How I'm running an autonomous Claude Code loop on a real shipped product

An indie developer describes using an autonomous Claude Code loop to handle routine work on a live Cloudflare Workers game without letting the agent touch production.

The workflow follows a MAPE-K pattern: read repo instructions and history, pick one prioritized task, plan carefully, test first, then open PRs against dev and auto-merge only on green CI.

To keep the product safe, the setup isolates the agent in a dev branch, records outcomes in a changelog for future runs, and gives extra caution to security-sensitive paths like leaderboard tamper resistance.

The post also shares implementation notes from the Balatro-style typing arcade, including PixiJS scene management in React and the stack around TanStack Start, Drizzle, Neon Postgres, Clerk, and React 19.