Build a visual AI benchmark with a human judging pass
I wanted to know which models could make a useful prototype, but a single design score hid too much. For this round of my How I AI benchmark, I had Claude Code separate four questions: can a model follow art direction, what does it make with almost no direction, can it organize a complicated wireframe, and can it improve an existing design without breaking it?
Claude Code changed the prompts and rubrics, ran the models through Vercel AI Gateway, and extracted their HTML into a browser review page. The run produced 40 prototypes. I could try them, score them without seeing the model names, and download my ratings. The agent handled the repeated runs and comparison-page assembly; I spent my time looking at the actual work.
Then I had it compare my ratings with the automated judge. They disagreed enough to make the judging process worth investigating. One apparently broken prototype turned out to be a problem with the review page’s iframe, which Claude Code reproduced and fixed. We ended with an adjustable comparison of human and automated scores. That was more useful to me than treating the first leaderboard as the answer.