Four steps, about five minutes, no recruiting.
This is the exact loop the homepage shows, walked through slowly, one step at a time: what you type, what the model actually does with it, and what you get back before you've finished your coffee.
Everything a real check needs, built in.
Skip the recruiting entirely
Describe your audience once. The panel forms immediately, no screener, no incentive budget, no waiting on availability.
Ground it in what customers actually said
Paste in reviews, tickets, or interview notes. Answers trace back to real language instead of pure guesswork.
See the bias before you trust it
Every panel ships a visible diversity check against a real Census age baseline, not a claim taken on faith.
Push further without starting over
Ask follow-ups, retry a persona, or regenerate a narrow panel with wider constraints, all without losing your work.
Describe who you're building for.
One sentence is usually enough: the audience, the product category, and — if you already know it — the specific question you need answered. “Solo accountants at small Indian firms who still track invoices in Excel” is a real brief we use in testing, not a simplified example.
There's no form to fill out and no fields to configure first. The brief is the only required input — everything after this step is optional.
Ground it in reality — optional, but it changes the answers.
Paste in real reviews, support tickets, or interview notes about this audience. The text is chunked and embedded, then the most relevant excerpts are retrieved and handed to each persona before it answers your question — the same retrieval pattern real search and recommendation systems use, applied to grounding a character instead of ranking documents.
Skip this step entirely and the panel still works — it falls back to what the model already knows about the category. Ground it and you get answers that reference the actual language your real customers use, not a generic approximation of it.
Interview a panel built to disagree with you.
Every panel gets hard constraints, not a vague request for diversity: no two personas share the same combination of age range, income tier, and adoption stance, and at least one persona is actively skeptical or resistant to the product category — not just “has concerns.”
Answers stream in per persona as they're generated, not all at once and not in a fixed order — whichever persona finishes thinking first appears first. Once you have answers, ask a follow-up and get a fresh round without starting over.
I don't trust apps that need internet all the time.
If it saves me an hour a week, I'm switching today.
The panel acts. You stay in control.
Personas act
Each persona answers independently, in its own voice, without seeing what the others said first.
You supervise
Retry a persona that missed the point, or flag an answer that doesn't ring true. Every change is visible.
You explore further
Chat one-on-one with any persona, in plain language, whenever the synthesis raises more questions than it answers.
See the diversity check before you trust anything.
Every panel ships with a visible breakdown of its own composition — income tier, adoption stance, age spread — plus a direct comparison against a real adult-population baseline for age.
If a panel skews narrow, it says so, and a one-click “rebalance” regenerates with wider constraints instead of asking you to start over and rewrite the brief by hand.
5 personas, four lean the same way.
94% similar — flagged.
58% similar — flag cleared.
It doesn't end at step four. It loops.
A synthesis pulls out themes and real disagreements across the panel's answers, a suggested follow-up question shows up when the panel split close to evenly, and every one of those can send you back to step three with a sharper question — all in the same sitting, without losing the panel you already built.
If the diversity check flags a panel as too narrow, iterate regenerates it with explicitly wider constraints instead of silently reusing the same skew.
Answers come back from the panel.
Themes and real disagreements get pulled out.
Shown when the panel split close to evenly.
What “about five minutes” actually breaks down to.
These are typical ranges, not a guarantee — actual timing depends on which model provider answers your request, panel size, and whether you're grounding with real material. Every attempt falls back across multiple providers automatically if one is slow or unavailable, so a single stalled request rarely becomes a stalled panel.
What actually changes, step by step.
- ×Recruit for weeks before you learn anything
- ×One shot, re-fielding means starting over
- ×Panel composition is a black box
- ×A static report you wait on
- ✓Panel forms in under a minute
- ✓Ask follow-ups in the same sitting
- ✓Visible diversity check, every run
- ✓A loop you can rerun right now
Before you start one yourself.
No — one honest sentence about the audience and product is enough to start. You can always run a variant with a sharper brief once you see the first panel's answers.
Run through it yourself — it's free to start.
No card, no sales call, no waiting on approval.
Start a free panel