aronprins/claude-loop ? reverse-engineered prompt

Build me a Claude Code skill that can take a claude-loop/prd.json file and turn it into working code story by story.

I want one fresh subagent for each user story, with the main session only coordinating, checking results, and moving on to the next story. Each story should be implemented, tested, committed, and then marked done in the PRD. Keep a running progress.txt log for anything important learned along the way, and use AGENTS.md files in the repo so future runs can pick up useful project knowledge automatically.

Please make it work as a self contained workflow inside the repo, with the runtime state living in a top level claude-loop/ folder, and archive finished runs when everything passes. If it helps, look up the current Claude Code docs online so the setup matches how the Task tool and skills work today.