ralph-loop-help — Claude Code skill: what it does, how to install

What ralph-loop-help does

The skill fires when you ask what Ralph Loop is, how it works, or how to use it. It carries a ready explanation of the Ralph Wiggum technique: one prompt goes to the agent over and over, and the self-referential part comes from the agent reading its own earlier work in files and git history rather than from feeding output back as input.

Each pass looks like this: the agent gets the same prompt, edits files, tries to exit, and a stop hook catches that exit and hands the prompt back. State lives in .cursor/ralph/scratchpad.md. Two things end the run — the agent printing a <promise> tag with the agreed phrase, or the --max-iterations cap being hit. With neither in place the loop keeps going, so set one before you start.

The technique suits tasks with a clear success criterion and greenfield work that needs refinement in steps. It gets in the way when a person has to make the design calls or the criterion is vague, because the agent has nothing to aim at. To stop a run early, ask the agent to cancel it; it removes the state file and reports how many iterations passed.