I use GPT-4 pretty consistently(set up a discord bot for myself). What I found myself doing was tending towards the most simple prompt that the LLM would still understand - If I asked a human expert the types of prompts I was giving GPT, I most likely would've gotten a clarifying question rather than an answer like the LLM was giving me, simply because I'm talking in such short and concise sentences.
I think the interesting thing is that the more concise a message is to a fellow human, the more work needs to be done by the other party in order to actually decode my message, even if it is ultimately understandable. Whereas with LLMs, shorter token length doesn't really matter: matrices of the same size are being multiplied anyways.