Proto-personas in User Representation

• Proto-personas are early user representations based on assumptions and limited data, capturing demographics, goals, behaviors, and context.
• They serve as lightweight tools for stakeholder alignment in workshops and enable cold-start personalization and synthetic population generation.
• Recent evaluations show efficiency gains of up to 97.5% and improved personalization metrics using both prompt-engineered and graph-mined approaches.

Searching arXiv for the cited proto-persona papers to ground the article in the current literature. Proto-personas are early user representations based on assumptions and limited data, commonly used to guide product definition and stakeholder alignment in early-stage Product Discovery; recent work also treats them as compact, natural-language descriptors mined from large-scale interaction data, and as seeds from which diverse synthetic populations can be generated programmatically (Ayach et al., 11 Jul 2025, SI et al., 7 Jun 2026, Paglieri et al., 3 Feb 2026). Across these formulations, proto-personas encode hypothesized demographics, goals, behaviors, motivations, frustrations, contexts, or aesthetic sensibilities without requiring full-scale user research, while remaining lightweight enough to support rapid ideation, cold-start personalization, and broad coverage of possible user types.

1. Definitions and conceptual boundaries

In the UX literature represented here, proto-personas are “early user representations based on assumptions and limited data,” and they capture hypothesized demographics, goals, behaviors, motivations, frustrations and context. A closely related formulation describes proto-personas as low-fidelity archetypes: short, often qualitative sketches of “who” a user might be, including demographics, goals, and attitudes. This distinguishes them from fully validated personas, which are not required by the workflows described in these papers (Ayach et al., 11 Jul 2025, Paglieri et al., 3 Feb 2026).

Recent work broadens the concept in two directions. First, ZIPP treats personas as concise descriptors of a user's identity and aesthetic sensibilities, mined from large-scale Reddit behavior and used to condition image generation in zero-shot settings. Second, “Persona Generators” define a programmatic function,

$$G_{(\phi,\theta)}(c, D, N) \rightarrow \{p_1 \ldots p_N\},$$

which expands a textual context $c$, diversity axes $D=\{d_1,\ldots,d_K\}$, and a desired population size $N$ into a set of text-based personas spanning relevant opinion and preference space (SI et al., 7 Jun 2026, Paglieri et al., 3 Feb 2026).

These formulations imply three distinct but connected roles for proto-personas: as workshop artifacts for stakeholder alignment, as interpretable conditioning variables for personalization, and as structured seeds for synthetic population generation. The common denominator is interpretability: in all three cases, the representation is textual, editable, and intended to expose assumptions rather than bury them in a latent profile.

Formulation	Inputs	Output
Lean Inception proto-personas	Product Vision, “Is/Is Not/Does/Does Not” matrix, fixed template	At least five proto-personas with fields such as Name, Age, Occupation, Goals, Behaviors, Motivations, Frustrations, Needs
ZIPP personas / proto-personas	Reddit interaction graph, attention-weighted text snippets, MLLM verbalization	A concise persona description capturing likely background, interests, and aesthetic sensibilities
Persona Generator	Context $c$, diversity axes $D$, population size $N$	$N$ fully realized, text-based persona descriptions

A common misconception is that persona artifacts are meaningful only after extensive user research. The materials here describe proto-personas precisely as pre-research or limited-data constructs, though later validation and refinement are repeatedly recommended (Ayach et al., 11 Jul 2025).

2. Prompt-engineered proto-personas in Product Discovery

A concrete prompt-engineering workflow is described for Lean Inception. The setup uses ChatGPT 4o-mini via the OpenAI web interface in an anonymous tab, with default parameters, approximately $0.7$ temperature, default max-tokens, and no system messages beyond context. The prompting pattern combines a Context Manager to lock in the Product Vision and “Is/Is Not/Does/Does Not” matrix, Template Injection to impose a fixed proto-persona schema, and Persona Role to instruct the LLM to adopt a “UX/UI designer” persona (Ayach et al., 11 Jul 2025).

