You could instruct the LLM to be more picky about user input in terms of grammar, style, and spelling.
I quite like that idea. Maybe you could use them for verbal exercises that stretch those linguistic muscles - e.g. getting them to explain a complex concept in precise language to an agent that is constantly trying to poke holes in their response. Though maybe that'd just end up being frustrating!