llm Steps

Make a single call to Claude.

Basic usage

steps:
  - llm:
      system: You are a helpful assistant.
      prompt: "What is the capital of France?"

Options

Option	Type	Default	Description
system	string	-	System prompt (instructions for Claude)
prompt	string	required	User prompt (the question/task)
silent	boolean	false	Capture response without printing

System prompt

The system field sets Claude's behavior and context:

steps:
  - llm:
      system: |
        You are a senior code reviewer.
        Focus on security issues and performance.
        Be concise and actionable.
      prompt: "Review this code: {{output}}"

Tips for system prompts:

• Be specific about the task
• Specify the output format you want
• Mention constraints (length, style, etc.)

Using with exec

A common pattern is to capture command output, then send it to Claude:

commit:
  description: Generate a commit message
  steps:
    - exec:
        command: git diff --cached
        silent: true
    - llm:
        system: |
          Generate a concise git commit message for these changes.
          Use conventional commits format (feat:, fix:, docs:, etc).
          Output only the commit message, nothing else.
        prompt: "{{output}}"

Silent mode

When silent: true, the response is captured but not printed:

steps:
  - llm:
      system: Generate a shell command to list large files.
      prompt: "Find files over 100MB"
      silent: true
  - exec:
      command: "{{output}}"
      confirm: true

This is useful when the LLM generates a command that you want to execute.

Variable interpolation

All template variables work in both system and prompt:

steps:
  - llm:
      system: |
        You are helping a user on {{os}}.
        Current directory: {{directory}}
        Current time: {{datetime}}
      prompt: "{{args.question}}"

Multi-step conversations

Each llm step is independent - there's no conversation history between steps. If you need multi-turn conversation, use agentic mode instead.

For sequential LLM calls, pass context explicitly:

steps:
  - id: analysis
    llm:
      system: Analyze this code for bugs.
      prompt: "{{output}}"
      silent: true
  - llm:
      system: Suggest fixes for these bugs.
      prompt: "{{steps.analysis.output}}"

Output as step result

The LLM's response becomes the step's output, available as {{output}} in the next step or {{steps.id.output}} if the step has an id.