I've written prompts, there's no "engineering" in copy-pasting the same keywords literally everyone else uses. Huge Frankenstein abominations of "Trending on Artstation" 30 synonyms for "high quality" and of course none of the tags have custom weights because that would actually take skill and understanding on how prompts work vs just puking what's popular to make some flim flam that fakes the illusion of competency.
I am a senior developer working on and with LLMs. I do spend a lot more time than I expected doing prompt engineering. Not putting quotes, after laughing at it I do think this is becoming a real thing. Understanding zero shot, one shot, few shots, understanding the quirks of the different models, forcing a format by providing part of the answer, compressing context to fit a context window...
The code around it is simple: some string manipulation, some database access. Actually maybe half the code around it I made it generate by GPT-4.
I don't know if it is displacing data scientist yet, but I am pretty sure that data scientists right now spend as much time writing prompts than writing code.