snarktank/ralph ? reverse-engineered prompt

I want to create an autonomous AI agent that can manage a coding project. The idea is to have a simple script that acts as a loop. This script will repeatedly call an AI coding tool like Amp or Claude Code to work on a list of tasks until they are all finished.

The system should work from a product requirements doc (PRD). I'll need a way to take a standard markdown PRD and convert it into a prd.json file that the agent can use as its to-do list. Each item in the list will have a status, like passes: false.

The main agent loop should then pick the highest-priority task that isn't done, implement it, run quality checks, and commit the code if it works. After a successful run, it should update the prd.json to mark the task as complete and also append any learnings to a progress.txt file for a simple form of memory. The whole process repeats until all tasks in the list are done.