Since the AI's answers are often sloppy, I do it in one go: add

Answer in 2 paragraphs of prose, not emojis, no lists.

Works well across models.