Agent-Skills.md

Agent Skills: Ralph Loop Patterns

|

UncategorizedID: phrazzld/claude-config/ralph-patterns

Repository

phrazzld/claude-config
phrazzld
31

Install this agent skill to your local

pnpm dlx add-skill https://github.com/phrazzld/claude-config/tree/HEAD/skills/ralph-patterns

Skill Files

Browse the full folder contents for ralph-patterns.

Download Skill

Loading file tree…

skills/ralph-patterns/SKILL.md

Skill Metadata

Name
ralph-patterns
Description
|

Ralph Loop Patterns

You're running in a Ralph loop - an autonomous issue-to-merged-PR workflow.

Completion Signals

Output these exact XML tags for the shell script to detect:

  • <promise>DONE</promise> - Task complete, exit successfully
  • <promise>BLOCKED</promise> - Cannot proceed, needs human intervention

Always include the exact XML tag. The shell script greps for it.

Progress Tracking

Ralph monitors these for "no progress" circuit breaker:

  • Git HEAD (commits)
  • PR state (reviews, CI status)

Make commits frequently to show progress. Avoid long stretches without commits.

Within a Ralph Loop

You have full access to:

  • /spec, /architect, /build, /refactor, /update-docs, /pr
  • Moonbridge MCP for Codex/Kimi delegation
  • Sentry, PostHog, GitHub MCPs
  • All configured hooks and skills

Delegation Pattern

Use Moonbridge for implementation:

mcp__moonbridge__spawn_agent({
  "prompt": "Implement [feature]. Follow pattern in [ref file].",
  "adapter": "codex",
  "reasoning_effort": "high"
})

Handling CI Failures

  1. Get CI logs: gh run view --log-failed
  2. Analyze failures
  3. Fix and push
  4. Wait for re-run (CI usually triggers automatically)

Handling Review Feedback

  1. Get PR comments: gh pr view --comments
  2. Identify critical vs stylistic feedback
  3. Address critical feedback first
  4. Push changes
  5. Respond to reviewers via commit message or PR comment

When to Output BLOCKED

Output <promise>BLOCKED</promise> when:

  • Missing required credentials or API keys
  • Ambiguous requirements that need human clarification
  • Conflicting constraints with no clear resolution
  • External dependency is down/unavailable
  • Permissions issues you can't resolve

Include a brief explanation after the tag:

<promise>BLOCKED</promise>
Reason: Need clarification on whether user auth should use JWT or session cookies.

Anti-Patterns

  • Don't loop infinitely on the same failing test
  • Don't make empty commits just to show progress
  • Don't ignore reviewer feedback hoping it goes away
  • Don't skip /codify-learning at the end (Ralph does this automatically)