Learning from Claude Skill Examples: anthropics/skills Patterns

If you want to build good skills, the fastest path is to read Anthropic's published official example skills. The anthropics/skills repo holds patterns that are actually used in production. Rather than explaining concepts, this guide distills practical patterns and authoring principles you can learn from the examples, based on official material. (As of July 2026, re-checked September 2026; re-checked September 2026.)

What official example skills exist

The repo's example skills (mostly Apache 2.0 open source) span creative, technical, and work tasks. Representative ones by category:

• Dev & technical: webapp-testing (verify local web-app UI via Playwright), mcp-builder (generate MCP servers), frontend-design
• Document creation: pdf, docx, xlsx, pptx — the skills that actually power Claude's document creation, under the repo's document-skills/.
• Creative & design: canvas-design, algorithmic-art, theme-factory
• Work & collaboration: doc-coauthoring, internal-comms, brand-guidelines, plus the meta-skill skill-creator

When starting out, use the repo's template-skill as a starting point.

Pattern 1 — core in SKILL.md, details in references/

Well-built example skills keep SKILL.md lean with just core instructions. Detailed docs and schemas go in a references/ directory, linked from the main file. For instance, skill-creator moves JSON structures for evals/grading into references/schemas.md. This makes progressive disclosure work properly — normally only the short core loads, and details enter only when needed. Executable scripts and resources are bundled in scripts/ and assets/ (optional).

Pattern 2 — design the workflow as stages (doc-coauthoring example)

A good skill guides a task through clear stages. The doc-coauthoring skill runs collaborative writing in three: ① gather context (Claude asks questions to capture background, scope, and tone) → ② refine & structure (iterate brainstorm/edit per section) → ③ reader testing (validate the doc with a fresh, context-free Claude to find blind spots obvious to authors but confusing to others). When sub-agents are available, it poses 5–10 likely reader questions to a fresh Claude and checks automatically. Baking a "stage → test → iterate" structure into a skill stabilizes output quality.

Authoring principles from the examples

The principles running through the example skills:

• Clear, slightly "pushy" description: skill-creator recommends an assertive description that makes when to use it obvious, rather than a vague one ("a dashboard that shows internal data"). The description isn't human documentation — it's the key signal Claude uses to pick a skill.
• Core in SKILL.md, details in references/
• Auto-activate when the trigger matches (direct invocation also works)
• Tell users where output files are saved: for files users must inspect (.docx, .xlsx), state the save location

Related reading

The basics and SKILL.md structure: Claude Code skills guide; step-by-step building: create a custom skill; use cases: skills by use case.

Related articles

• Claude Code Skills (Agent Skills) Guide: Extend Claude with SKILL.md
• Claude Code Skill Not Recognized — Troubleshooting Guide
• Claude Code Installation Guide: native vs npm, Windows Included
• Claude Code vs Cursor: Key Differences and Which to Choose
• How to Use Claude Code: Real Usage After Install (Commands, VS Code)