When something is wrong, the instinct is to fix the thing in front of you: recolor the button, redo the one bad output. The higher-leverage move is almost always one level down: fix the system that produced it, so everything it makes next is better. AI made that trade cheap enough to be the default. A worked example: instead of nudging one bad generated image, rewrite the criteria the tool runs on, and the whole set improves at once. Point the tool at the tool. The discipline is spending your attention where it compounds, on the machine, not the symptom, and knowing which symptoms are worth going upstream for.