The prompt template contains five activities: the model is told it is a UX/UI designer helping create proto-personas for a Lean Inception; Product Vision text is inserted; the “Is/Is Not/Does/Does Not” matrix is inserted; a schema is provided requiring fields including Name, Age, Occupation, Marital status, Education, Background, Goals, Behaviors, Motivations, Frustrations, and Needs; and the model is instructed to generate at least five proto-personas using only that template. This framing is explicitly designed to constrain output structure while grounding generation in workshop artifacts.

The workshop integration is equally structured. Preparation gathers Lean Inception artifacts, converts them to plain text, and prepares an execution script and questionnaire. In divergence, each participant pastes the context and template into ChatGPT 4o-mini, records start and stop time, copies the generated proto-personas into a form, and completes TAM and empathy items on a 5-point Likert scale. In convergence, a moderator collects all individual outputs and the group selects one reference persona per slot, refining each by merging elements from other outputs until five final proto-personas are agreed. Follow-up interviews then examine TAM ratings, empathy perceptions, and impact on Lean Inception (Ayach et al., 11 Jul 2025).

The reported efficiency result is explicit: mean AI generation time $T_{AI}$ is $c$0 minutes with $c$1, compared with a typical manual proto-persona time of approximately $c$2 minutes per session, giving $c$3 minutes and a percentage gain of approximately $c$4. Effectiveness is not reduced to a single formal score, though an informal quality function is proposed:

$c$5

Acceptance is high on descriptive TAM measures: Perceived Usefulness $c$6 agreed, Perceived Ease of Use $c$7 agreed and $c$8 neutral, Attitude Toward Use $c$9 agreed, Behavioral Intention to Use $D=\{d_1,\ldots,d_K\}$0 agreed and $D=\{d_1,\ldots,d_K\}$1 neutral, and Enjoyment of AI tools $D=\{d_1,\ldots,d_K\}$2 agreed and $D=\{d_1,\ldots,d_K\}$3 neutral. Cognitive empathy is also reported as uniformly strong at $D=\{d_1,\ldots,d_K\}$4 agreement, whereas affective and behavioral empathy vary substantially across participants (Ayach et al., 11 Jul 2025).

Thematically, the generated proto-personas motivated product adaptations, served as a central reference in defining features and user journeys, and guided MVP scoping and prototyping even when stakeholders were absent. At the same time, participants reported generalization and domain-specificity limits, especially in legal-tech reporting, as well as inconsistent capture of accessibility and non-functional requirements.

3. Graph-mined proto-personas for zero-shot personalization

ZIPP operationalizes proto-personas as compact, natural-language summaries of identity and aesthetic preference inferred from user–community interactions and image-posting behavior. Its input is a bipartite Reddit interaction graph in which users $D=\{d_1,\ldots,d_K\}$5 and subreddits $D=\{d_1,\ldots,d_K\}$6 are nodes, edges $D=\{d_1,\ldots,d_K\}$7 record comment plus post counts $D=\{d_1,\ldots,d_K\}$8, subreddit nodes are initialized with a $D=\{d_1,\ldots,d_K\}$9-dimensional OpenAI text-embedding of the subreddit description, and user nodes are initialized with small Gaussian noise. The encoder consists of two layers of GATv2, each with $N$0 attention heads; PMI-normalized edge weights are encoded in $N$1; and final embeddings are normalized as $N$2 and similarly for $N$3 (SI et al., 7 Jun 2026).

Training jointly optimizes two InfoNCE objectives with in-batch negatives. The graph-structure term aligns each user with the subreddits they frequent:

$N$4

The visual-behavior term aligns user embeddings with CLIP-image embeddings extracted from images actually posted by users:

$N$5

The total loss is

$N$6

with $N$7. The stated rationale is that $N$8 captures “who they talk to,” while $N$9 captures “what they post,” so the learned space reflects demographic, topical, and aesthetic senses of identity.

After training, proto-persona verbalization proceeds by gathering first-layer attention weights $c$0, normalizing them to $c$1, allocating a total token budget $c$2—for example $c$3 tokens—across top subreddits in proportion to $c$4, selecting the user’s highest-scoring posts and comments until each subreddit’s budget is filled, concatenating these snippets into a context buffer $c$5, and feeding $c$6 plus a role-play system prompt into GPT-4o or another MLLM. The instruction is: “Write a concise, coherent persona description capturing this user’s likely background, interests, and aesthetic sensibilities.” The resulting one-sentence descriptor is then used to rewrite future prompts (SI et al., 7 Jun 2026).

