In my experience, no amount of output style customization gets it anywhere near the output I'd like.

I've had more success running a prose linter[1] with a ruleset[2] that gets rid of most of the egregious stuff. You can also pretty easily write custom rules for it.

And sometimes, when I'm fed up enough, I just throw all the prose in the codebase at Antigravity/Gemini to rephrase. The result is so much more readable.

[1]: https://vale.sh/

[2]: https://github.com/tbhb/vale-ai-tells

You can change it ... and Claude will maybe pay attention to what you changed 25% of the time.

"You can tell an AI model to write in a different style" and "the AI model will actually write in the style you tell it to" are two different things that often don't intersect