The discipline is definitely one of fumbling about, trying different adornments to your spell.

That said, it's easier if you are both an engineer and an accomplished practitioner of the English language. The models will recognize both structure data and examples, and paired with a detailed explanation of what you want, most times you'll get something useful.

I spend more time getting the damn things to be *consistent* with their output, because we still have to parse and use the results in a meaningful way.

Goodside is a great person to follow for ideas and techniques.

With respect to longevity, people are already building ways to use the AI to talk to the AI, langchain being a good example of such.