AgentSkillsCN

Ralph

拉尔夫

SKILL.md

Ralph

Run Ralph - the autonomous AI coding agent that plans and builds features iteratively.

How to Use

When this skill is invoked, use run execute to call Ralph:

code
run execute "plan ./plans/my-feature.md"
run execute "build ./plans/my-feature.md"
run execute "build ./plans/my-feature.md 20"

Commands

Plan Mode

Analyzes codebase and creates a task list, then exits:

  • plan <plan-file> - Create task list for a plan file
  • plan ./plans/feature.md - Example with relative path

Build Mode

Implements tasks iteratively until completion:

  • build <plan-file> - Build until RALPH_DONE marker
  • build <plan-file> <max-iterations> - Build with iteration limit

Utility Commands

  • status - Show Ralph version and check notification setup
  • test notifications - Test notification configuration
  • help - Show Ralph usage information

Examples

code
# Plan a new feature (analyzes and creates task list)
run execute "plan ./plans/add-dark-mode.md"

# Build a feature until complete
run execute "build ./plans/add-dark-mode.md"

# Build with max 20 iterations
run execute "build ./plans/add-dark-mode.md 20"

# Check Ralph version
run execute "status"

How Ralph Works

  1. Plan Mode: Runs once, creates <plan-name>_PROGRESS.md with task list
  2. Build Mode: Iterates until RALPH_DONE appears in progress file
  3. Progress Tracking: Updates progress file after each iteration
  4. Notifications: Optional Slack/Discord/Telegram alerts (configure via ~/.ralph.env)

Progress File

Ralph creates a progress file named <plan-name>_PROGRESS.md that tracks:

  • Current status (IN_PROGRESS or RALPH_DONE)
  • Task list with completion status
  • Notes and blockers

Notes

  • Ralph uses Claude CLI in pipe mode with Sonnet model
  • Plan files should be markdown specs describing the feature
  • Progress files are created in the current working directory
  • Use --test-notify to verify notification setup before long builds