This implements Geoffrey Huntley's iterative bash loop methodology where each coding task runs in a fresh AI agent process to avoid context window bloat. You write specs with clear acceptance criteria, run the loop, and it works through them autonomously until each outputs a DONE signal. The clever bit is that state lives on disk between iterations (specs, history logs, plans) while the agent itself stays fresh. Best for multi-feature work where you want the AI grinding through tasks unattended. Requires YOLO mode and solid acceptance criteria or it'll spin. The setup includes a project interview that focuses on vision over tech stack, which is a nice touch.
npx -y skills add fstandhartinger/ralph-wiggum --skill ralph-wiggum --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