PersonaGen: Agile Persona Generation

• PersonaGen is a web-based tool that converts raw user feedback into structured persona templates for agile software development.
• It leverages GPT-4 for data cleaning, integration, and analysis, while a Neo4j knowledge graph organizes and recommends persona attributes.
• Evaluations indicate its potential to streamline requirements engineering despite challenges with non-functional requirements and varied user expertise.

Searching arXiv for PersonaGen and closely related persona-generation papers to ground the article with current literature. PersonaGen is a web-based tool that generates persona templates from user feedback to support requirements engineering in agile software development. It combines GPT-4 for feedback processing with a Neo4j-backed knowledge graph for structuring and recommending persona attributes, and it is intended to automate the cleaning, integrating, and analyzing of feedback so that teams can obtain structured persona artifacts for agile ceremonies and design work. In the reported study, PersonaGen was evaluated through qualitative feedback from a small-scale user study involving student software projects, and the results were mixed, with particular challenges around persona-based educational practice and non-functional requirements (Zhang et al., 2023).

1. Motivation and problem setting

PersonaGen is positioned in a context where personas are widely used in agile to understand users, drive user stories, and uncover diverse and inclusive requirements. The motivating problem is not the conceptual value of personas, but the operational difficulty of maintaining and updating them when requirements change rapidly. The paper states that many teams, especially startups, perceive persona work as time-consuming, and that there is a gap in tools that directly turn raw feedback into actionable personas within agile workflows (Zhang et al., 2023).

Within this framing, PersonaGen aims to automate the heavy lifting associated with persona construction from feedback. The feedback sources explicitly referenced include app reviews, interviews, surveys, and usability testing. The intended outcome is not merely a textual summary of users, but a structured persona template that can be consumed during requirements analysis, backlog refinement, user story writing, and sprint planning. A plausible implication is that PersonaGen treats persona generation as a requirements-engineering artifact pipeline rather than as a purely UX documentation exercise.

The tool also reflects a specific view of agile practice. Persona artifacts are treated as instruments for clarifying who user stories are for, validating acceptance criteria, and highlighting both functional and non-functional requirements that matter to specific user segments. At the same time, the study shows that non-functional requirements remained difficult to analyze in practice, which limits any interpretation of PersonaGen as a complete solution for agile requirements analysis (Zhang et al., 2023).

2. Architecture and end-to-end pipeline

PersonaGen is implemented as a web application with an HTML, CSS, and JavaScript frontend, a Java Spring Boot backend, Neo4j as the knowledge graph database, GPT-4 accessed via OpenAI for language processing, and CSV export for processed feedback (Zhang et al., 2023).

Its pipeline is described in four stages. First, user feedback is collected. In the evaluation reported in the paper, the feedback came from student software projects, although the tool is described as applicable to app reviews, interviews, surveys, and usability testing. Second, GPT-4 is used for cleaning, integration, prediction, and analysis of raw feedback, producing “well-processed” data suitable for downstream persona generation. These outputs are saved in CSV format. Third, a knowledge graph is constructed to strengthen connections among persona-related attributes and to support classification and recommendation of personas. Fourth, persona templates are generated from those structured attributes (Zhang et al., 2023).

The technical description is intentionally high-level. The paper does not disclose concrete prompt structures, prompt templates, or scoring functions used with GPT-4. It also does not provide algorithmic pseudocode, formal definitions, graph schemas, or equations. The graph description is categorical rather than schema-level: Neo4j stores nodes and edges representing persona attributes and requirements, which are then used for classification and recommendation. For an academic reader, this absence is material, because it means that PersonaGen is documented as a systems concept and prototype rather than as a formally specified algorithmic method.

3. Knowledge graph and persona template structure

The knowledge graph is the structural core of PersonaGen. The paper states that it strengthens connections among persona-related attributes and supports classification and recommendation of personas. The node types explicitly mentioned are user requirement nodes, requirement type nodes, demographic nodes, job role nodes, and user-defined nodes. Edges connect personas to these attribute nodes and also connect attributes to one another, such as linking a persona to specific requirement types or demographics (Zhang et al., 2023).

This graph is not presented with a formal labeled property graph schema or explicit edge labels. Instead, its role is functional: it provides a way to relate extracted or inferred requirements, requirement categories, demographic descriptors, and job roles so that persona templates can be classified and similar personas recommended. This suggests a hybrid architecture in which GPT-4 supplies transformed semantic material from feedback while the graph supplies persistence, structure, and relational continuity across iterations.

The generated persona templates contain the fields explicitly mentioned in the paper: demographic information, visualization, feedback, motivations, and requirements for applications derived from user feedback. The visualization is present in the tool, but the paper does not include its detailed design. The figure “An Example of Persona Template” is referenced, but its textual contents are not reproduced, and no formal template schema is given. As a result, PersonaGen’s template should be understood as a structured but only partially specified artifact.

A recurrent limitation concerns non-functional requirements. The reported evaluation found that teams struggled to analyze non-functional requirements, especially accessibility. The current pipeline focuses on general attributes and requirements, and future work is described as integrating accessibility requirements and more human-centric aspects into the templates (Zhang et al., 2023).

4. Role in agile requirements engineering and design practice

PersonaGen is explicitly situated in agile workflows. In Scrum and requirements analysis, development teams can use it to transform feedback into persona templates that inform backlog refinement, user story writing, and sprint planning. The paper also states that personas can help validate acceptance criteria and identify gaps across user segments. In sprint review and retrospective settings, new feedback can be gathered, the knowledge graph updated, and personas regenerated to reflect changing requirements (Zhang et al., 2023).

