Reviewing a Site Through Synthetic Users

A ready-to-use prompt plus instructions for using it. It runs any live webpage - a landing page, product page, pricing page, signup form - past 5 fictional but realistic users from your target audience. Each one walks through the page top to bottom and reports: what caught their attention, what stayed unclear, which doubts never got resolved, and where they would have closed the tab.

This is meant to help you think through the page before showing it to real people or rewriting the copy. You can copy the prompt below as is, replacing the text in square brackets [...] with your own.

What you get out of it

• A map of the page: everything on it, in order top to bottom
• 5 personas (fictional users) with different attitudes - from interested to skeptical
• For each persona, a walkthrough of the page with notes: what they liked, what confused them, what doubt remained, where they would have left
• A summary analysis: what repeated across several people, where they disagreed, what was surprising
• A verdict on each block of the page: works / trips people up / loses the person
• Top 5 copy edits, each with an explanation of why
• A mandatory warnings block: what you can trust in this simulation, and what you can't

What you need beforehand

• Claude Code, Cursor, or a regular Claude chat (whichever you prefer, it's the same prompt)
• A link to the live page you're reviewing. The page must be accessible without a login wall (no enter-your-password screen), otherwise the assistant won't be able to read its text.
• An understanding of your target audience: who these people are, what they do, what industry, how technical they are. You need to write this into the prompt - the assistant shouldn't invent your audience for you.

Terms you'll run into (explained once):

• ICP (Ideal Customer Profile) - the ideal customer profile: who this product/page is actually for.
• Synthetic users - characters invented by the assistant who play the role of your customers. They're not real people or data, just a plausible rehearsal of a reaction.
• Persona - one such fictional user with their own role, context, and attitude.
• Conversion - the share of people who complete the target action on the page (clicked the button, signed up, bought). Conversion leaks - meaning people are leaving before reaching the action.

The prompt (copy it whole)

Review this page with synthetic users. This is a calibration, not research:
you play my customers so I can think about the page before showing it to real people.

Page: [paste the link to your page]

My target audience (ICP): [describe: who these people are, what they do, what industry,
company size and stage, how technical they are, key context].

The page's goal action: [e.g. sign up / request a demo / buy / join the waitlist].
What worries me most: [e.g. is the price clear / do non-technical people feel out of place /
is it clear who this is for]. If nothing worries you, write "none".

Do this:

1. First read the actual text of the page at the link, block by block, top to bottom
   (hero, who it is for, program/how it works, price and dates, about the author/trust,
   testimonials, FAQ, buttons). Do not invent anything. If the text cannot be read
   (login wall, page did not load) - stop and tell me, do not make up text for the page.

2. Build 5 distinct personas from my audience. They must genuinely differ (context, pressure,
   level of skepticism), not be one person under different names. Mandatory:
   at least one - an open skeptic ("just another course/tool, probably a scam"),
   at least one - with a real barrier (budget, language, access). Do not make them all warm leads.

3. Walk each persona through the page top to bottom, 8-12 steps. Speak in their real voice:
   people doubt, skip around, contradict themselves, abandon the tab halfway.
   If it reads like a smooth marketing interview - rewrite it rougher, real life is not like that.

4. In each walkthrough, explicitly tag:
   [LIKES] - what genuinely landed (quote the specific phrase/block)
   [UNCLEAR] - friction, jargon, vague outcome
   [DOUBT] - an objection that stayed, even if the person is interested overall
   [EXIT] - the spot where the person would close the tab (if any)

5. After all personas, give a summary analysis:
   - what repeated across two or more people (shared strengths and shared friction points)
   - where personas diverged and why
   - what was surprising (something the page owner probably did not expect)
   - a verdict on each page block: works / trips up / loses the person
   - top 5 copy edits, each with one line on why
   - overall verdict: the page conveys value to my audience / partially / no, with reasoning

6. End with a mandatory block "WARNINGS - READ BEFORE YOU CHANGE ANYTHING":
   - what the simulation probably got right (known pains of this audience)
   - what it almost certainly got wrong (real willingness to pay, internal budget approval,
     how the author comes across in person, trust in testimonials, current price sensitivity)
   - what you still must validate with real people and real analytics
     (funnel data, heatmaps, session recordings, actual customer interviews)
   - red flags: if someone's line sounds suspiciously smooth and "marketing-y" -
     the model is playing along, do not trust it

Write in English. Persona lines - natural spoken English, not stiff or corporate.
No em dashes, no emoji.

How to use it

1. Grab the link to the page you want to check. It can be your landing page, pricing page, signup form - or even a competitor's page, to see it through your customers' eyes. The key requirement: the page must be publicly accessible, with no login screen.
2. Describe your audience in your own words in the [...] field. The more specific, the better. Not "managers," but "product managers at enterprise software companies, 5+ years of experience, paying for training out of their own pocket, not the company's, not technical, don't code." If you leave it vague, the assistant will invent an audience, and the whole review will miss the mark.
3. Fill in the page's goal action and what worries you. This points the personas toward the spot where you suspect a problem. If nothing in particular worries you, say so.
4. Paste the whole prompt into Claude Code, Cursor, or Claude chat and send it.
5. The first thing the assistant will do is read the actual text of your page. If it tells you the page won't open (login wall, blank page), give it the page text manually: copy the page's content and paste it into the chat. Without the real text, the review is fiction, and the assistant will say so honestly.
6. How to read the results. Don't fixate on individual lines - they're fictional. Look for repetition: if two or three different personas trip up in the same spot, that's a signal. The most useful parts are the verdict by block (works / trips up / loses) and the top 5 copy edits. Start there.
7. If you want it faster or deeper, ask: "run the personas as parallel agents." The assistant will then read the page once and hand it off to several helpers at the same time, then compile a combined review. Same result, just faster.

Important notes about this review

This is a calibration, not research. You're getting a plausible rehearsal of a reaction, not data. Never make launch, ad budget, or pricing decisions based on a single review like this.

Synthetic people are good at guessing an audience's known pains, but they almost always get the important part wrong: whether the person will actually pay, whether their budget gets approved internally, whether they'll trust the testimonials, how you come across in person.

Use it to find weak spots in your copy and prepare questions - then validate everything with real people and real analytics (funnel data, heatmaps, session recordings, customer interviews). And don't use this technique for sensitive audiences where lived personal experience matters (patients, vulnerable groups) - there the model will produce plausible-sounding nonsense.