How it worked:
1. Asked Claude to interview me about the game design 2. From that conversation, it generated ~248 structured tasks as JSON 3. Created a simple loop: Claude picks a task, implements, tests, commits, repeats 4. Throttled to 3 tasks, 15min pause, to not kill my API quota The script: claude --permission-mode acceptEdits "@specs/prd-v1.json \ 1. Read PRD. Pick highest-priority task. \ 2. Implement. Run tests. \ 3. Update progress. Commit. Push. \ ONLY ONE TASK AT A TIME." Result: https://jarls-production.up.railway.app Code: https://github.com/fcavalcantirj/jarls The loop concept: https://www.anthropic.com/engineering/effective-harnesses-fo...
Not perfect — had bugs, ugly screens. But 95% was autonomous.
The bottleneck isn't coding anymore. It's thinking clearly upfront.