anthropics/skills

What It Does

Anthropic's official anthropics/skills repo is the canonical collection of agent skills for the Claude ecosystem. It contains 17 production and example skills spanning document creation (DOCX, PDF, PPTX, XLSX), developer tools (MCP builder, webapp testing, skill creator), creative skills (frontend design, algorithmic art, theme factory), and enterprise skills (brand guidelines, internal comms, doc co-authoring). Several of these (the document skills) power Claude.ai's built-in document capabilities in production. This isn't just a demo repo; it's the reference implementation for what "skills" look like in the Anthropic ecosystem.

The Good

Breadth and depth of skills. 17 skills covering the full spectrum from creative to technical to enterprise. The document skills (docx, pdf, pptx, xlsx) are especially impressive. They're 230 to 590 lines each, packed with real production knowledge. The DOCX skill alone is a mini-masterclass in docx-js, covering page sizes, DXA units, table rendering gotchas, tracked changes XML, and the critical "never use unicode bullets" rule.

The claude-api skill is a monster. 324 lines of meticulously organized documentation covering every SDK (Python, TypeScript, Java, Go, Ruby, C#, PHP), model versioning (Opus 4.7, 4.6, Sonnet 4.6, Haiku 4.5), thinking/effort parameters, prompt caching, managed agents, and a detailed reading guide. This is essentially Anthropic's API docs distilled into a skill. For any agent working with Claude's API, this is gold.

Skill-creator is legitimately useful. This is a meta-skill, a skill for creating skills. It includes eval infrastructure (run_eval.py, run_loop.py, aggregate_benchmark.py), a review viewer (generate_review.py), and a full iterative workflow: draft → test → evaluate → improve → repeat. The description optimization loop with train/test splitting is sophisticated.

Scripts that work once you've solved the dep mystery. with_server.py (webapp-testing) runs out of the box. The others (package_skill.py, evaluation.py, validate.py) run cleanly *after* you've installed four undocumented Python deps: see Smoke Test Results for the dep list and what fails without it. The scripts themselves are solid; the install story is rough.

Honest licensing. The document skills (docx, pdf, pptx, xlsx) are explicitly marked "Proprietary" while the example skills are Apache 2.0. Clear separation. No confusion.

Plugin marketplace integration. The .claude-plugin/marketplace.json defines proper plugin bundles (document-skills, example-skills, claude-api) that install via /plugin install in Claude Code. Clean distribution model.

The Bad

Script import issues. Running package_skill.py directly (python scripts/package_skill.py) fails with ModuleNotFoundError: No module named 'scripts'. You have to use python -m scripts.package_skill from the skill-creator directory. The README doesn't document this clearly.

Four undocumented Python dependencies, no requirements.txt. Our sandbox testing surfaced four missing deps the repo doesn't list anywhere: pyyaml (blocks skill-creator's package_skill and quick_validate), anthropic SDK (blocks mcp-builder/evaluation), mcp (also blocks mcp-builder/evaluation), and defusedxml (blocks docx/validate). The full required preinstall is in the Setup section. Plus all the office libs (python-docx, openpyxl, pypdf, python-pptx) for the document skills' scripts. For Anthropic's official reference repo, this is a glaring gap: there's no requirements.txt, no skill-level dep manifests, no mention in the README.

Inconsistent skill depth. Some skills are 590-line production bibles (docx); others are 32-line stubs (internal-comms). The brand-guidelines skill is 73 lines. While there's nothing wrong with short skills, the quality gap is noticeable.

Spec is just a redirect. The spec/agent-skills-spec.md file contains only "The spec is now located at https://agentskills.io/specification." For an official repo, you'd expect the actual spec to be vendored or at least summarized. A single-link file feels incomplete for an offline-capable skill system.

Template is too bare. The template skill is just 6 lines: name, description, and "Insert instructions below." Given how sophisticated the real skills are, a richer template with examples of common patterns (frontmatter, sections, reference file pointers) would help newcomers.

No versioning on individual skills. The marketplace.json has a single version (1.0.0) for the whole repo. Individual skills don't track their own versions. Makes it hard to know if a specific skill has been updated.

Smoke Test Results

We ran two sandboxed test passes in clean, isolated Linux sandboxes (ubuntu 24.04 / aarch64). The fresh pass shows what a first-time user encounters. The with-deps pass shows what works once you've reverse-engineered the missing dependency list.

Run A. Fresh sandbox, no Python deps preinstalled

$ git clone --depth 1 https://github.com/anthropics/skills && cd skills
✅ Cloned 15M, ~2 seconds.

$ python skills/webapp-testing/scripts/with_server.py --help
✅ usage: with_server.py [-h] --server SERVERS --port PORTS ...

