I reinforced in the prompt so it wouldn't say system messages, but the model is frequently related to this as well, some models just aren't good for conversational purposes (Usually GPT ones)