Great writeup! One observation I've had lately is that one could spend hours crafting the perfect CLAUDE.md or Agents.md, but this only controls how the agent writes the code. It doesn't control how much context the backend dumps into the conversation on every tool call.