Prefigure
AI population testing, inside Claude

Know if it works before you build it.

Ask Claude to test your Figma prototype. Prefigure builds a population of AI users modeled on your audience, lets them use it for a simulated week, and brings the metrics — with a fix list — back into the chat.

Watch a sample run →Connect Claude
THEN, IN CLAUDE: “Test this prototype with 300 people who keep quitting habit apps.”
SplashREACHED300Welcome297Sign in270Onboarding 1 · Tr…209Onboarding 3 · Ea…183Onboarding 4 · Mo…169Choose your first…132Set your goal119Today100−61left at sign in33% set up a first habit“I got as far as asked to create an accountbefore being shown anything and thenclosed it.”INES · LEFT AT SIGN INLIVEMain Habits — onboarding
SplashREACHED300Sign in270Choose your first…132Today10033%
The cost of finding out

Validating one idea used to take a quarter. Now it takes an afternoon.

The usual way
Build something testableWEEKS OF ENGINEERING
Recruit 100 target usersWEEKS + INCENTIVES
Wait for D+77 DAYS, MINIMUM
Ask them whyINTERVIEWS
Months
per question
With Prefigure
Ask Claude 1 min→Confirm what it read 1 min→Simulate a week minutes→Read the report
An afternoon
per question
Illustrative for a small team · drawn to true scale, the Prefigure bar would be thinner than a pixel
No engineering

It runs on the prototype you already have. Nothing to build, deploy or instrument.

No recruiting

The population is generated to match your audience. No screeners, no no-shows, no gift cards.

No waiting on the calendar

Seven days of coming back — or not — simulated in minutes.

Why it works on day one

Skip the three slowest parts of user research.

01

Your AI already has the file.

Connect Prefigure to Claude once. Ask it to test a prototype and it opens your Figma file with your own access — every screen, and where every tap leads — then checks what it read with you before a single person runs.

  • Figma prototypes, the app you are building with Claude Code, or a live site
  • Buttons that do nothing are kept: they are unmet expectations
  • Nothing to export, no public link, no token — and nothing in your file is changed
figma.com/proto/kQ8vN2/Main-Habits 10 SCREENS · 23 TAPS2 LEAD NOWHEREVALUE: FIRST HABIT SET
02

People like your users. None recruited.

Describe who it is for. Prefigure samples a population across the ways those people actually differ — how skeptical, how rushed, what they use today — instead of 500 copies of one persona.

  • The split is decided in code first, then each person is written
  • Every person keeps the segment they came from
  • Meet all 500 before you spend a minute of simulation
Pragmatists 225 Early adopters 125 Skeptics 100 Innovators 50 Grace Oduya, 27 · nurse “If it takes a minute I’ll do it on my 3am break. Longer, no.”
03

Your metrics, and A/B, in an hour.

Define what success means — reach a screen, tap a button, pass a funnel, come back on day 7. Then A/B two versions with the same people and the same luck, so the only difference is the product.

  • Any metric built from what people did, with a confidence interval
  • Paired A/B: a winner, or an honest “too close to call”
  • Week-long retention counted from behaviour, never asked for
DAY 0 500 DAY 1 38% DAY 3 28% DAY 7 21% 7 DAYS OF BEHAVIOUR → ABOUT AN HOUR 1 DOT = 5 PEOPLE
What you get

One report. Where they left, why, and what to fix first.

Read a full sample report →
PREFIGURE REPORT DOLIKIM — ONBOARDING V3 · 500 PEOPLE · 7 SIMULATED DAYS · SEP 23, 2026 You lose 2 in 5 people before they’ve seen your product. Still using it on day 7 105 Reached value, then left 100 Left before reaching value 295 41% reached value TARGET WAS 60% 21% still there on day 7 COUNTED, NOT ASKED 64 could explain the promise OUT OF 100 18% would pay $4.99/mo LOW CONFIDENCE
A
The verdict, in one sentence

Before a single chart: what is going wrong and how big it is.

B
Every number opens into people

Every drop comes with the reason and the words of the people who left.

C
A fix list, not a slide deck

Ordered by what it costs you to ignore — each one re-runnable.

Can a simulation be trusted?

Simulated — and honest about it.

Counted, not generated

No model writes a number. Every metric is counted from what the population actually did, one tap at a time.

Confidence on every metric

Where people left: high. Whether they would pay: low. The report says which is which, on the number itself.

Calibrated to your usersSOON

Connect your analytics and Prefigure tunes the population to behave like the people you already have.

Questions founders ask first.

Do I need to build anything?

No. A clickable Figma prototype is enough — the same one you would show a colleague. If you are building with Claude Code, it can hand over the app instead.

Which AI does it work with?

Claude today — Claude desktop, claude.ai and Claude Code. Add one address and ask in plain words. ChatGPT is next.

What happens to my design?

Claude reads it with your own Figma access and sends Prefigure only what the study needs: the screen images, their text and where each tap leads. Nothing in your file is changed. Images are kept for 30 days, and you can ask Claude to delete a study at any time.

How is this different from asking an AI for feedback?

Feedback is an opinion about a description. Prefigure puts 500 people through your actual screens, one tap at a time, and counts what they do. What they say comes attached to where they were when they said it.

Can I trust the numbers?

Trust the direction and the size of a gap. Every metric carries its own confidence, and behaviour is counted, never generated. Run version A, change one thing, run version B — the difference is the finding.

What does a study cost?

Watching the sample is free, and every connection runs three studies a month for free. Pro — no monthly limit, 30-day retention and A/B on every study — is in early access.

Find out tonight.

Ask Claude. Have a report before you would have finished writing the recruiting email.

Watch a sample run →Connect Claude