The sample proto-personas are stylistically dense rather than demographically exhaustive. Examples include “A user who loves muted, vintage portraiture with soft shadows and grainy textures,” “A street-photographer drawn to high-contrast neon-lit urban scenes and cinematic wide-angle compositions,” and “An anime fan favoring pastel color palettes, character-focused compositions, and soft bloom effects.” Each persona typically encodes a demographic vibe, a topical interest, and aesthetic sensibilities. This suggests a compressed representation optimized for controllability in generation rather than for the fuller biographical schemas used in Lean Inception.

ZIPP is explicitly motivated by cold-start limitations in existing personalization methods: prior approaches require dense interaction histories or per-user fine-tuning and collapse context-dependent preferences into a static representation. In contrast, ZIPP conditions generation on natural-language personas without user-specific training data or weight updates, and in the few-shot setting it matches or exceeds fine-tuned baselines trained on more than $c$7 examples per user (SI et al., 7 Jun 2026).

4. Persona Generators and support-coverage-oriented synthesis

“Persona Generators” shift proto-persona work from single-artifact creation to scalable population synthesis. The core claim is that most prior approaches prioritize density matching—replicating what is most probable—rather than support coverage—spanning what is possible—thereby leaving long-tail behaviors underexplored. The proposed framework therefore evolves lightweight generator code that expands minimal descriptions into diverse persona populations tailored to arbitrary contexts (Paglieri et al., 3 Feb 2026).

Optimization uses an AlphaEvolve-style loop in which the object being searched is the generator code $c$8. For each candidate, the framework samples a batch of questionnaires, invokes the generator to produce personas, simulates those personas’ responses to items via Concordia to obtain numeric embeddings $c$9, computes diversity metrics, and averages them to obtain fitness. Separate elite sets are maintained for each metric; an LLM acts as a mutation operator to modify Stage 1 or Stage 2 of the code; poorly performing candidates are replaced; and periodic extinction events reset low-performing islands from the best elites. The paper reports approximately $D$0 iterations (Paglieri et al., 3 Feb 2026).

The framework evaluates six complementary diversity metrics over persona embeddings $D$1: support coverage estimated by Monte Carlo sampling; convex hull volume $D$2; mean and minimum pairwise distances; dispersion, defined as the largest empty region in the response space; and density matching via KL divergence relative to a Sobol reference. The use of both coverage and density metrics is methodologically important: high-probability realism and broad support are treated as distinct desiderata rather than interchangeable ones.

A concrete example begins from the seed context “Elderly residents of a rural Japanese village, 2010,” with diversity axes $D$3 and $D$4. Stage 1 samples five points in $D$5 via Sobol or stratified sampling and prompts the LLM for one-line descriptors such as “Persona A: 82-year-old, strong village bonds, low tech curiosity, very traditional.” Stage 2 then expands each descriptor in parallel into one or two paragraphs with name, bio, formative anecdote, and current attitudes. The output is a set of unique personas varying along all three axes (Paglieri et al., 3 Feb 2026).

On $D$6 held-out questionnaires with $D$7 per scenario, the evolved generators outperform Nemotron-Personas, a Concordia formative-memory baseline, and a Name-only baseline on all six metrics. Coverage exceeds $D$8 versus approximately $D$9 for the best baseline, convex hull volume increases by $N$0–$N$1 over seeds, mean pairwise distance rises steadily, and dispersion and KL improve. The reported gains also generalize to held-out test questionnaires and to two downstream tasks: comedy writing and conflict resolution (Paglieri et al., 3 Feb 2026).

5. Evaluation regimes across proto-persona research

The three lines of work employ markedly different evaluation logics. In Lean Inception, evaluation emphasizes workflow efficiency, stakeholder acceptance, and empathy. Efficiency is directly timed; acceptance is captured through TAM-style agreement distributions; empathy is decomposed into cognitive, affective, and behavioral components; and effectiveness is assessed thematically rather than via inferential statistics. The paper explicitly reports no inferential tests such as $N$2-tests or $N$3-values, relying instead on Likert distributions and interview synthesis (Ayach et al., 11 Jul 2025).