$ cd skills/skill-creator && python -m scripts.package_skill ../frontend-design
❌ ModuleNotFoundError: No module named 'yaml'

$ python skills/mcp-builder/scripts/evaluation.py --help
❌ ModuleNotFoundError: No module named 'anthropic'

$ python skills/skill-creator/scripts/quick_validate.py --help
❌ ModuleNotFoundError: No module named 'yaml'

$ python scripts/package_skill.py --help
❌ No such file or directory (wrong cwd: must run from skills/skill-creator)

$ python skills/docx/scripts/office/validate.py --help
❌ ModuleNotFoundError: No module named 'defusedxml'

Pass rate: 1 of 6. The 5 failures are not bugs in the skills: they're missing dependency declarations. There's no requirements.txt at the repo root, the README doesn't list deps anywhere, and you discover them one ModuleNotFoundError at a time.

Full sandbox log →

Run B. Sandbox with deps preinstalled

After pip install pyyaml anthropic mcp defusedxml python-docx openpyxl pypdf python-pptx, the same six tests look very different.

$ python skills/webapp-testing/scripts/with_server.py --help
✅ usage: with_server.py [-h] --server SERVERS --port PORTS ...

$ cd skills/skill-creator && python -m scripts.package_skill ../frontend-design
✅ Skill is valid!
✅ Successfully packaged skill to: frontend-design.skill

$ python skills/mcp-builder/scripts/evaluation.py --help
✅ usage: evaluation.py [-h] [-t {stdio,sse,http}] [-m MODEL] ...

$ python skills/skill-creator/scripts/quick_validate.py --help
❌ SKILL.md not found (script doesn't implement --help)

$ python scripts/package_skill.py --help
❌ No such file or directory (still: wrong cwd)

$ python skills/docx/scripts/office/validate.py --help
✅ usage: validate.py [-h] [--original ORIGINAL] [-v] [--auto-repair] ...

Pass rate: 4 of 6. The 2 remaining failures are real script issues, not dep issues: quick_validate doesn't implement --help, and package_skill.py only works when invoked as a module from a specific directory.

Full sandbox log →

What the two runs tell you

The scripts in this repo are solid code: once you've figured out the deps. The 3-test jump from Run A to Run B is entirely about the undocumented dependency list, not about the scripts themselves. For Anthropic's official reference repo, where readers reasonably expect a working pip install story, this is the single biggest gap. Setup below has the full preinstall command.

Run C. Functional verification (does it do what it claims?)

Install testing only proves "the script started without crashing." We also ran four functional assertions against the actual outputs of skill-creator, the central meta-skill:

$ python -m scripts.package_skill ../frontend-design
✅ produced frontend-design.skill (5,967 bytes, valid ZIP archive)
✅ archive contains SKILL.md at the expected path

$ python scripts/quick_validate.py ../frontend-design
✅ Skill is valid! (exit 0 on a shipped example skill)

$ python scripts/quick_validate.py <broken-skill-no-frontmatter>
✅ No YAML frontmatter found (exit 1 on a constructed-broken input)

Functional pass rate: 4 of 4 for skill-creator. The validator correctly accepts good skills and rejects broken ones with a useful error. package_skill.py produces real, installable .skill archives: not just zero-exit success messages.

Not functionally tested - and this is genuine, not aspirational: docx/validate.py would need valid and intentionally-malformed .docx fixtures that we'd have to construct; mcp-builder/evaluation.py needs a live MCP server endpoint we don't have. Those skills work in the install sense (Run B passed). We can't yet say "we verified their output matches the docs." That's why the page shows functional ~ and not functional ✓.

Full functional log →

Setup Walkthrough

1. Clone: git clone --depth 1 https://github.com/anthropics/skills.git (~14 MB, clean)
2. Install the undocumented Python deps before you run anything. This single line covers everything our sandbox tests needed: pip install pyyaml anthropic mcp defusedxml python-docx openpyxl pypdf python-pptx. Without this, 5 of 6 scripts in the repo fail with ModuleNotFoundError.
3. Read: Start with README.md, which is well-structured with links to docs.
4. Install in Claude Code: /plugin marketplace add anthropics/skills then /plugin install document-skills@anthropic-agent-skills.
5. For Claude.ai: Skills are already available on paid plans.
6. For API: Use the Skills API quickstart guide.
7. Custom skills: Use the template/ directory or the skill-creator skill.

No API keys needed for the skill definitions themselves. Some skills (claude-api, mcp-builder) require API keys at runtime for their scripts.

Alternatives

1. obra/superpowers (193K stars). The most-starred agent skills repo. Methodology-focused rather than skill-specific. Complementary, not a replacement.
2. sickn33/antigravity-awesome-skills (37K stars). 1,400+ installable skills across Claude Code, Cursor, Codex CLI, Gemini CLI. Community-driven alternative for breadth.
3. addyosmani/agent-skills (42K stars). Production-grade engineering skills curated by a Google Chrome engineer. More focused on coding best practices.