skills and checks installed copies against canon. film-excellence installs a skill and templates and is advisory.

Packs: skills that enforce

Installed into 19 repositories across the FrankX, Starlight, and Arcanea estate as of 2026-10-09. Re-run an installer to upgrade. web-excellence, routine-contract and agent-infrastructure move an installed skill that differs to .claude/skills/.replaced/<name>, keeping only the most recent set-aside copy; estate-guard and film-excellence overwrite their skill in place. In every case, commit or copy a customised skill before re-running. All but risk-classifier accept --dry-run to show the diff first.

git clone --depth 1 https://github.com/frankxai/claude-skills-library /tmp/csl
/tmp/csl/packs/web-excellence/install.sh "$PWD"
/tmp/csl/packs/estate-guard/install.sh "$PWD"

Then one line in the repo's CLAUDE.md: Website / web-design work goes through the web-release-gate skill first.

Why this library

The largest skills repositories are excellent at judgment: how to think about a design, a diff, a test. They are weak at proof. None of them measures anything, produces evidence at real breakpoints, refuses to let work be called done, or tells you where a vendored skill came from and whether it has moved upstream since. This library is built around those gaps.

The honest gap on our side: evals. All 114 in this library read well. None yet ships evidence that it still behaves under pressure. That is the next moat and the first bet in docs/STRATEGY.md.

Skill categories

Counts are the groupings in docs/CATALOG.md, regenerated from the tree by scripts/generate_catalog.py. 114 skills across 11 categories.

Browse the full catalog, or the generated site at docs/index.html.

A few to start with

How the pieces fit

flowchart LR
    SIS["Starlight Intelligence System<br/>substrate: memory, agents, SIP"] --> ACOS["Agentic Creator OS<br/>canonical skill source"]
    ACOS -->|"mirror the strongest"| LIB["claude-skills-library<br/>114 skills · 7 packs"]
    UP["Upstreams<br/>Vercel · Emil Kowalski · Anthropic · ..."] -->|"pinned, licensed, drift-checked"| LIB
    LIB -->|"/plugin · npx skills add · cp"| CC["Claude Code"]
    LIB -->|"runtimes/"| RT["Cursor · Codex · Gemini CLI<br/>OpenCode · Antigravity"]
    LIB -->|"packs/*/install.sh"| REPOS["19 estate repos<br/>hooks + CI + tests"]

Starlight is the substrate. ACOS is where skills are authored and battle-tested first. This repo is the public, install-from catalog, and the packs are how its standards reach every other repo.

Skills vs. MCP vs. prompts

Use a skill to teach the agent how to do something well. Use MCP to give it access to a system. Use a prompt for a one-off. Many skills here pair with an MCP server.

Works with six runtimes

Agent Skills are a portable format. One adapter guide per runtime lives in runtimes/.

The rationale for MCP as the universal bridge is in MULTI_RUNTIME.md.

Anatomy of a skill

free-skills/<skill-name>/
├── SKILL.md          # required: frontmatter + instructions
├── references/       # optional: deep docs loaded on demand
├── scripts/          # optional: executable helpers the agent runs
└── assets/           # optional: templates, fonts, icons used in output

---
name: mcp-architecture          # lowercase, hyphenated, 64 chars or fewer
description: Design and implement MCP servers... Use when architecting an MCP server...
---

The description is the most important line. It is what the model uses to decide when to load the skill. State what it does and when to use it, with trigger keywords.

Authoring standard: spec/README.md. Start here: template/SKILL.md.

Quality and validation

Zero-dependency scripts enforce the standard on every push and pull request (validate.yml, pack-tests.yml, catalog-check.yml):

python3 scripts/validate_skills.py           # frontmatter + structure, exits non-zero on failure
python3 scripts/generate_catalog.py --check  # docs/CATALOG.md matches the tree
python3 scripts/generate_site.py --check     # docs/index.html matches the tree
python3 scripts/check_counts.py              # README counts match the tree
python3 scripts/check_links.py               # no broken internal links
node scripts/build-catalog.mjs --check       # catalog/skills.json matches the marketplace
node --test scripts/tests/absorb.test.mjs    # the four absorption gates still refuse

Inbound external work runs through ABSORPTION.md: license, provenance, distinctness, attestation. node scripts/absorb.mjs refuses rather than warns. What has entered is recorded in ABSORBED.md.

