Is there a paper describing how @bcherny trained Claude code 🤯