This is autonomous loop mode for Claude Code that keeps running the same prompt until your task is done. You set up a state file with clear completion criteria (usually a TODO.md checklist), and it iterates until it hits ALL_TASKS_COMPLETE or your max iteration cap. Named after Ralph Wiggum because it never stops despite confusion. Best for well-defined work like migrations, refactoring, or getting tests to pass. Not for exploratory work where you need judgment calls. The state file tracks iterations and you get checkpoint reminders every 5 loops to review progress. Basically supervised autonomy: you can let it run overnight on clear tasks, but you're meant to check in periodically. Includes safety caps and git tracking so nothing goes off the rails permanently.
npx skills add https://github.com/belumume/claude-skills --skill ralph-loop