The worst PRDs I have written were complete. Every field filled, every screen described, and still useless, because they recorded decisions without recording why.
The three questions
A PRD earns its keep when a new engineer can answer, without asking: what problem is this for, what counts as done, and what did we decide not to do.
Acceptance criteria are the spec
If a line cannot be tested, it is a preference. Writing criteria first exposes the parts of the problem still unresolved, which is exactly the part worth arguing about early.
Keep the cuts visible
A short "not now" list saves more time than any diagram. It stops the same scope returning every second sprint with a new name.