Not a bad idea - however

> Caveman only affects output tokens — thinking/reasoning tokens are untouched.

The problem is the thinking. But could help to tune my system prompt for Kimi.