skill-coach

Skill Coach

When to Use

Use this skill when the user asks to:

• Create a new agent skill
• Edit or update an existing skill
• Generate skill content with AI assistance
• Distribute skills to their configured agents
• Check the status of their skills

Instructions

Creating a New Skill

1. Ask the user what the skill should do and which agents it targets
2. Run coach init skill to scaffold the skill directory
3. If the user wants AI-assisted authoring, run coach generate <skill-name> instead — this scaffolds and authors in one step
4. After creation, run coach lint <path> to validate

Editing a Skill

1. Run coach edit <skill-name> to open in the user's editor
2. Coach will automatically validate on save
3. For AI-assisted editing, use coach generate <skill-name> --prompt "description of changes"

Distributing Skills

1. Ensure distribution targets are configured: coach config get distribute-to
2. If not configured, help the user set them: coach config set distribute-to claude,cursor
3. Run coach sync to symlink all managed skills to agent directories

Skill Format Reference

A valid SKILL.md has YAML frontmatter and a markdown body:

• name (required): lowercase alphanumeric with hyphens, max 64 characters
• description (required): what the skill does, max 1024 characters
• license (optional): MIT, Apache-2.0, or ISC
• allowed-tools (optional): list of tools the skill needs access to

The body should include:

• When to Use: trigger conditions — when should an agent activate this skill?
• Instructions: step-by-step guidance for the agent
• Constraints: what the agent should NOT do

Security

• Never include secrets, API keys, or credentials in skills
• Never include destructive commands without explicit user confirmation
• Always scope allowed-tools to the minimum needed
• Run coach scan <path> to check for security issues before distributing

Constraints

• Do not modify skills outside of coach-managed directories
• Always run coach lint before distributing a skill
• If coach lint reports issues, fix them before proceeding
• Do not distribute skills that fail security scanning with High or Critical findings