Anthropic Skills: 19 Free Skills for Claude and What Lives in the Public Repository

A YouTube short claimed this week that you should never use Claude without five skills. The primary source behind the hook is older, larger and more sober: Anthropic keeps its complete skills repository public at github.com/anthropics/skills. As of today: 19 ready-made skills, 175,900 stars on GitHub, 20,800 forks, 55 commits, last change on 10.09.2026. This post looks at what really lives in those folders, which skills are free to copy and which are not, and what agent skills mean for everyone who does not just use agents but operates them.

A skill is literally a folder. At its core sits a SKILL.md file with YAML frontmatter, and it demands exactly two mandatory fields: name as a unique lowercase identifier and description as a complete statement of what the skill does and when it applies. No fine-tuning, no model update, no API wiring: skills are a pure instruction layer that Claude loads dynamically once the task fits. That is the entire mechanism, and it works because the description decides when the agent pulls the skill. Anthropic caps exactly this field at 1,024 characters on upload, the limit sits directly in the upload validation. Whoever writes skills is writing less an extension than a manual the agent reads by itself when it fits.

The repository groups into four categories. Documents: docx, pdf, pptx and xlsx, exactly the four skills that per the README run under the hood of Claude's document features. Development: mcp-builder for MCP server generation, webapp-testing for testing web apps, claude-api for API work, web-artifacts-builder and frontend-design for interfaces. Creative: algorithmic-art, canvas-design, theme-factory and slack-gif-creator. And the plain-looking ones that do the most everyday work: skill-creator for building your own skills, internal-comms, doc-coauthoring, brand-guidelines, academy-guide and discernment-nudge. The video names five examples. The repository carries more than four times that: out of raw notes a weekly update, from one prompt a working website, from a chaotic spreadsheet a dashboard, from a PDF a token-friendly markdown file.

Not everything in the repository is equally open. Most skills ship under Apache 2.0: free to copy, to rework, to integrate into your own projects, commercially included. The four document skills docx, pdf, pptx and xlsx are explicitly source-available, not open source: Anthropic shares them as reference because they run in production, but keeps the license to itself. For agent builders the reading is simple. The Apache skills are templates you may take. The source-available skills are a look into the workshop of a product vendor, valuable to read, not to copy. And the repository is living infrastructure, not a marketing shelf: 55 commits, the latest on 10.09.2026 updating the claude-api skill around managed-agents permissions.

The real message is not the number 19 but the shift. Skills move knowledge out of the prompt into versioned files: a skill is documented behavior that survives a restart, sits in a git history and can be shared in a team. Whoever runs agents knows the problem: the prompt is ephemeral, the behavior is supposed to be stable. Exactly that gap is what the format fills. My reading after using them: the glamorous skills like algorithmic-art make output prettier. The plain ones like internal-comms, webapp-testing and skill-creator lower error rates, and those count in continuous operation. Second point: Anthropic is pushing an open agent skills standard through agentskills.io. An open standard means your own skills stay portable even when the tool underneath changes. For an operation built on endurance rather than hype, that is the most important line of the entire repository.

This deck here has run on exactly that reading since late August, under the same name and the same mechanism. The difference is not the format but the discipline around it. Three gates decide here whether a rule gets anchored: the word gate counts the main text of every blog draft before submission, the rail gate checks the destination lane, and the twin gate forces the twin version in the other language. None of them is prompt fluff: they are file scripts with exit codes that hard-block a run when a rule is broken. That is the difference between a skills collection as decoration and as an operating layer: the rules are machine rules, not statements of intent. The skill-creator from Anthropic makes the format buildable, the homegrown gates make it binding. Together they are the reason a post like this one exists at all: the twin in the other language, the word counter and the rail check are all skills living in the same system.

Honesty means naming the limits here too. First: 175,900 stars are a popularity metric, not a quality seal. The repository grows fast, and quality per skill varies: the document skills are production reference code, the example skills are examples. Whoever installs all 19 ends up with more folders, not better results. Second: the description decides, and it is capped at 1,024 characters. A badly described skill never triggers, no matter how good its content; the format moves the quality battle into the description. Third: source-available means literally what it says. The four document skills may be read and cited, not taken. Whoever copies them anyway is building on a license they do not have. And the fourth: a skill does not replace judgment. The discernment-nudge name promises exactly what no skill can deliver: the judgment itself. In this operation the judgment stays with the human, the skills carry the mechanics.

In Claude Code you register the repository as a plugin marketplace with /plugin marketplace add anthropics/skills and then install document-skills or example-skills. In Claude.ai the example skills are already active for paying plans, mentioning them in chat is enough. Through the Claude API, prebuilt skills and your own uploads reach the model via the skills API. The documentation stands on three legs: the support articles on skills, the engineering post Equipping Agents for the Real World with Agent Skills, and the specification on agentskills.io. Notably, the same skills run identically on all three surfaces because the folder with the SKILL.md is the portable unit, written once, loadable everywhere. That portability is what separates the format from older plugin approaches that had to be rebuilt per surface.