It's perfectly reasonable for a process like this to be initiated by an expert, 80% created according to the input and initiation, with a full review at the end.

The expert could further imbue their expertise in what's created.

It goes without saying that basic generation attempts of text will only return the average of the corpus and not much more.