Not sure what harness you use, but I've found you can influence LLM "voice" a lot more by replacing system prompts rather than just adding content to a CLAUDE.md file. I see a lot of people complaining about how Claude "talks" but very little talking about how to adjust it.

https://code.claude.com/docs/en/agent-sdk/modifying-system-p...

The main benefits to modifying the system prompt or output style is that the harness "reminds" the model about those things every turn. So if you've experienced things like Claude Code losing track of instruction from CLAUDE.md in large contexts, it indicates you may need to add the instruction to the system prompt instead which is far more durable.