I know multiple people that carefully prompt to get that done. The model outputs in direct token order, and can't turn around, so you need to make sure that's strictly followed. The system can and will come up with post-hoc "reasoning".