Synthetic Users as a Design Tool
Every designer knows the moment when you’ve gotten too close to your own work. You’ve been in the same flow for days, you’ve run it past everyone within arm’s reach, and somewhere along the way you started designing for yourself instead of for the person who’ll actually use it.
I hit that wall on a solo project redesigning the first-time application for a government service.
The goal was easy to say and hard to do: make it understandable. This is exactly the kind of thing a government agency should make easy. Someone shows up ready to do something good, and you want them to get all the way through, not give up halfway through the form because the red tape wore them down.
But the only people I could talk to were people who had already made it through. And it’s a strange thing to go looking for the other group. You can’t post a flyer that says, “Are you someone who might need this someday and doesn’t know it yet?” The person I most needed to design for was the person I had the least access to.
So I started where I could. I interviewed a handful of recent applicants about where they got stuck, where they second-guessed themselves, what they had to call someone to ask. The people I most needed to reach, the ones who took one look at the form and quietly gave up, were never going to show up in my interviews. But successful applicants remembered being confused, and that got me close.
Then I built a synthetic user out of what they told me: a first-time applicant I could sit with, ask questions, and bounce ideas off whenever I wanted.
I used Claude, but you could do the same thing with a custom GPT. More important than the tool you use is understanding what this kind of user is good for, what it isn’t, and the one rule I don’t break.
Real People First
I have to talk to real people before I build a synthetic user. Always. I can build a persona out of my own assumptions, but then I’m just reinforcing what I already believe, which is a bad way to design anything.
What the synthetic user does is scale a perspective I earned honestly. I could only get a few recent applicants, and “a few” doesn’t stretch across weeks of iteration. The synthetic user took what they taught me and made it available on demand, so that, as a team of one, I could keep poking at my blind spots.
A synthetic user doesn’t produce a new perspective. It reflects one back at you. Feed it your assumptions, and it hands them back dressed in user-sounding language, which feels a lot like validation but is really just an expensive way to talk to yourself. Feed it real research, and it reflects that instead.
Though, even calling it a “mirror” is generous. An actual mirror won’t make things up, but ask a synthetic user something your research never covered, and it will improvise something confident and plausible. The reflection is only trustworthy inside the frame of what real people said. Past that edge, it’s inventing.
One Synthetic User Per Purpose
The version of this I’m proudest of came from a company I worked at before Don’t Panic Labs. I built a synthetic user on top of our ideal customer profile, or ICP. That’s the portrait of who actually buys: how big their operation is, who inside it touches the product day to day, and what those people are trying to get done with our product. All of that was research-backed, and it still wasn’t enough on its own, because the ICP was the same across everything.
What made it most useful was layering specific research on top for a specific purpose. I was working on two very different moments: launching a payments feature, and preventing churn through cancellation. Same ICP underneath, two distinct synthetic users, each shaped by the research I’d done in that topic area.
A synthetic user grounded in “our customers, roughly” tells you roughly nothing. One grounded in what people said while they were canceling catches things.
What They’re Good For
Working solo, the synthetic user keeps a user perspective in the room when no user can be. Colleagues give great feedback, but it’s best-practice feedback, not user feedback. And it’s easy to get myopic when you’re heads-down. To forget a perspective you already know, because you’re too far inside your own head to surface it.
It also sped up the product marketers I worked with. Instead of running a content idea through a full A/B test to discover the obvious problems, they could pressure-test it against a synthetic user first and walk into testing with a stronger starting version. This didn’t replace A/B testing; the test is still the arbiter. It meant what they brought to the test was sharper than what they’d have written alone.
Where They Don’t Belong
If I have an assumption I need to validate, I go find a real user. The moment I need to learn something I haven’t already uncovered, a synthetic user is the wrong tool. It can’t discover. It will gladly generate a confident answer it has no grounding for if you let it, but a fluent guess isn’t a finding. Asking a synthetic user to validate an assumption is asking the mirror to show you something that isn’t in the room.
I’ve seen people use them to generate assumptions and hypotheses to test, which I think is a great use. Then you go validate those with a real person.
The Rule I Don’t Break
Everything the synthetic user knows has to trace back to something a real person told me. Build it from research, keep it specific, and use it to sharpen the questions you take to real people. That’s what keeps me from mistaking my own reflection for a finding.
