LEGOUI: Designing with UI-DSL Bricks to Balance Transparency and Controllability

LegoUI addresses the 'black-box' nature of current generative UI (GenUI) tools, which often force users into a cycle of trial-and-error prompting without insight into the model's reasoning. The authors propose a staged framework that decomposes interface creation into discrete, interpretable steps. By using a UI-specific domain-specific language (UI-DSL), the system externalizes intermediate design decisions—such as component selection, layout structure, and interaction behavior—as persistent, editable items. This allows users to inspect, accept, reject, or modify the model's assumptions before a final interface is rendered.

In a technical evaluation using 40 real-world design prompts, the requirement analysis stage of LegoUI achieved over 95% accuracy in capturing explicit requirements with near-complete coverage and zero redundancy. User studies involving 15 participants demonstrated that this staged approach significantly enhances perceived transparency and controllability. Participants reported that they could better align the final output with their original design intent compared to traditional one-shot generative systems. However, the study also noted that this increased control comes with a trade-off: users face higher interaction costs, as they must spend more time evaluating and managing the intermediate specification items.

This research shifts the paradigm of AI-assisted design from passive, end-to-end generation toward a collaborative, co-design process. By making the 'reasoning' of the AI visible and editable, LegoUI empowers beginner designers to maintain agency over their creative work. It provides a practical solution for integrating generative AI into professional workflows where precision, consistency, and the ability to iterate on specific design dimensions are critical.