The single best prompting trick I've found: ask the model to think step by step BEFORE answering, in the same message.

Not "give me the answer," but "reason through it, then give me the answer." Accuracy on anything multi-step goes up dramatically. It's the difference between a gut reaction and actually thinking.