Use this seriously — and call out when it's lying to us.
Define a persona, hand it to Claude, and run real research before you run real research — and learn when to trust it, and when it's lying to you.
Pre-test every design decision against 5 personas in 30 minutes. No scheduling. No recruiting. No no-shows. The leverage is real — but so is the trap.
Sharpen your thinking, catch blind spots, pressure-test a flow — all before you book a single real interview.
Teams who skip real research because the AI "already told them" what users think. Three weeks of fake confidence, one hour to be humbled.
Use this seriously — and call out when it's lying to us.
An AI persona you define in rich detail, then interview or usability-test. Claude plays the character. You play the researcher.
Use Claude's reasoning about a character to generate hypotheses and pressure-test assumptions — with a clear research purpose. Not creative writing.
A persona becomes a system prompt; you run a scripted conversation. The "agent" plays a user — not a software builder.
Run the same interview with 5 personas. Iterate the prompt. Re-run. The leverage compounds with every cycle.
Same Claude Desktop. Same persona. Different research moments — one for hypotheses, one for friction.
Open-ended questions about motivations, expectations, fears. Run before the persona sees the product. Surfaces mental models, hesitations, trust signals.
Show a screen or real URL. Narrate a task, observe think-aloud reactions. Run inside the product via Playwright. Finds friction, labels confusion, exposes edge cases.
Two stops, one journey — Claude Desktop runs both, switching modes seamlessly.
Thin personas give you thin opinions. Rich personas give you specific, surprising, actionable reactions.
"A 30-year-old woman in Riyadh. Shops online. Middle income."
Generic, safe, averaged opinions. Nothing sharp. Nothing you didn't already guess.
"Sara, 34, Riyadh. Marketing manager, mother of two. Got burned last year by a fake Instagram seller — since then she only buys from big names."
Specific reactions: "the missing review count worries me", "I wouldn't trust a site that asks for payment before showing the return policy."
Recent experience · emotional baseline · preferences. The last three fields are where synthetic users earn their keep.
This isn't a ChatGPT gimmick. Industry research leaders are actively publishing about synthetic users — both sides of the debate.
"Synthetic Users: If, When, and How to Use AI-Generated Research." Not dismissing, not endorsing. NN/g publishing on this signals the field is taking it seriously.
"How Synthetic Customers Bring Companies Closer to the Real Ones." Treats synthetic users as legitimate augmentation of customer research.
Gartner-cited tool. Claims 85–92% parity with real interviews on thematic overlap and depth of insight. Parity numbers exist — that alone makes it measurable.
"The Synthetic Persona Fallacy" + "The Challenges of Synthetic Users." The research community is actively wrestling with this — live debate = real topic.
A healthcare design team's three-week synthetic research practice — flattened in a single hour with real humans.
Copy this structure for any persona. The fields at the bottom — recent experience, emotional baseline, preferences — are where the signal lives.
## Sara Al-Otaibi, 34, Al-Malga district, Riyadh Demographics: Marketing manager at a mid-size agency, 6 years in. Mother of Layan (7) and Faisal (4). Bilingual — code-switches Arabic/English. Context: Hybrid 9-to-5. Online shopping = mostly Namshi for kids' clothes and Carrefour Online for monthly groceries. Has never tried panda.sa. ## Where the signal lives ↓ Recent experience: Last Ramadan, burned by a fake perfume seller on Instagram (SAR 400 via bank transfer, vanished). Since then: no bank transfers, no "new sites", checks for visible consumer-protection cert. Emotional baseline: Tired. Doing this between two meetings, on her phone, with her 4-year-old pulling her sleeve. Patience for forms = zero. Reads 2- and 3-star reviews first (5-stars lie, 1-stars are drama). Preferences: Saved addresses non-negotiable. Apple Pay, then Mada. Abandons cart immediately if surprise delivery fee at final step, or if forced to create an account before seeing the total.
Online grocery store of Panda — one of Saudi Arabia's largest supermarket chains. Two missions, one real site, both before you ever talk to a real shopper.
Three reference personas to reuse — or invent your own. Three different mental models, three different sets of friction.
Three different mental models — three different sets of friction. That's the whole point of running multiple personas.
Copy this. Swap the persona. Swap the URL. Run it. Same template works for any product, any task.
You are [PERSONA NAME]. Stay in character throughout. [PASTE RICH PERSONA HERE] I want you to do a think-aloud usability test on this site: URL: https://panda.sa/en Task: Find essential groceries for a family of four for the week (milk, bread, rice, vegetables), add them to cart, review totals, and attempt to proceed to checkout. Visit the site using Playwright. At EVERY screen: 1. Describe what you see (in [PERSONA]'s voice) 2. Say what you'd click next and why 3. Call out anything confusing, frustrating, or surprising 4. Note anything that would make you abandon the task Take a screenshot at each key step. # IMPORTANT - Dismiss the initial login popup with the Escape key - Stop at the final checkout login wall — do not create an account - Stay in character. Don't summarize "what a user might think" — BE [PERSONA].
Two activities. Group work. Same Claude Desktop. Different research moments — interview first, usability test second.
One app, two modes. Claude Desktop handles both. But knowing what NOT to use it for is the difference between a sharp researcher and a fooled one.
Perfect the AI Interviewer you built in Session 3 into a real research tool a junior researcher could open Monday morning and use.