Example skill

Example Skill

This is the skill body. Everything below the frontmatter is the instruction set Claude follows once the skill is invoked. Write it the way you would brief a sharp new teammate: concrete, ordered, and unambiguous.

Delete this whole file's contents and write your own. The notes below are a guide to the format, not content to keep.

How a skill is invoked

• Automatically, when the user's request matches your description.
• Manually, by typing /<plugin-name>:<skill-name>. For this template that would be /example-skill:example-skill once installed.

Plugin skills are always namespaced by the plugin name, so pick names that read well together.

How to structure the body

Use plain Markdown headings. A reliable shape:

1. Trigger phrases: restate the exact phrases that should fire the skill, so behavior is predictable.
2. Workflow: numbered steps. Tell Claude what to read, what to compute, and in what order. Be explicit about parallel vs. sequential work.
3. Output format: show the exact shape of the result. If it is long, put it in a references/ file (see below) and point to it.
4. Rules / guardrails: what to never do, what to confirm before doing, how to handle missing data.
5. Examples: a few "user says X" to "skill does Y" lines remove ambiguity faster than paragraphs of prose.

Bundling supporting files

A skill is a folder, not just one file. You can ship:

example-skill/
├── SKILL.md            # this file, the entry point
├── references/         # longer docs the skill reads at runtime
│   └── output-format.md
└── scripts/            # helper scripts the skill can run
    └── do-thing.sh

Reference bundled docs by their relative path (e.g. "see references/output-format.md") and Claude will read them. If a step runs a bundled script or executable, reference it with the ${CLAUDE_PLUGIN_ROOT} variable (e.g. ${CLAUDE_PLUGIN_ROOT}/scripts/do-thing.sh) so the path still resolves after the plugin is installed into Claude Code's cache. Never use ../ to reach outside the skill folder.

Reading the user's own data

If your skill needs the user's files or live data, do NOT bake real data into the skill. Instead:

• Read from files the user keeps in their repo (document the expected paths).
• Pull live data from MCP servers the user connects themselves.
• Ship only fictional placeholders under an examples/ folder, and write a SETUP.md telling the user how to wire up their own.

See plugins/daily-brief/ in this repo for a full worked example of that pattern.

Checklist before you publish

• description is specific and names trigger phrases.
• Folder renamed; name in frontmatter matches the folder.
• .claude-plugin/plugin.json updated (name, description, keywords).
• No real data, emails, names, or credentials anywhere.
• Registered in .claude-plugin/marketplace.json.
• Installed and invoked once locally to confirm it loads.