Anthropic skills: the official GitHub repo, the document skills and which ones to install

When people search for "Anthropic skills" they usually mean one of two things: the public GitHub repository where Anthropic publishes its own skills, or the four document skills that let Claude produce real Word, Excel, PowerPoint and PDF files. They are the same thing seen from two sides. This guide covers what is actually in the official collection, how to install it in each product, which skills earn a place in a team setup, and what the licences let you do.

If you are new to skills as a concept, start with the Claude skills pillar guide, which covers how a skill works and how to write one. If you are deciding between a skill and an MCP connection, read Claude skills vs MCP. This page sticks to Anthropic's own collection.

The anthropics/skills repository

The official repo is github.com/anthropics/skills, described as the "Public repository for Agent Skills" and actively maintained. The top level has four things worth knowing about:

• skills/ holds the skills themselves, one folder each with a SKILL.md.
• spec/ holds the Agent Skills specification. The open standard itself lives at agentskills.io.
• template/ is a minimal starter skill.
• .claude-plugin/marketplace.json turns the repo into a Claude Code plugin marketplace called anthropic-agent-skills.

The README is explicit that these skills are "provided for demonstration and educational purposes only" and that what Claude does in its own products may differ from the code shown. Take that seriously: treat the repo as a set of well-built reference implementations, not as a supported product with an SLA.

What is in the collection

As of this writing the skills/ folder contains 19 skills. The marketplace groups them into five plugins:

The split matters. example-skills is a grab bag: some entries are genuinely useful at work, others (algorithmic art, Slack GIFs) exist to show what the format can do. The document skills are a different category entirely. They are the same skills that power Claude's file creation in claude.ai, published so developers can see how a complex production skill is structured.

The document skills: docx, xlsx, pptx, pdf

Each document skill is a SKILL.md plus a scripts/ folder of helpers.

docx creates new Word documents with the docx npm library, and edits existing ones by unzipping the file and changing word/document.xml directly, since that library cannot open existing files. It reads documents with pandoc, handles tracked changes and comments, and ships an accept_changes.py script for producing a clean copy.

xlsx uses openpyxl for formulas and formatting and pandas for bulk data. Its hard rules are the useful part: use formulas rather than hardcoded results, and never ship a file with formula errors. Because openpyxl writes formulas without cached values, the skill runs a recalc.py script that recalculates every formula through LibreOffice and reports errors as JSON.

pptx builds decks with pptxgenjs, reads them with markitdown, and renders slides to images through LibreOffice and Poppler so Claude can check the visual result instead of trusting the XML.

pdf covers extraction with pdfplumber and pypdf, creation with ReportLab, merging and splitting with qpdf, OCR with pytesseract, and a set of scripts for filling both fillable and non-fillable forms.

The practical consequence for Claude Code users: these skills expect tools on the machine. In claude.ai and the API they run in Anthropic's sandbox, where the dependencies are preinstalled. On your laptop, the docx, pptx and xlsx workflows lean on LibreOffice, Poppler and pandoc, and the PDF skill on qpdf and Tesseract for OCR. Install what is missing before you judge the skill.

Installing in Claude Code

Claude Code reads skills from ~/.claude/skills/ (personal) and .claude/skills/ (per project), and from installed plugins. Anthropic's own docs note that the pre-built document skills are not part of Claude Code out of the box; only the Claude API skill comes bundled. The repo's marketplace fills that gap. Inside a Claude Code session:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

You can also run /plugin, choose "Browse and install plugins", pick anthropic-agent-skills and then the plugin you want. After that, Claude picks a skill up automatically when your request matches its description ("turn this CSV into a formatted Excel workbook"), or you can invoke it by name. Plugin skills are namespaced, so the PDF skill is reachable as /document-skills:pdf, and usually by the bare /pdf if nothing else claims that name.

