Improving LLM-Assisted Domain-Specific Design Tasks Through Domain-Structured Persona Prompting

Large language models (LLMs) are increasingly being used as assistants for creative design tasks. In knowledge-intensive, highly constrained design tasks, LLM prompt configuration may affect model responses across user needs, accessibility principles, technical feasibility, and implementation conditions, thereby affecting design output quality. To examine this relationship, this paper investigates the effects of a standard LLM assistant, divergent persona prompting, and a Domain-Structured Persona Prompting (DSPP) framework on design outcomes for age-friendly health monitoring wearable devices. Fifty-three participants with design backgrounds completed design tasks under these three prompt configurations. The resulting proposals were then evaluated by experts, and further analyzed through user experience questionnaires, domain knowledge invocation coding, and LLM-as-a-Judge-assisted validation. Results indicate that DSPP enhances the overall quality of design proposals, with particular advantages in age-friendliness, practicality, and implementation feasibility; DSPP-generated proposals also more frequently incorporate standards-based rationale, older user characteristics, and age-friendly design principles. This research shows the potential of DSPP in LLM-assisted professional tasks with complex requirements and constraints.