I don't think that is controversial.

My system CLAUDE.MD spells out my preferences in stack and architecture, and how I expect the agent to interact with me.

I have it generate a summary doc in large repos so it doesn't need to slurp in the world to review a PR or work on a corner of it.