LLMs are good enough at writing interface code that it raises a question: what if you let one write the UI on the fly, while someone is using the app, instead of ahead of time?
To find out, I built a small app that generates a product review form tailored to whatever you're reviewing. Type "DeLorean sports car" and it asks about acceleration and handling; type "wool sweater" and it asks about fit and itchiness. You enter a product, the LLM works out what kind of thing it is, and it builds a form to match — a stable skeleton (overall rating, pros and cons) with category-specific questions grown on top, all rendered in real time.
The idea came from a Google demo back in December 2023, where an AI generated an interactive widget in response to a request — running a tiny software-development cycle on the spot. Review forms felt like a good, simpler test case: the fields change from product to product, but the basic structure stays put.
It works, with the rough edges you'd expect from an experiment. The LLM is surprisingly good at picking relevant questions, and generation is fast enough to feel nearly live. It also asks the occasional redundant question, trips over ambiguous products ("Apple" — the fruit or the company?), and has no memory between sessions, so the same product can produce a different form each time.
I built this mostly to see whether it would work at all. It does — and I suspect some of these ideas belong in Flavorful, or anywhere a malleable interface would help. The source is on GitHub.