The useful version of this skill looks nothing like the tricks that circulate. It is closer to writing a shot description for a crew: subject first, then camera, then light, then the two things that must not happen. Vague adjectives are the main cause of expensive variance.
It is also the least durable part of any pipeline. Models change and prompt habits stop working; plates, identity sheets, gates and ledgers do not. Anything you want to still own in two years belongs in those, not in a prompt.
What makes a good AI image prompt?
One clear subject, one named camera and lens behaviour, one lighting setup, and specific nouns instead of mood adjectives. Then a negative list containing only failures you have actually seen.
Is prompt engineering still a useful skill?
Yes, but it is a small part of production. The leverage is in the pipeline around the prompt: what is locked before generation starts and what gets checked afterwards.