Runs Claude in a loop until your task is actually done, either in headless mode spawning child sessions or using hooks to intercept session exit. You get two strategies: structured mode breaks down complex projects into a phased task list, while lightweight mode just repeats the same prompt Ralph-style until tests pass or bugs are fixed. It tracks progress across sessions, detects completion via promise tags or checkbox counting, and includes budget controls so you don't accidentally burn through API credits overnight. The hook mode is clever for avoiding nested session issues. Honestly most useful for the kind of grunt work you'd otherwise babysit across multiple manual sessions.
npx skills add https://github.com/feiskyer/claude-code-settings --skill autonomous-skill