How this works: 1. Users type what kind of page they want and Jev decides what each line is (headline, fact, assertion etc.) 2. Jev makes choices for each layer of the page (opening, stage, light, depth etc.). Parameterizing design is hard (impossible?) but this took some tuning with Claude 3. A full page is constructed with a single call of about 90 questions.

Of course using an LLM would be more prudent for step 1, but I wanted to check the limits of Jev's understanding (it works pretty well in my opinion!).