Anthropic skills

Anthropic's own skills for documents, design, art, internal comms and building new skills and MCP servers.

Install

Claude Code
```
/plugin marketplace add anthropics/skills
```
Then /plugin install document-skills@anthropic-agent-skills or example-skills@anthropic-agent-skills.
Claude apps

Already available on paid Claude plans.

The reference set of skills from the company that makes Claude. It is the best place to see what a well-built skill looks like, and several of these are the default way to work with files in Claude.

What’s in it

Documents — docx, pptx, xlsx and pdf read, create and edit Office files and PDFs.
Design — frontend-design, theme-factory, canvas-design, brand-guidelines and web-artifacts-builder for interfaces, themes and visual pieces.
Building — skill-creator for writing and testing your own skills, mcp-builder for MCP servers, and claude-api as an API reference.

Before you install

The high scanner scores are a good lesson in reading scans: the document skills run Python to manipulate files, which is exactly what a static scanner is built to flag. That makes them expected, not safe by default. Install the plugin that matches what you need rather than everything.

Official

Published by the platform or the maker of the thing itself.

Access: Runs on your machine Works on local files and folders; no access to your online accounts.
Review: Reviewed — install with care
Scan: 100/100 SkillSpector 2.12.0 said do not install at 3337550
Checked: Sep 25, 2026

Each of the 19 skill folders was scanned on its own. Scores ranged from 0 to 100. Six scored 100: claude-api, docx, mcp-builder, pptx, skill-creator and xlsx. claude-api had the most findings (59 HIGH), and all of them are in its Markdown API docs. There were no CRITICALs and 106 HIGHs, and all were read. Most are false positives. 59 are "incomplete analysis" notes on files the scanner only partly read, such as long reference docs, HTML templates, a PDF and a tarball of 45 shadcn UI components. The "hidden instructions" hits are byte-order marks in bundled Office XML schemas. "rm -rf" and "sudo" are doc examples and the out.docx cleanup step. "Anti-refusal" is "errors don't apologize" and "always answer the question first". Credential, env-variable and context-leakage hits are API docs and scripts that pass your environment on to soffice or `claude -p`. The HTML templates load only p5.js from cdnjs and Google Fonts. CAUTION is for real behavior the skills need to do their jobs. docx, pptx, xlsx and pdf run bundled Python scripts on your files, and the three Office skills call LibreOffice. xlsx recalculates and saves the workbook in place, using a macro installed in a throwaway LibreOffice profile. Where AF_UNIX sockets are blocked, as in some sandboxes, the shared soffice.py compiles a small LD_PRELOAD shim with gcc in the temp folder, and it reuses one if it's already there. skill-creator runs `claude -p` subprocesses, which bill your account, and writes temporary command files into .claude/commands, which it deletes afterward. Its review viewer serves on 127.0.0.1:3117 and first kills whatever is listening on that port. webapp-testing runs the server commands you give it through a shell. web-artifacts-builder installs pnpm globally if it's missing and pulls unpinned npm packages.

Publisher: Anthropic
Source: github.com/anthropics/skills
Works with: Claude apps, Claude Code
Categories: Documents & content, UX, UI & design systems, Agent building, Evals
License: Per skill — most ship their own terms in LICENSE.txt; the four document skills (docx, pdf, pptx, xlsx) are proprietary.
Cost: Free

Curated by Andrew Borden-Chisholm. Notes on building with these tools are on the blog.