If you only want one skill from example-skills, the plugin is all or nothing. The Apache-licensed skills can instead be copied folder by folder into .claude/skills/ and committed with the project. Do not do that with the document skills; see the licence section below. For more on setting up Claude Code itself, see how to install Claude Code.

Using them in Claude.ai and the API

Claude.ai. You do not install the document skills here. They are used automatically when Claude creates files, provided code execution is on. On individual plans that is Settings, Capabilities; on Team and Enterprise an owner first enables code execution and Skills in the organisation settings. Skills from the repo beyond the document set are, according to the README, already available on paid plans. To add your own, zip the skill folder and upload it under Customize, Skills.

Claude API. Attach pre-built skills by ID in the container parameter and include the code execution tool, whose sandbox the skills run in:

{
  "model": "claude-opus-5-5",
  "max_tokens": 4096,
  "container": {
    "skills": [
      { "type": "anthropic", "skill_id": "xlsx", "version": "latest" }
    ]
  },
  "tools": [
    { "type": "code_execution_20250825", "name": "code_execution" }
  ],
  "messages": [
    { "role": "user", "content": "Build a 12-month budget workbook with formulas." }
  ]
}

Up to 20 skills can be attached per request. API skills run with no network access and no runtime package installation, which is the right default for anything touching business data. Custom skills uploaded through /v1/skills are shared across the workspace. Note that skills do not sync between surfaces: a skill uploaded to claude.ai is not available in the API or in Claude Code, and the other way round.

Which ones are worth using

Our opinion after using the collection on real work:

• document-skills: yes. The single most useful thing in the repo. If your team generates reports, board decks or finance workbooks from Claude Code, this is what stops broken formulas and slides that look fine in XML and terrible on screen. The xlsx skill alone justifies it for anyone in finance and reporting.
• skill-creator: yes. It walks through writing a skill, runs evals against it and helps tune the description so the skill triggers when it should. Use it before you write your team's first internal skill.
• mcp-builder: yes, for engineers. A solid guide for building MCP servers in Python (FastMCP) or TypeScript. Pair it with our MCP security best practices before anything goes near production data.
• webapp-testing and frontend-design: worth a try in front-end projects. The first drives Playwright to check local web apps; the second pushes Claude away from generic-looking UI.
• claude-api: already there if you use Claude Code, since it ships bundled.
• doc-coauthoring and internal-comms: situational. Useful structure for specs and status updates, but most teams will want their own versions with house style.
• The rest: fun, instructive, rarely needed at work.

Licence notes

Read the licences before you fork or redistribute anything.

Most skills carry an Apache 2.0 LICENSE.txt. You can copy, modify and ship them, including inside a company plugin, as long as you keep the licence and notices.

The four document skills are different. Their LICENSE.txt says use is governed by your agreement with Anthropic and, among other restrictions, forbids reproducing the materials, creating derivative works, and distributing them to third parties. In practice: installing them through the official marketplace and using them is what Anthropic intends; vendoring a modified copy of the xlsx skill into your internal skill pack is not. If you need a house-style variant, write your own skill that sits alongside the official one rather than editing it.

Governance: skills are code, treat them like code

Anthropic's documentation is blunt about this: use skills only from trusted sources, because a skill can direct Claude to run code and call tools in ways that do not match its stated purpose. The official repo is about as trusted as it gets, but the same discipline should apply to every community skill someone on your team finds: review the scripts, pin a version, and know what data the agent can reach while it runs.

Two details matter for EU teams. First, Anthropic states that Agent Skills is not covered by zero data retention arrangements; skill definitions and execution data follow its standard retention policy. Second, skills decide how Claude works, while MCP connections decide what it can reach. That second part is where most of the risk sits. At Walma we run that layer in the customer's own Azure tenant in an EU region: every model, the MCP connections, policy, cost control and logging in one place. See the AI Hub for how it fits together, or the Claude Code guide for running the agent itself under policy. If an installed skill never seems to fire, the troubleshooting checklist in our skills guide covers the usual causes.