Thank you. Models are clumsier than I expected in mafia games. Despite my clear instructions to 1) limit max output tokens, 2) avoid thinking out loud, and 3) use the <think></think> tag for internal thoughts, they sometimes behave misleadingly.
You could try issuing two prompts: one to "think", and another to "say to the group". Not as nice as having a model that can properly format its response, but you'll get more accurate results.
