Aren't LLMs trained on and optimized for human code? In general, anything that is concise is more effective as context for the LLM, whether it be your CLAUDE.md or code, since LLMs are meant to model human language.
Humans can share context outside the code, while LLMs need it to be more explicitly structured inside. And if we keep the slightly verbose good practices that already help humans understand the intent, we kind of get the best of both worlds.