I asked my AI to make an icon. A chess pawn, flat design, two-tone green. Simple enough.
It came back wrong. The highlight on the pawn's head was just... a random white swoosh. It looked like clip art from 2008. I couldn't explain why it was wrong, but I knew it was.
So I asked a designer.
She looked at it for maybe three seconds and said: "The highlight has to follow the curve of the shape it's on. Same center, smaller radius. It's a concentric arc, not a decoration."
That's one of those rules that seems obvious once someone says it out loud. Highlights aren't arbitrary. They're geometry. The curve of light on a sphere follows the sphere. The curve of light on a cylinder follows the cylinder. If your highlight doesn't share a center point with its parent shape, it reads as fake.
I fixed the icon. It looked right immediately.
Then I did something I wouldn't have thought to do six months ago. I opened the skill file — the markdown document that tells my AI how to generate icons in this style — and I added the rule:
Highlight arcs on curved surfaces MUST be concentric with the outer shape — same center point, smaller radius.That's it. One line. Now every icon this system generates follows a principle that took a trained designer years to internalize.
Here's what's interesting about this: I'm not teaching the AI to draw. I'm teaching it to see. There's a difference. Drawing is mechanical. Seeing is knowing why something looks wrong before you can articulate the geometry.
The designer didn't give me a formula. She gave me a way of looking. I translated that into a constraint. The constraint lives in a text file. And now the machine produces work that passes a designer's eye test — not because it understands light, but because someone who does wrote down the rule.
I've been collecting these moments. A copywriter's instinct about sentence rhythm. A product manager's heuristic for when a feature is "done." A photographer's rule about where the eye goes first. Each one becomes a line in a file somewhere.
The AI doesn't get better by training on more data. It gets better because the people around me keep teaching it things they spent years learning. And unlike a junior designer who might forget, this one remembers every rule it's ever been given.
I'm starting to think the most valuable AI skill isn't prompting. It's listening to experts and knowing which of their instincts can be written down.