create-skill

/create-skill — Skill Builder

First: Read LEARNINGS.md in this skill's directory.

Arguments: $ARGUMENTS

Process

Step 1: Understand

If $ARGUMENTS provides a skill name and description, proceed. Otherwise ask:

1. What should this skill do? (one sentence)
2. When should it trigger? (what would the user say/do?)
3. What resources does it need? (scripts, reference docs, templates?)

Keep questions to one message. Don't over-interview.

Step 2: Scaffold

Run the init script:

python3 .claude/skills/create-skill/scripts/init_skill.py <skill-name> --path .claude/skills

This creates the directory with template files. If --path is omitted, defaults to .claude/skills in the current workspace.

Step 3: Write SKILL.md

Replace the template content. Follow these rules:

Frontmatter (required):

• name: kebab-case, max 64 chars. Becomes the /slash-command.
• description: Comprehensive. This is how Claude decides when to activate. Include trigger phrases, contexts, use cases.

Frontmatter (optional, add only when needed):

• allowed-tools: Restrict tool access (e.g., Read, Grep, Glob, Bash)
• argument-hint: Autocomplete hint (e.g., [filename], [--dry-run])
• fork: true: Run as sub-agent (use for data-heavy skills that would bloat main context)
• disable-model-invocation: true: Manual /name only, no auto-trigger

Body rules:

• Start with "Read LEARNINGS.md" instruction
• Use $ARGUMENTS placeholder for user input
• Imperative form ("Run the script", not "You should run the script")
• Keep under 500 lines. Split longer content into references/ files.
• Reference bundled resources with relative paths + describe when to read them
• Match freedom to fragility: exact scripts for brittle ops, text guidance for flexible tasks

Body structure (adapt as needed):

# /name -- One-Line Summary

**First:** Read `LEARNINGS.md` in this skill's directory.
**Arguments:** `$ARGUMENTS`

## Step 1: [Name]
[What to do, expected outcome]

## Step 2: [Name]
[What to do, expected outcome]

## Rules
[Constraints, gotchas, things to never do]