setup-ralph_skill

Overview

This skill sets up Geoffrey Huntley’s Ralph Wiggum autonomous coding loop in any directory, creating the proper file structure, prompts, and safety/backpressure mechanisms. It provides commands and templates to initialize the loop, enable automated commits and optional GitHub backups, and enforce tests and safety rules. The goal is to let the agent iterate reliably while you observe and tune the environment.

How this skill works

The setup creates a deterministic loop: a planning prompt produces a prioritized TODO list, a building prompt implements one task and runs validation, and the loop restarts with fresh context. It installs project files (PROMPT_plan.md, PROMPT_build.md, plan state, validation scripts), example CI-style tests for backpressure, and optional GitHub push hooks via the gh CLI. Safety rules and isolated test execution are included to prevent destructive operations.

When to use it

• Initialize a new autonomous Claude/Anthropic loop in a repository or directory
• Try Ralph when you want continuous, single-task iterations with minimal context bleed
• Add deterministic backpressure via tests, linters, and typechecks
• Create a safe sandbox for agent-driven development with remote backups
• Prototype autonomous coding workflows before scaling to multiple agents

Best practices

• Use one task per loop to maximize context efficiency and keep prompts focused
• Provide strong backpressure: failing tests should block commits and force rework
• Keep PROMPT_plan.md and PROMPT_build.md separate; plan must not implement code
• Run tests in temporary isolated directories to limit side effects
• Enable automatic GitHub push only after confirming gh CLI is authenticated; set RALPH_BACKUP=false to disable

Example use cases

• Bootstrap a new project directory with Ralph structure and starter prompts
• Convert an existing repo into an autonomous loop that iterates on one task at a time
• Experiment with different backpressure strategies (unit tests, builds, linters) to guide agent behavior
• Customize build prompts to enforce company coding standards and safety rules
• Use the loop as a reproducible environment for prompt engineering and failure-mode observation