On giving AI taste

Ask a model for a card and you usually get one that looks finished in the preview: rounded corners, a soft black shadow, twenty-four pixels of padding and a hardcoded white background. It also looks like every other card on the internet, which makes sense once you remember the training set is the median of public design. Accept the first pass and you ship that median. That’s less a failure of the model than a failure of review. Taste is the work after generation, when you read the output, name what’s wrong and change it before anyone else has to live with the draft. The median card The default CSS looks harmless: .card { border-radius: 12px; box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1); padding: 24px; background: white; }Each line fails a different system decision. The radius ignores a sharper product surface; the shadow invents its own light instead of using elevation tokens; the padding misses the spacing scale; and the background breaks the moment dark mode appears. None of that shows…