Generating Proto-Personas through Prompt Engineering: A Case Study on Efficiency, Effectiveness and Empathy
In this paper, we propose and empirically investigate a prompt engineering-based approach to generate proto-personas with the support of Generative AI (GenAI). Our goal is to evaluate the approach in terms of efficiency, effectiveness, user acceptance, and the empathy elicited by the generated personas. We conducted a case study with 19 participants embedded in a real Lean Inception, employing a qualitative and quantitative methods design. The results reveal the approach’s efficiency by reducing time and effort and improving the quality and reusability of personas in later discovery phases, such as Minimum Viable Product (MVP) scoping and feature refinement. While acceptance was generally high, especially regarding perceived usefulness and ease of use, participants noted limitations related to generalization and domain specificity. Furthermore, although cognitive empathy was strongly supported, affective and behavioral empathy varied significantly across participants. These results contribute novel empirical evidence on how GenAI can be effectively integrated into software Product Discovery practices, while also identifying key challenges to be addressed in future iterations of such hybrid design processes.
We explore a prompt-engineering-based approach using LLMs to generate proto-personas during early-stage Product Discovery, where user data is limited, with a specific focus on empathy and usercentered design [7, 17, 18, 22]. Our approach complements manual persona generation methods [14] by offering a rapid alternative, especially when detailed user data is unavailable, aligning with Liu et al.’s call for structured persona creation processes [25].
Synthesis . Our prompt-engineering approach cut protopersona creation from days to under six minutes, automating repetitive steps and accelerating output. Participants reported increased productivity, richer context, and more strategic discussions.
Effect 1: Motivation to make adaptations to the product to meet certain proto-personas. Developing empathy with the proto-personas (not necessarily using AI) is a factor that contributes to the motivation to help them [22]. The theme Factors that facilitated the development of empathy with proto-personas reinforces this aspect. P7 expresses it: “What contributes most to identification is behavior (...) For example, (...) the demanding persona relies on the accuracy of the data. We also feel it, even as researchers”. This connection helped participants to think about the difficulties of the proto-personas, which culminated in the desire to solve these needs.
Another aspect that possibly contributed to the motivation to make adaptations was the fidelity of the AI-generated proto-personas to the real-world context—through attributes like education, age, and occupation. This aspect is consolidated by the Conformity of the Proto-personas with the Project Context theme. As P6 notes: “And it creates that persona with something that you will already deal with on a daily basis, in your routine. And I think that makes you think that it is a real person...”. This perceived realism fostered empathy and motivation among participants, reinforcing their engagement in designing for these users [30].
Finally, we perceived that during the approach there was an encouragement of participation, as reinforced by the theme Team Collaboration and Alignment. P8 comments: “So people wanted to talk, this desire to share became greater and I think that perhaps in a process without this people would be more withdrawn”. The fact that the proto-personas were generated by an LLM may have influenced participants’ willingness to critique them. Since the personas were not created by a human team member, participants might have felt more comfortable expressing criticisms [8]. In this sense, having greater activity in the discussion of proto-personas is beneficial to better understand the proto-personas [46]. This better understanding, caused by participation, is a possible cause for the motivation to make adaptations of the product [30].
Effect 2: Proto-personas generated as a central reference in refining functionalities and user journeys. The aforementioned alignment with user expectations and context, Conformity of the Proto-personas with the Project Context theme made the AI-generated proto-personas valuable not just for initial understanding but also as concrete references in subsequent LI stages. P1 reflects on how these proto-personas influenced functionality decisions: “If I am building a system for an older audience... I will be concerned with larger fonts (...) I think the proto-persona helps us there”. The contextual compliance extended even further, enhancing the dialogue with stakeholders during prototyping. P12 emphasizes: “I think this is something that the Chat proto-personas got right, because it has something that customers were already talking about”. The alignment with stakeholder expectations made the proto-personas more actionable, turning them into useful inputs for the definition of screens and functionalities.
Another factor was the base for discussion that the approach provides to refine the proto-personas — the usefulness of the approach. Since one of the main functions of proto-personas is to serve as a basis for discussion [28], the approach aligns with this intended workflow. P15 supports it: “And it sets a ground, the basis of what will be worked on, everyone is on the same level from there. (...) So it speeds up discussions, consensus”.
On the other hand, despite some conformity in the proto-personas, the participants reported some generalization - Inconsistencies and Generalizations in Proto-persona Generation - or partial conformity - Partial conformity of proto-personas concerning the project context - of the LLM regarding the project. P3 brings this frustration: “...the way it was generated here, they are personas for a system, which analyzes legal documents, but our system is for attorneys, and it is for attorney advisors, so... It has needs that are a little different from those that were raised here”. P12 discusses the lack of specialization of proto-personas: “The part about needs, objectives and behavior for everyone (...) it seems like it was the same thing written in a different way.” Indeed, LLM often struggles with complex tasks involving semantic nuance [40] or lack of nuanced understanding in specialized domains [2] - as the legal - domain. That aspect opposes the novelty presented by the LLM in other studies as Marques et al. [29]. This LLM limitation may have caused more generic or inconsistent proto-personas.
Increments for Approach Improvements -, generally pointing to the construction of a tool or even greater interactivity in the approach. P9 comments: “I believe that perhaps a platform, an application like that that already had the basic prompts, basic prompts already inserted, right, in the ChatGPT”.
The tendency for participants to agree is reinforced mainly by the theme of Approach Efficiency. P13 argues: “The issue of one or two days of work was reduced by one hour, so it was very good”. This time saving is also related to the theme of Approach Usefulness, more specifically, its use as a basis for discussion. Most participants brought up the argument that having something to start from (the result of the approach) greatly speeds up the refinement of proto-personas.
Synthesis . Results show broad TAM acceptance: most participants found the approach useful, easy to use, and valuable for generating context-rich proto-personas, speeding up design, and supporting collaboration. Despite minor concerns about generalization, prerequisites, privacy, and empathy gaps, participants expressed strong intent to reuse and adopt similar tools
Affective empathy. This dimension of empathy relates to scenarios where participants can see themselves in the proto-persona’s place, defined as “feeling the same affective state as another person” [18]. Affective empathy presented the lowest level of agreement among participants, based on the Likert scale data.
8 CONCLUSION AND FUTUREWORK
The approach proved efficient and effective, reducing the time and effort required to create proto-personas while enhancing productivity and practicality. It enabled the generation of realistic, contextaware personas that supported collaboration and informed decisionmaking, even with limited stakeholder input. Participants positively received the method across all TAM dimensions, with no disagreement reported; neutral feedback was linked to overgeneralization, occasional confusion, or limited emotional connection. Empathywise, the approach strongly fostered cognitive empathy, showed mixed results in affective empathy, and revealed opportunities for improvement in behavioral empathy, particularly in strengthening engagement and alignment with user behaviors.
Future work. Explore developing a more intuitive tool to support the approach, integrating image generation, and improving non-functional aspects such as accessibility. Enhancing affective and behavioral empathy remains a challenge, suggesting hybrid strategies like narrative techniques or stakeholder reviews. Adaptive prompting with domain-specific cues may improve contextual fidelity. Further studies should also examine GenAI’s role in human– AI co-creation, support collaboration without direct stakeholder access, streamline convergence phases, and test the approach across diverse persona profiles to assess generalizability.