The most important rule in my palm-reading prompt is a refusal

When I started on the palm reading for Numenist, the hard part wasn't palmistry. It was that a vision model will happily describe lines it can't actually see. Hand it a blurry photo and it will still find a strong heart line and a forked head line, and report them with total confidence.

So before the reader says a word, the prompt makes it check four things: the whole palm and every fingertip are in frame, the light is even, the heart, head, and life lines can each be traced on their own, and the photo is in focus. If any check fails, it asks for a better picture and says plainly what it can and can't see. The prompt names guessing from a blurry photo as the number one risk in AI palm reading, and after testing it, I agree.

The camera does the same job up front. It coaches the person to move closer or hold steady, and it won't take the shot until every fingertip is in view.

I think the lesson travels well beyond palms. One of the most useful things you can teach a model is when to stop.