Do you mean like structured outputs? Unfortunately here the model is guided to explicitly tell you when you violate the rules and why, it can confuse it's system rules with the game rules and say you're not allowed to ask a question about copyrighted material etc.
Mind explaining?
Curious as well
If you constrain the model to a JSON schema, most frivolous refusals go away.
And if you finetune on a few formatted examples the effect is even greater