Vibeleaderboard
Index / app
Visit github.com
Category
AI Agents
Rank
Type
APP
Builder
anthropics
Date

About

Anthropic's public repository of Agent Skills — modular, composable capability files that extend what Claude and other coding agents can do.

What it does

A corpus of worked example skills from the team behind Claude, each one a folder holding a short instruction document with a name and description up front, plus optional scripts and reference material. An agent keeps only the name and description in memory at all times; the full instructions load when a matching task appears, and bundled files load only when the instructions call for them. The same tree doubles as an installable marketplace of three ready-made bundles, several of which back a shipping document pipeline.

Why it's ranked here

This is the reference corpus for a skill format that other agent runtimes have since adopted, and it is unusual in shipping the real implementations behind a production document pipeline rather than toy samples. Any bundle installs with one command, each skill declares its dependencies, and the document skills are dense with specific, checkable failure modes instead of generic advice. Two caveats are visible in the repository itself: several skills are source-available rather than open source, and the format specification lives elsewhere.

What's good

The document skills read as failure catalogues, and that is where the value sits. The spreadsheet instructions name the exact modern Excel functions that need a special prefix, warn that spill-style functions can produce silently truncated results with zero reported errors, and note that a clean recalculation only proves formulas evaluate, not that they are correct. The word-processing skill treats documents from outside parties as untrusted, telling you to delete symlinks from unpacked archives, and its commenting helper parses XML defensively and escapes text by default. This is operator knowledge, written down.

Tradeoffs

Licensing is not uniform: the word-processing, PDF and spreadsheet skills declare themselves proprietary while the readme describes many others as Apache 2.0, so the collection cannot be adopted wholesale. The readme also says these are demonstrations and real behaviour may differ. The bundled evaluation harness pins an older model by default, executes only the first tool call in a response so parallel calls are dropped, and grades answers by exact string match, workable for IDs and numbers and little else. One art skill instructs the model to praise its own output, a self-description rather than a quality control.

How to use it well

Install one bundle rather than the whole tree; the marketplace manifest lists which skills each carries. When writing your own, spend the effort on the description, since that is what fires the skill. Copy the document skills' pattern: enumerate the specific ways a task fails and the exact command that verifies the output. Skip the deliberately empty template and start from the skill that teaches skill authoring; its paired runs with and without the new skill are the part worth copying, since a skill that changes nothing is not earning its context. Treat bundled scripts as black boxes driven through their help flag.

Technical notes+

The marketplace manifest at .claude-plugin/marketplace.json defines three plugins, document-skills with 4 skills, example-skills with 12, and claude-api with 1, all with strict set to false and source './'. spec/agent-skills-spec.md is a single line pointing to agentskills.io; the spec text is not in the repository, and template/SKILL.md holds only a name, a description and one heading. skills/skill-creator/SKILL.md documents the three loading levels (metadata always in context, SKILL.md body on trigger, bundled resources on explicit read), advises keeping SKILL.md under 500 lines with a table of contents for reference files over 300 lines, and notes Claude tends to undertrigger, so descriptions should be a little bit pushy. skills/docx/scripts/comment.py writes the four comment parts plus relationships and content-type overrides, prints marker XML to paste into the document manually, parses with defusedxml.minidom, and XML-escapes comment text unless --raw is passed. skills/mcp-builder/scripts/evaluation.py runs a tool-use loop against a live server over stdio, SSE or HTTP with default model claude-3-7-sonnet-20250219, selects a single tool_use block per assistant response via next(...) so additional parallel tool calls are not executed, and grades the extracted response text by exact string equality against the qa_pair answer. skills/docx/SKILL.md, skills/pdf/SKILL.md and skills/xlsx/SKILL.md carry proprietary license frontmatter; skills/xlsx/SKILL.md names the six post-2007 functions needing the _xlfn. prefix.

Observed

License
Mixed: the word-processing, PDF and spreadsheet skills declare a proprietary license in their frontmatter, while the readme states many other skills are Apache 2.0
Interfaces
Installable as a Claude Code marketplace with three plugin bundles covering document skills, example skills and the Claude API skill
Specification
The in-repo spec file is a one-line pointer to agentskills.io; the specification text itself is not in the repository
Skill format
Only two frontmatter fields are required, name and description
Template
The bundled template contains only a name and description plus a single heading

Read from README.md, spec/agent-skills-spec.md, template/SKILL.md, .claude-plugin/marketplace.json, skills/skill-creator/SKILL.md, skills/docx/SKILL.md, skills/pdf/SKILL.md, skills/mcp-builder/SKILL.md, skills/webapp-testing/SKILL.md, skills/xlsx/SKILL.md, skills/algorithmic-art/SKILL.md, skills/docx/scripts/comment.py, skills/mcp-builder/scripts/evaluation.py.

Intel on Agent Skills

More in Intel

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.