Synthetic user feedback

Personas that answer back.

Active Personas turn a static persona document into an LLM-powered stand-in for a real user. Ask it what it thinks of a screen, run a full heuristic evaluation, or send it to walk through your live website. Get feedback in the persona's own voice, on demand, without scheduling a single interview.

Get started See how it works

Built on the research prototype from "Active Personas for Synthetic User Feedback" (Simaremare & Edison, ICSOB 2025).

A public-transit app home screen, the kind of interface an Active Persona evaluates

@ingrid-1 (claude)

The buy-ticket button is clear, but I can't tell which zone I'm in until I tap through. That's a guess I don't want to make.

@claudio-1 (gpt)

I ran the 20-item heuristic pass: 3.4 / 5 on "visibility of system status". The map helps, the labels don't.

A team working together on laptops around a table

A Panel is a shared space for your team and its Active Personas — attach a screenshot, @mention a persona, ask what it thinks.

How it works

Write a Persona

Describe a target user: goals, tech proficiency, accessibility needs, tone, background. Plain text, no model involved.

Activate it

Bind the Persona to an LLM configuration and an impersonation approach. The same Persona can run against several models to compare tone and criticality.

Talk in a Panel

A Panel is a shared chat space for your team and one or more Active Personas. Attach screenshots, @mention an AP, ask what it thinks.

Evaluate or $try

Run the built-in 20-item Nielsen heuristic questionnaire, or send an AP to drive your live site step by step and report what it did, saw, and felt.

What you can do

Panels

One primitive for every conversation: one user and one AP, a whole team with several APs, or a broadcast to all of them at once.

Multi-model comparison

Instantiate one Persona against Gemini, Llama, GPT, and Claude side by side. The research shows the model shifts how critical the feedback is.

Heuristic evaluation

The paper's 20-statement, Nielsen's-10-heuristics questionnaire, answered with a Likert rating and a rationale per statement, comparable across personas and models.

Live web interaction

An AP drives a real browser through your site, screenshot by screenshot, using a registered credential bound to its allowed origins, then reports back in persona.

Reports

Export a Panel as structured JSON: the screenshots shown, the questions asked, every reply, and the Likert ratings, scoped to a whole Panel or a single run.

Audit trail

Every action is logged, every agent turn has a verifiable hash-chained process trace, and credentials are redacted before anything is written down.

Grounded in research

The concept comes from a design-science study that had eight Active Persona instances (two personas across four LLMs) answer a 20-item Nielsen's-heuristics questionnaire against a transit app, then compared that synthetic feedback against real user interviews and app-store reviews. The finding: Active Persona feedback aligns well with human feedback, and both the persona definition and the choice of LLM measurably change how critical it is.

Simaremare, M. & Edison, H. (2025). Active Personas for Synthetic User Feedback: A Design Science Study. ICSOB 2025 / LNBIP 574.

Start a conversation with a persona

Create an account, write a persona, and ask it what it thinks. It answers in minutes, not weeks.

Create an account

Two people reviewing interface sketches next to a laptop

This application is part of a research project supported by