The official repo from the format's author: 19 skills plus spec and template

Why it is the reference, not just a repo

This is Anthropic's own skill implementation: 177,290 stars / 21,004 forks, 19 skills under ./skills, plus ./spec (the Agent Skills specification, now living at agentskills.io) and ./template (the scaffold for a new skill). If you want to know how a SKILL.md is actually supposed to be written, read this rather than any second-hand tutorial — the format's author is here.

The repo's own definition is worth memorising verbatim: folders of instructions, scripts, and resources that Claude loads dynamically. Progressive disclosure comes straight from that: SKILL.md carries routing information only, heavy content lives in references/ and scripts/ and gets read on demand.

Two licensing tiers, and this is the trap

Scope	Licence	Notes
skills/docx, skills/pdf, skills/pptx, skills/xlsx	source-available, not open source	The very code powering Claude's production file-creation capability, shared as a reference for complex skills
The other 15 (algorithmic-art, canvas-design, frontend-design, mcp-builder, webapp-testing, theme-factory, brand-guidelines, internal-comms, doc-coauthoring, claude-api, skill-creator, web-artifacts-builder, slack-gif-creator, academy-guide, discernment-nudge)	Apache-2.0	Forkable into your own product

There is also an explicit disclaimer: these skills are for demonstration and education only, the implementations and behaviours you get from Claude may differ, and you must test in your own environment before relying on them for critical tasks. Read that line before treating the official repo as a production dependency.

Use pptx as the teaching case: what a production skill looks like

skills/pptx is the most engineered of the four document skills, and it splits three operations across entirely different technical paths:

• Create via pptxgenjs;
• Edit via unzip → modify XML → zip, bypassing any render engine so masters and themes survive;
• Read via markitdown, which extracts content to Markdown.

Supporting scripts are thumbnail.py, add_slide.py, clean.py, validate.py, plus a soffice.py wrapper around LibreOffice. It even encodes pptxgenjs gotchas into the instructions, and these were clearly paid for:

- pptxgenjs mutates the option object in place → never share one across add* calls
- a combo series using secondaryValAxis needs BOTH valAxes and catAxes with 2 entries each,
  otherwise PowerPoint treats the chart as corrupt and discards it
- never reorder the children of <p:presentation>

That is the answer to "what belongs in a skill": not general knowledge, but the traps the model walks into on its own and that fail silently. The second one is the archetype — the chart vanishes in PowerPoint and the agent has no idea it broke anything.

All three entry points work

# Claude Code: register the repo as a plugin marketplace first
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

After install, skills fire on mention, e.g. "Use the PDF skill to extract the form fields from path/to/some-file.pdf". The other two routes: these example skills are already available on paid Claude.ai plans (and you can upload custom ones), or mount them programmatically through the Skills API.

The highest-value way to use this repo

Treat template/ plus skill-creator as scaffolding, read pptx/xlsx end to end as the model answer for a complex skill, and only then write your own. The seven production skills behind this site follow exactly that shape: SKILL.md holds the iron laws and the process state machine, heavy lifting goes to scripts/, and specification knowledge sinks into the body — instead of everything being crammed into one prompt.