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.