Ralph Won't Stop Until Your Spec Is Shipped
Ralph is an autonomous agent loop that does one thing: run your AI coding tool over and over until every item in your PRD is done. You give it a product requirements doc, it spawns Claude Code or Amp, completes what it can, commits, and starts a fresh session for the next task. No human babysitting required.
The clever part is how it handles context limits. Each iteration is a clean slate with full context budget, but memory persists through git history and a progress.txt file. After every cycle, Ralph writes what it learned to AGENTS.md, which the next iteration reads automatically. It's self-teaching across sessions, which means iteration 10 is meaningfully smarter than iteration 1.
15,900 GitHub stars tell you developers want this. The pattern is dead simple but the execution matters: PRD generation from natural language, automatic story decomposition, learning accumulation. You describe what you want built, Ralph figures out the user stories and grinds through them one by one.
Works with Amp by default, Claude Code with a flag. Ten iterations per run, configurable. The whole thing is a shell script and some skill files. Based on Geoffrey Huntley's Ralph pattern. Sometimes the best tools are the simplest ones.
https://github.com/snarktank/ralph
← Back to all articles
The clever part is how it handles context limits. Each iteration is a clean slate with full context budget, but memory persists through git history and a progress.txt file. After every cycle, Ralph writes what it learned to AGENTS.md, which the next iteration reads automatically. It's self-teaching across sessions, which means iteration 10 is meaningfully smarter than iteration 1.
15,900 GitHub stars tell you developers want this. The pattern is dead simple but the execution matters: PRD generation from natural language, automatic story decomposition, learning accumulation. You describe what you want built, Ralph figures out the user stories and grinds through them one by one.
Works with Amp by default, Claude Code with a flag. Ten iterations per run, configurable. The whole thing is a shell script and some skill files. Based on Geoffrey Huntley's Ralph pattern. Sometimes the best tools are the simplest ones.
https://github.com/snarktank/ralph
Comments