While the goals each person has in using LLMs are all over the map, I often find that GPT-4 is already very well-tuned without having to provide meta instructio

While the goals each person has in using LLMs are all over the map, I often find that GPT-4 is already very well-tuned without having to provide meta instructions. Yes, it can be flowery and overly apologetic, but all of those extra instructions use up tokens and likely distract it from giving the best possible answer to your question. Perhaps there is a distinction between using LLMs vs experimenting with LLMs, here. Experiments are often fascinating, but I can hit up GPT-4 with questions that jump right into advanced circuit design, and 90% of the time it meets me where I am without any special coaxing required.