The default style really annoys me with its verbosity and overuse of numbered lists. Describing a tone and length of responses and anything else you want - think step by step, first principals, show don’t tell - all work pretty well.

You can even ask it to only reply with a Boolean true or false - which certainly doesn’t play to its strengths

Actually, that's a by-product of RLHF. A base model is usually not that verbose.