The tool is also described as useful for persona-based design. Designers and product owners can use persona content such as demographic information, motivations, feedback excerpts, and application requirements to guide UI/UX decisions and ensure coverage of diverse users. The knowledge graph supports continuity across projects and iterations by connecting role, demographic, and requirement-type attributes over time. This is especially relevant in agile settings where requirements evolve quickly and persona artifacts can otherwise become stale.

Practical usage guidance in the paper is correspondingly procedural. Teams ingest feedback, run GPT-4 processing, populate the knowledge graph with nodes representing requirements, requirement types, demographics, and job roles, generate persona templates, and then iterate as requirements evolve. Users may add custom node types as needed. The source code is reported as available at the project repository, which positions PersonaGen as a reproducible tool artifact in addition to a research proposal (Zhang et al., 2023).

A common misconception would be to treat PersonaGen as replacing stakeholder validation or training in persona use. The paper does not support that interpretation. It states that PersonaGen helps by automating data processing and offering pre-structured personas, but does not replace training on persona use in agile ceremonies and UX/RE practices.

5. Evaluation findings, limitations, and threats to validity

PersonaGen was evaluated through small-scale user studies involving feedback from three student software projects, totaling 13 third-year undergraduate students. The analysis was qualitative; no quantitative metrics or performance numbers were reported. This is a defining characteristic of the evidence base for the system (Zhang et al., 2023).

The findings were mixed. On the positive side, some participants judged the GPT-4-generated personas as more accurate than their own independent analysis, which they attributed to their limited qualitative analysis experience. Experience varied across participants, and some students with industry experience demonstrated a better grasp of persona concepts and applications. The principal challenges reported were lack of education and practical knowledge about persona-based practices in student projects, and difficulty analyzing non-functional requirements, especially accessibility.

The paper’s limitations follow directly from this evaluation design. The technical description is high-level, with no detailed algorithms, prompts, schemas, or quantitative performance metrics. The study is small-scale, qualitative, and focused on student projects, so generalizability to industrial contexts remains to be demonstrated. Threats to validity are implied by the limited sample size, the participant profile, the absence of controlled comparison to alternative tools or baselines, and the fact that GPT-4 behavior may vary with prompts and model versions that are not disclosed.

The practical cautions listed for use are similarly significant. Potential failure modes include LLM hallucination or overgeneralization from limited feedback, biased personas if feedback is not representative, misclassification in the knowledge graph due to ambiguous attributes, overreliance on generated personas without stakeholder validation, and insufficient coverage of non-functional requirements such as accessibility, performance, and privacy. The paper does not address privacy or ethics, though it recommends removing personal identifiers before sending feedback to GPT-4 and complying with data protection requirements. This suggests that PersonaGen’s research contribution is strongest as an initial tool concept for persona bootstrapping, not as a complete governance-ready pipeline (Zhang et al., 2023).

6. Relation to prior and subsequent persona-generation research

PersonaGen is compared in the paper to several earlier strands of automated persona work. Jung et al. (2017) generated personas from aggregated social media data; Branco et al. (2020) focused on automatic persona generation from questionnaires using clustering; and Watanabe et al. (2018) examined data-driven personas in B2B software development with emphasis on retrospective significance rather than automated construction. Against this background, the paper claims novelty in using GPT-4 for cleaning, integrating, predicting, and analyzing user feedback, and in constructing a knowledge graph through requirements, requirement types, demographics, and roles to enable persona classification and recommendation (Zhang et al., 2023).

Subsequent literature has treated PersonaGen as representative of a single-shot LLM approach that transforms user feedback into personas in one pass. The clearest contrast is PerGent, which replaces one-pass prompting with an iterative critique-refinement loop involving a generator agent, a critic agent, and an orchestrator. In an industrial evaluation at Kinaxis, PerGent is reported to achieve the highest expert approval rate, reproduce a larger proportion of expert-authored content than baselines, and add substantial new content, albeit at higher computational cost (Amini et al., 8 Jun 2026). This later comparison clarifies PersonaGen’s place in the literature: it is an early LLM-plus-knowledge-graph system for agile persona generation, but not an agentic or industrially benchmarked orchestration framework.

Broader later work pushes persona generation in different directions. HACHIMI formalizes theory-aligned and distribution-controllable student persona generation using orchestrated agents and a neuro-symbolic validator, producing a 1 million persona corpus for Grades 1–12 (Jiang et al., 5 Mar 2026). “Persona Generators: Generating Diverse Synthetic Personas at Scale” instead optimizes support coverage across diversity axes using an AlphaEvolve loop over generator code (Paglieri et al., 3 Feb 2026). These systems are not extensions of PersonaGen in a narrow software-engineering sense, but they illustrate a broader shift from single-pass generation toward controllability, validation, orchestration, and large-scale synthetic populations.

Future work identified for PersonaGen itself is more modest and more local to requirements engineering. The paper lists better handling of accessibility requirements, integration of more human-centric aspects into persona templates, and use in educational practice for UI/UX design and requirements engineering courses. A plausible implication is that PersonaGen’s enduring significance lies less in final-system performance than in articulating a concrete LLM-plus-knowledge-graph workflow for turning user feedback into persona templates inside agile software development (Zhang et al., 2023).