The Anatomy of a Generative UI Session.

Recent advances in generative AI have collapsed the cost of software development, enabling a new generation of tools, such as Lovable, v0, and Bolt, that let anyone build an application through natural language to fit their own needs. The longstanding HCI challenge of letting people without design or coding skill create interactive systems therefore shifts from one of development to one of design. Accounts of design theory, both iterative and reflective, traditionally assume a trained designer refining a stable artifact of their own making, where each change is local and reflects the author's intent. In a generative tool, each of these conditions takes a different form. (i) The user's request operates at a different level of abstraction than the code it changes, so the resulting change is non-local, a structural property of the generative medium. (ii) The change is produced by the model, so it contingently reflects the model's reading as much as the author's intent. (iii) The designer's knowledge, the work of deciding what the interface should contain, how it should be arranged, how it looks and behaves, and whether it serves what the person set out to do, is shared between the model and the user.

To trace how far these accounts extend into this new setting, and where the generative medium requires revisiting them, we built Mark, a custom AI-assisted prototyping web application that reproduces the core features of commercial tools. We recruit 100 users without design or development expertise and task them with building and revising prototypes. We give each participant three tasks: a from-scratch brief in a domain held constant across participants and outside their expertise (T1); a modification task on a shared starting artifact (T2); and a from-scratch brief generated on the fly within the participant's own domain of expertise (T3). We map each prompt onto an established taxonomy of UI prototyping prompts and each interface diff onto a standard taxonomy of interface elements. We analyze how the relationship between the prompt and the interface unfolds, where and how far the generated artifact moves beyond what the prompt asked, and how this depends on the level of the prompt and the kind of element changed (RQ1). We trace whether the result matches the user's request, how users recover when it does not, and what becomes of the changes the user never requested (RQ2). Comparing T1 and T3, we measure how domain expertise shapes the way users steer the model (RQ3). Study 1 collects a behavioral record, which allows us to measure how sessions unfold at scale, but does not show what participants noticed in a generation, why they chose a recovery move, or when they judged the result done. In a second, moderated study, ten non-experts from the same population complete the same three tasks and then review a replay of their own session, reconstructing what they were noticing and deciding at moments selected from the log by fixed rules. Through this reconstruction we recover the reasoning behind the behavior that Study 1 records. Additionally, we recruit ten trained designers to complete the identical protocol, letting us compare how designers and end users read, judge, and steer generation (RQ3). Together, the two studies give the anatomy of a generative prototyping session.