Should an AI design your whole app screen at once, or build it piece by piece so you can tweak each part?

This explores what you gain and lose when an AI builds a whole interface page in one pass, compared with assembling an interface step by step from smaller widgets that the user can refine. The collection has no study that tests these two approaches against each other. It does have good evidence on each one, and together that evidence shows a clear trade-off: one-shot pages look impressive, and composed widgets are easier to inspect and steer.

The case for one-shot pages is strong on preference. When people compare a full web page generated by an LLM with an ordinary markdown chat reply, they choose the page 83% of the time. Those pages match expert-built pages in quality about half the time, and only the newest models can produce them at all Do full web pages beat markdown chat for LLM responses?. The weakness is that a polished page can hide what it left out. A benchmark of generative UI tools found that about a quarter of stated design rationales never made it into the output. For functional requirements the figure rose to 34%, and the tools recognized only half of the UX principles written into the prompts Do generative UI tools actually implement their stated design rationales?. One-shot generation gives you something complete-looking, and checking what is actually missing takes real effort.

The iterative widget approach has a different profile. Task-specific generated interfaces beat plain chat in over 70% of cases, and a large part of that comes from structured representation combined with step-by-step refinement, which lowers cognitive load Do generated interfaces outperform text-based chat for most tasks?. The TaskArtisan study shows the cost. Widgets made analysis clearer, but they also made it rigid, and they added prompting overhead. Users hit a trade-off between malleability and specification: interfaces that are easy to use are hard to reshape mid-task, and flexible interfaces ask non-programmers to think like engineers Do generated analysis UIs really work better than chat?. Composing widgets one at a time doesn't remove that tension. It moves it to every step.

Research on AI agents that operate existing interfaces offers a lateral view. Agents improve when one hard composite task is split into separate jobs: planning apart from grounding (working out which on-screen element an instruction refers to) Can structured interfaces help language models control GUIs better?, and understanding the screen apart from choosing the action Why do vision-only GUI agents struggle with screen interpretation?. That pattern suggests full-page generation may suffer from the same overload: layout, content, interaction logic and design intent all have to be produced in a single pass. A related point is that LLMs write token by token with no built-in pause to reflect or revise Does AI text generation unfold through temporal reflection?. A one-shot page is a first draft with no revision built in, while iterative composition adds the revision loop from outside. There is a cost on the user's side too. AI context is already mutable and short-lived in ways people struggle to keep track of How does AI context differ from conventional software context?, and an interface that keeps changing underneath them makes that worse. A finished page, whatever its gaps, at least stays put.

The main gap in the collection is a direct comparison: the same tasks, with one-shot and composed interfaces, measuring both preference and how many requirements actually get implemented. Until that exists, the evidence suggests one-shot pages work best for presenting information, and iterative composition works best for ongoing work where the user needs to see and correct what the AI built.