Vibe Design: Human-in-the-loop Agentic Framework for UI Design with Large Language Models

Advances in LLMs are driving structural changes across the software development landscape, particularly in software product development processes. These transformations extend beyond programming and increasingly influence established UI design workflows. Conventional UI design relies on iterative cycles of prototyping and user testing. However, such processes are often time- and resource-intensive, creating tension with real-world environments that demand rapid iteration. Although prior research has explored AI-assisted approaches to support UI design, most studies focus on isolated functionalities or depend on code-based inputs that are rarely available in practical design settings. To address these limitations, this study proposes Vibe Design, an LLM-based agentic framework that enables iterative interface prototyping and usability test simulation without relying on code-based data. The framework integrates LLM-assisted prototyping and usability test simulation grounded in the Persona-Scenario-Goal methodology. An empirical study was conducted using three types of user data, namely demographic profiles, persona descriptions, and interview transcripts, to evaluate simulation performance. Results indicate that the interview-based agent achieved the highest performance, reproducing 76% of the usability themes identified in human usability testing. By integrating the Persona-Scenario-Goal methodology and Human-in-the-Loop mechanisms, this study demonstrates both the practical applicability and methodological implications for LLM-based UI design agents. While the findings highlight the potential of AI-driven user simulation to reduce repetitive validation costs, they also indicate that higher-order judgment and creative reframing remain the domain of human designers. These results underscore the need for complementary human-AI collaboration in the reconfiguration of contemporary design processes.