Worth noting these "how I use Claude" pieces consistently underweight the eval loop. Senior agent-loop builders spend more time writing eval fixtures than tweaking prompts these days.