What's actually in a moodboard

"Make it good" isn't a spec. Neither is "make it feel premium." AI will generate a hundred directions and can't tell you which one is right, because nobody has said what right means.

I wanted to know if that could be extracted. So I built a pipeline that takes a moodboard, measures what's actually in the images, clusters them, and writes out a thesis: the values, the rules it follows, the things it never does.

Seventy-six references collapsed into eight channels. The forbidden list turned out to be the useful half. No gradients. One accent color. Never a floating drop shadow. Those are the rules I'd been enforcing for years without saying them out loud, and they're far easier for a machine to follow than any positive instruction.

It's on GitHub. The design of this site came out of it, though I overruled the output twice, which is its own finding: the thesis was right about structure and wrong about warmth.

What I keep coming back to is that a specific point of view compresses better than a general one. Telling a model to make it beautiful gives you the average of everything. Telling it never to use a gradient gives you something that looks like a person made it.