This is a state machine wrapper around specification-driven development that breaks task implementation into single-step iterations to avoid context window explosion. Instead of running everything at once, you invoke a Python script that shows you the next command to execute (implementation, review, cleanup, sync), you run it manually, then loop back. State persists in fix_plan.json with transitions like init → choose_task → implementation → review → fix → cleanup. Supports task range filtering with --from-task and --to-task flags. The ralph-loop name comes from applying Geoffrey Huntley's "Ralph Wiggum as a Software Engineer" technique, which apparently means doing one dumb thing at a time until you're done. Honest take: this is workflow automation for people who keep blowing up their context window trying to do too much in one shot.
npx -y skills add giuseppe-trisciuoglio/developer-kit --skill ralph-loop --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills