So, its actually a lot of language tweaking to get just the right context/task description/data embeddings so the LLM (GPT3/4) gets it right >=90% of the time, which surprisingly often is better than actual humans, and in many cases there are also ways to detect imperfection and simply retry automatically which increases success chances even further.
The fetching/formatting/submitting data part (the manual coding) is getting easier over time, but the prompting remains somewhat, and I so far had no luck with any kind of recursion to let the LLM design its prompt, since ultimately all the specifics needed in the context has to somehow got into the context, which is me engineering it into big string structures.
probably doesn't sound shiny, but step-by-step making jobs irrelevant in businesses without sacrificing customizations. I think of it as a silent revolution thats happinging in many places now, ultimately making myself redundant, but hey the ride is fun!