In ZIPP, evaluation targets alignment between generated images and user preference distributions. ZIPBench is described as the first zero-shot image personalization benchmark and contains $N$4 real users linked between Reddit and the Civitai AI-art platform, $N$5 images they actually generated, and graph-mined proto-personas for each user. Automated metrics include CLIPScore, PIGReward, and CMMD, where CMMD measures distributional divergence between a user’s ground-truth images and persona-conditioned generations. A typical zero-shot comparison on SDXL plus GPT-4o reports CLIPScore increasing from $N$6 to $N$7 and PIGReward improving by approximately $N$8, while CMMD decreases from approximately $N$9 to approximately $N$0; in few-shot personalization with five examples, CMMD reaches $N$1 versus $N$2 for ViPer. Across four benchmarks and $N$3 LLMs spanning five model families, persona conditioning yields consistent gains of $N$4–$N$5, and human evaluation with $N$6 participants and $N$7 pairwise comparisons reports a $N$8 win rate over generic generation and $N$9–$0.7$0 over fine-tuned baselines (SI et al., 7 Jun 2026).

Persona Generators evaluate a different target: diversity coverage of synthetic populations. The six-metric suite—coverage, hull volume, mean distance, minimum distance, dispersion, and KL divergence—formalizes whether a generated population spans relevant regions of the latent response space rather than merely concentrating around high-probability modes. This is a distinct evaluation philosophy from both prompt-engineered workshop artifacts and personalization-oriented persona conditioning (Paglieri et al., 3 Feb 2026).

Taken together, these evaluation regimes show that “quality” is not a single construct. In one setting it means time reduction and stakeholder utility; in another it means lower preference distributional divergence and improved human preference judgments; in a third it means support coverage and long-tail exploration. This suggests that proto-personas should be assessed relative to their operational role rather than by a universal metric.

6. Applications, limitations, and unresolved tensions

The application space described in these papers is broad but internally coherent. In Product Discovery, prompt-generated proto-personas are used during Lean Inception to support divergence and convergence, align stakeholders, and guide MVP scoping and feature refinement. In image generation, graph-mined personas serve as conditioning variables that personalize text-to-image outputs in zero-shot and few-shot regimes. In synthetic evaluation and UX exploration, Persona Generators function as “proto-persona factories” that produce broad sets of edge-case and long-tail users for downstream testing (Ayach et al., 11 Jul 2025, SI et al., 7 Jun 2026, Paglieri et al., 3 Feb 2026).

Several limitations recur. The Lean Inception case study tested only ChatGPT 4o-mini, identified domain generalization issues, and found that affective and behavioral empathy were uneven; accessibility and non-functional requirements were not consistently captured, and participants needed prompt familiarity and domain context. Persona Generators identify mode collapse, too many diversity axes, radius $0.7$1 mis-calibration, and over-optimization on stated preferences as practical pitfalls. ZIPP, while improving cold-start personalization, depends on mining and verbalizing user behavior at scale, so its own best-practice recommendations emphasize popularity-correct edge normalization, capped token budgets to reduce hallucinations, evaluation of persona diversity through CMMD and demographic equity through IPF reweighting, and privacy protection through filtering and de-identification before LLM verbalization (Ayach et al., 11 Jul 2025, Paglieri et al., 3 Feb 2026, SI et al., 7 Jun 2026).

A further tension concerns realism versus breadth. The Persona Generators framework explicitly prefers support coverage over density matching, whereas Lean Inception workflows favor contextual alignment and richness, and ZIPP optimizes alignment with observed aesthetic distributions. This suggests that proto-personas can be designed either to approximate likely users or to stress-test possible users, depending on whether the task is stakeholder communication, personalization, or robustness analysis.

Another persistent issue is the relation between interpretability and validity. All three approaches privilege editable textual descriptions over black-box vector profiles, which supports inspection and reuse. Yet each also recommends human oversight: stakeholder review in Lean Inception, domain-specific prompt grounding, and privacy-aware verbalization or post hoc evaluation in large-scale pipelines. The plausible implication is that proto-personas remain most reliable as hybrid artifacts—machine-generated or machine-mined, but situated within human validation and task-specific evaluation.