We built a thing that takes one photo of a child and paints that child into a pre-written 10-page picture book. Thirty books, eleven pages each, on a free image model. Every page has to look like the same kid.
Getting from "mostly works" to "every page of every book is clean" took about three weeks, and most of what I learned contradicted what I thought I knew about prompting. Here are the parts that generalize.
The headline finding
At one point we had a page — a doctor's office scene — that kept rendering two children and two doctors. Standard debugging instinct: be more explicit. So the prompt grew:
EXACTLY ONE child
