This is an autonomous iteration system that runs until an external verification command passes, not just until Claude says it's done. You give it a task and a verify command (like `npm test` or `curl`), and it loops indefinitely, tracking every attempt in a JSONL file so it won't repeat failed approaches. The hook runs your verify command independently after Claude claims completion. If verification fails, the loop continues with the rejection fed back in. After 3-4 rejections it forces reassessment, after more it makes you question the problem itself. It's basically autoresearch's oracle isolation pattern applied to general development work. You'd use this when you want unattended iteration on something with a clear pass/fail test, and you're comfortable letting it run overnight.
npx -y skills add tanweai/pua --skill pua-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