Prompts

A prompt is a saved prompt template within a bundle. Prompts help standardize common AI interactions and are automatically exposed as skills (slash commands) in Claude Code.

Prompt Structure

Prompts are defined within bundles:

prompts:
  code-review:
    description: "Review code for best practices"
    content: |
      Review this code for adherence to best practices.

      Focus on:
      - Error handling
      - Type annotations
      - Documentation
      - Performance considerations

  refactor:
    description: "Suggest refactoring improvements"
    content: |
      Analyze this code and suggest refactoring improvements.

      Consider:
      - SOLID principles
      - Code clarity
      - Testability

Claude Code Skills Integration

Prompts are automatically exposed as Claude Code skills. When you define a prompt, it becomes available as a slash command:

# In Claude Code, use the prompt as a skill:
/code-review
/refactor

Skill Configuration

Control how prompts appear as skills:

prompts:
  code-review:
    description: "Review code for best practices"
    content: |
      Review this code...
    plugins:
      llm:
        claude-code:
          enabled: true              # Default: true (opt-out model)
          description: "Review code" # Shown in /help
          argument_hint: "<file>"    # Autocomplete hint
          allowed_tools:             # Restrict available tools
            - Read
            - Grep
          model: "sonnet"            # Override model for this skill

Skill Fields

Field	Default	Description
enabled	true	Set to false to hide from skills
description	prompt description	Short description for /help
argument_hint	none	Hint shown during autocomplete
allowed_tools	all	Restrict which tools the skill can use
model	default	Override the model (sonnet, opus, haiku)

Disabling a Skill

To keep a prompt but not expose it as a skill:

prompts:
  internal-prompt:
    description: "Internal use only"
    content: |
      This prompt is used programmatically, not as a skill.
    plugins:
      llm:
        claude-code:
          enabled: false

Using Prompts

As Claude Code Skills

# In Claude Code session, just use the slash command:
/code-review
/refactor

# With arguments:
/code-review src/main.go

Via CLI

# Run a prompt by name
scm run --saved-prompt code-review

# Include with prompt reference
scm run -f my-bundle#prompts/code-review

List Available Prompts

# List all prompts
scm prompt list

# Show prompt details
scm prompt show my-bundle#prompts/code-review

Editing Prompts

# Edit prompt content in your editor
scm prompt edit my-bundle#prompts/code-review

Prompts vs Fragments

Aspect	Fragments	Prompts
Purpose	Context/instructions	Specific actions/requests
Usage	Combined with user input	Standalone skills or combined
Typical content	Guidelines, patterns, standards	Review requests, generation tasks
Claude Code	Injected as context	Exposed as slash commands

Fragments provide context that's always available. Prompts provide specific actions you invoke when needed.

Using Together

# Fragment provides context, prompt defines the action
scm run -f python-standards --saved-prompt code-review

# In Claude Code with skills:
# 1. Context from fragments is already injected
# 2. Just invoke the skill:
/code-review

Examples

Code Review Skill

prompts:
  review:
    description: "Comprehensive code review"
    tags: [review, quality]
    content: |
      Perform a comprehensive code review:

      1. **Correctness**: Logic errors, edge cases
      2. **Security**: OWASP top 10, input validation
      3. **Performance**: N+1 queries, unnecessary allocations
      4. **Maintainability**: Naming, complexity, documentation
      5. **Testing**: Coverage gaps, test quality

      Provide specific line references and suggested fixes.
    plugins:
      llm:
        claude-code:
          description: "Comprehensive code review"
          argument_hint: "<file or directory>"

Test Generator Skill

prompts:
  gen-tests:
    description: "Generate unit tests"
    content: |
      Generate comprehensive unit tests for the specified code.

      Requirements:
      - Use table-driven tests where appropriate
      - Cover happy path and error cases
      - Mock external dependencies
      - Include edge cases
    plugins:
      llm:
        claude-code:
          description: "Generate unit tests"
          argument_hint: "<function or file>"
          allowed_tools:
            - Read
            - Write
            - Grep

Documentation Skill

prompts:
  document:
    description: "Generate documentation"
    content: |
      Generate documentation for the specified code:

      - Function/method signatures with descriptions
      - Parameter explanations
      - Return value descriptions
      - Usage examples
      - Error conditions
    plugins:
      llm:
        claude-code:
          description: "Generate docs"
          model: "haiku"  # Use faster model for docs