snarktank/ralph

Ralph is an automation loop that runs AI coding tools, specifically Amp or Claude Code, repeatedly in a batch until an entire list of software tasks is finished. Instead of manually prompting an AI assistant one task at a time, you write a Product Requirements Document (a structured list of features and user stories), convert it into a JSON task file called prd.json, and then run Ralph. It picks the next unfinished story, launches a fresh AI coding session to implement it, runs quality checks (type checking and tests), commits the result to Git, marks the story as done, and then repeats, up to a configurable number of iterations. Each iteration starts with a clean context window (no memory of previous sessions), so Ralph keeps state using three simple files: the git commit history, a progress.txt file where each run appends learnings, and the prd.json file that tracks which stories pass. This design means even large features, spanning many context windows, can be completed autonomously. You would use Ralph if you want to automate a series of well-defined coding tasks end-to-end, stepping back while the AI works through a backlog. It requires either the Amp CLI or Claude Code installed, plus jq and a git repository. The project is written in TypeScript and shell script.