How to prompt an AI study builder (examples that work)

What a study builder actually does with your prompt

Here's the flow inside Sera, because the prompt advice only makes sense against it.

You submit a prompt, pick a template, and optionally attach artifacts — a URL, a Figma prototype, a PRD, screenshots. The builder reads all of it and confirms what it understood. Then it drafts a set of research goals in the background and hands them to you for confirmation.

You edit the goals. Then the builder authors the full study around them: the screener, the tasks participants perform on your stimulus, and the interview guide the AI moderator follows.

Every part of your prompt maps to a stage of that pipeline. The stimulus link drives what the tasks get built around. The problems you name shape the drafted goals. The end-questions shape the interview guide — and later, the synthesis. Write the prompt with that mapping in mind and the draft comes back close to right on the first pass.

The three parts of a prompt that works

1. A link to the stimulus, if you have one. When there's a thing to react to — a live URL or a Figma prototype link — this is the most valuable line in the prompt: it tells the AI exactly what participants will see and what to build tasks around. No stimulus yet? Skip the link and lead with goals. The AI is just as capable of generating an interview study around hypothetical situations — how people handle a scenario today, how they'd react to an idea you describe — as it is around a page.

2. What to look for. Name the problems you suspect or the risks you want probed. "I think the pricing tiers read as enterprise-only" is a gift to the goal-drafting step — it becomes a testable goal instead of a guess.

3. The questions you want answered by the end. Not interview questions. Outcomes. "By the end I want to know whether people understand what this costs" gives the builder a finish line, and it gives the synthesis a spine.

One line worth adding to almost any prompt: ask the AI to look for anything else worth evaluating — gaps or discovery opportunities you didn't name. You wrote the prompt from inside your own framing; that sentence invites the builder to flag what the framing missed.

Notice what's missing: audience specs. Don't put recruiting or targeting instructions in the prompt — who participates is handled by the screener and your recruiting setup, not by prose in the brief. Prompts that try to direct recruitment ("interview people who churned last month") ask for something the prompt can't deliver, and they crowd out the parts that matter.

Example prompts you can adapt

Testing a live page:

Test our pricing page: https://yourapp.com/pricing.
I think the three-tier layout hides the annual discount, and the
"Contact sales" tier makes the whole page feel enterprise-only.
By the end I want to know: which plan people think is meant for
them, whether they notice annual pricing at all, and what would
stop them from starting a trial today.

Testing a Figma prototype:

Test this checkout prototype:
https://www.figma.com/proto/aBc123XyZ9/checkout-redesign?node-id=1-2045&starting-point-node-id=1%3A2045
Look for problems in the shipping step — I suspect the address
form and the delivery-option cards compete for attention, and
that the order summary disappears too early.
By the end I want to know: where people hesitate, whether they
understand what flexible delivery costs, and whether they'd
trust this checkout with a card number.

Testing a flow end to end:

Test our signup flow: https://yourapp.com/signup.
Watch for confusion on the workspace-naming step, and whether
the empty dashboard after signup gives people anything to do.
By the end I want to know: what people expect to happen right
after signup, where they'd give up, and what one change would
get them to their first project fastest.
Also flag anything else in the flow you think is worth
evaluating that I haven't mentioned.

No stimulus yet — exploring a problem space:

We're considering building an expense-approval feature for
team leads. There's nothing to show yet.
Interview team leads about how they approve expenses today:
what tools they use, where the process stalls, and what a
rejected expense actually costs them in time.
By the end I want to know: whether approval delays are a real
pain or a minor annoyance, and what would make them switch
from their current process.

Same skeleton every time. Link if you have one, goals if you don't. Suspicions. End-questions. Three to six sentences, and the builder has everything it needs to draft goals worth confirming.

For prototypes, paste the /proto/ share link from Figma's Share dialog — that's the playable version participants can actually click through. A /design/ link points at the editor, not the prototype.

Prompts that fail

The failures are more consistent than the successes.

The vague ask. "Test our app." No link, no suspicions, no end-questions. The AI must guess all three, and it will — confidently. You'll get a polished, professional draft of a study about the wrong thing. Fix: pick one page or one flow, link it, and say what worries you about it.

The kitchen sink. "Test the new dashboard, and also the pricing page, and also see what people think of the brand." Three studies wearing one trench coat. The real constraint is time: around 30 minutes is a good interview length, and that usually means one or two topics and their associated stimuli. Spread it across three, and every topic gets a third of the depth. Fix: focus each study. AI-run studies are fast and cheap — run the others next.

The leading frame. "Confirm that users love the new dashboard: https://yourapp.com/dashboard." A verdict seeking a citation. The AI drafts from your framing, so a prompt built to confirm produces goals built to confirm. Fix: name the doubt instead. "I think the new dashboard buries the export feature — find out what people can't find."

One of these is worth memorizing: the leading frame. "Confirm" and "validate" are the most expensive words you can put in a research prompt.

A vague prompt doesn't produce a vague study. It produces a specific, confident, professional-looking study about the wrong thing.

The five-minute review

The prompt gets you a draft. A short review gets you a study — and in practice the draft comes back tight. Coverage against your prompt is usually very close on the first pass; the review is a double-check, not a rescue.

Read the drafted goals against your end-questions and confirm the tasks land participants on the part of the page or prototype you flagged. If something's off, edit it like a doc.

The review is also where the best builds happen. A useful one: ask the AI to add some quantitative evaluation questions — ratings and comparisons that give the synthesis numbers to work with alongside the open-ended answers.

Then launch. A study now costs a day, not a quarter — so the right response to an imperfect prompt is usually to run it, read five transcripts, sharpen the prompt, and run it again. Write the best brief you can. Don't let perfecting it become the research that never ships.

Frequently asked questions

What should I include in a prompt for an AI study builder?

Three things: a link to the thing you want tested (a live URL or a Figma prototype link) — or, if nothing exists yet, the situation you want explored — the problems or risks you want the study to probe, and the specific questions you need answered when it's done. You can also ask the AI to flag anything else it thinks is worth evaluating.

How long should a prompt for an AI research tool be?

Three to six sentences plus a link. Shorter and the AI fills the gaps with guesses; longer and you're usually writing the study yourself. One sentence for the stimulus, one or two for what to look for, one or two for the questions you want answered. Extra context belongs in attached artifacts, not prose.

Should I paste a URL or a Figma prototype link in my prompt?

Paste whichever is the real thing you want reactions to. A live URL tests what exists today; a Figma prototype link tests what you're about to build. Either way, the link is the most important line in the prompt — it tells the AI exactly what to build tasks around.

Should I write the interview questions in my prompt?

You can. If you already have a script or draft questions, include them — the builder generates research goals from them and uses them to inform the full study. What matters most is stating the outcomes: the questions you want answered by the end. Give the builder both and review the drafted guide it produces.

Can I paste my PRD or product spec as the prompt?

Attach it, don't paste it. A spec describes the product; a prompt names the test. Attach the PRD as an artifact so the builder can mine it for context, then write three sentences on top: the link, what you're worried about, and what you need to know. Specs alone produce broad, unfocused drafts.

What happens after I submit my prompt to an AI study builder?

The builder reads your prompt, template, and any attachments, confirms what it understood, and drafts a set of research goals in the background. You review and edit those goals before anything else is written. Once you confirm them, it authors the full study — screener, tasks, and interview guide — around your stimulus.