IIUC this doesn't make the LLM think in caveman (thinking tokens). It just makes the final output show in caveman.