“Make it nicer” will lead to unpredictable and controversial changes, likely to clash with the project owner’s vision. The owner would be better off using the LLM as a collaborator, if anything, to decide where to take the design.

However, UX can also be improved by some simple heuristics that could be seen as mechanistic by the standards of modern LLMs.

Performance and responsiveness, for example, contribute to UX even when the UI remains fixed. I’d wager Fable-level LLMs could improve perf on most open-source GUIs by benchmarking typical use cases and progressively applying standard tricks like caching, buffering, batching network calls.