End every serious prompt with 'list your assumptions'

An AI tool rarely says "I don't know". If your prompt leaves a gap, it fills it with something reasonable and carries on. The answer reads fine and the guess is buried somewhere in the middle.

Add this to the end of the prompt:

After your answer, add a section called "Assumptions" listing everything
you assumed because I didn't tell you, and a section called "Not sure
about" for anything a person should check. If you assumed nothing, say
so.

It won't catch everything. The model can only list the assumptions it knows it made, and it'll miss some. But what it does list is often worth seeing: a year it picked for you, an audience you never named, UK law when you meant somewhere else, a feature that's only in the newest version, a number carried over from your text instead of worked out.

Each one takes a few seconds to answer with "yes, that's right" or "no, fix that". Finding the same assumption inside a finished document takes a lot longer.

For anything that matters, go a step further: "Mark each claim as either from the text I gave you or from your general knowledge." Anything in the second pile needs a source before it goes in a report.

A wrong assumption stated confidently is still wrong, so treat the list as a prompt for your own review, not a guarantee. If you'd rather catch the gaps before it writes anything, get it to ask you questions first.