[!NOTE] Pinned upstream, 2026-10-10. affaan-m/ECC at 4eb71d92a39cab44ad40ac9d8a6a5ccb4029d6c2 (MIT) is a dependency of the Claude config repo, not a prose dump in free-skills/. Patterns that cleared the four gates are listed in ABSORBED.md. The port receipt is docs/ecc-port-report.md. Watching a repo is not the same as absorbing it.

Catalog site

scripts/generate_site.py renders the catalog to docs/index.html. It deploys as a static site to Vercel with the committed vercel.json, or to GitHub Pages from the docs/ folder.

Strategy

The plan to make this the reference skills library, in one page: docs/STRATEGY.md. The short version:

1. Evals before volume. Ship the eval suite, the passing cases, and the failing ones, for the top skills first.
2. Packs are the product. Enforcement, provenance, and tests are what a skills directory cannot give you.
3. Absorb patterns, not prose. The best skills from other libraries enter through the four gates, with credit, or are depended on, never pasted.
4. One canonical copy. Skills that exist in several estate repos get mirrored here and installed back from here.

Community

• Ask a question, share a working example, or suggest an idea.
• Report a bug or request a skill with the affected skill and a small, shareable example.
• Contribute a fix or skill. Evidence from real use and clear source attribution are what get a PR merged.
• Report a vulnerability privately.

See SUPPORT.md for routing and CODE_OF_CONDUCT.md for participation. Remove credentials and private material before posting.

Contributing

1. Read CONTRIBUTING.md and the authoring standard.
2. Copy template/ to free-skills/<skill-name>/ and write your SKILL.md.
3. Run python3 scripts/validate_skills.py and python3 scripts/generate_catalog.py until both are clean.
4. Open a pull request.

Related repositories

• frankxai/skills and frankxai/creator-skills: the curated, actively maintained lanes for AI architects and for creators. This library is the full catalog behind them.
• frankxai/agentic-creator-os: the canonical, larger source for skills, agents, and commands. New skills are authored and battle-tested there first.
• frankxai/Starlight-Intelligence-System: the substrate. Persistent context, memory, and the multi-agent layer the skills run on.
• frankxai/starlight-design-intelligence: the design and editorial kernel the web-excellence pack enforces.

FAQ

Does loading 114 skills bloat my context? No. Only each skill's name and description is preloaded for routing. The body loads when a skill triggers, and references/ load only when read. A library of 114 costs almost nothing until a skill fires.

How does the right skill get picked from a library this large? By the description. Every skill states what it does and when to use it with explicit trigger keywords. If one mis-triggers, sharpen its description. PRs welcome.

Do I have to install all of them? No. Install the plugin for everything, npx skills add a curated lane, or copy the folders you want.

Is this affiliated with Anthropic? No. It is an independent, MIT-licensed library built on Anthropic's open Agent Skills format. Skills under free-skills/anthropic/ are upstream reference skills that retain their original licenses.

Can I use these outside Claude Code? Yes. See runtimes/ for Antigravity, OpenCode, Codex, Gemini CLI, and Cursor.

Would you pay for a skill that proves it works?

Every first-party skill here is free and MIT. That is not changing. Skills under free-skills/anthropic/ keep Anthropic's own terms.

A skill is prose until something tests it. The idea on the table: a pack where every skill ships with its eval suite, the cases it passes, and the cases it fails. There is no paid product today. Before building one, say whether it should exist, and name the band.

Tell us what it is worth

The first fifty people to answer get one skill with its full eval suite and run output, free. At fifty answers the eval harness is published open-source regardless.

License

MIT. See LICENSE. Imported skills retain their upstream licenses, recorded in THIRD_PARTY_NOTICES.md and, with one exception, in each vendored folder; web-design-guidelines carries its MIT record in the pack's SOURCES.md until the file is vendored.

Star the repo · Catalog · Issues · Discussions