3Edit Web Builder iconVibe coding guide

How to test a vibe-coded app before launch

A plain checklist that works with any tool.

Test your AI-built app the way a new visitor would before you launch: a plain checklist for pages, forms, sign-up, payments, phones and the production build.

About 7 minutes to readLast updated

By Humayun Rashed, founder of 3Edit

Why test an app the AI built for you

You described what you wanted, the AI built it, and the preview looks right. A preview that looks right isn’t the same as an app that works for every visitor, though. AI coding agents build quickly and change things quickly, and a change in one place can quietly affect another.

A short, planned test run before launch finds the problems your first customers would otherwise find for you. You don’t need to be a developer: most of it is using your own app the way a stranger would, plus a few checks a developer would also make.

The pre-launch checklist

Work through these in order and tick each one off. When something fails, describe the exact steps and what happened to your AI agent, fix that one thing, then carry on.

CheckWhat to doWhy it matters
Every pageOpen every page from the menu and the footer, and press every link and button.Broken links and empty pages are easy to miss while you are building.
Every formFill in each form with good details, then with mistakes: an empty field, an email without an @.A form should accept good input and explain clearly what went wrong with bad input.
Sign-up and sign-inCreate a new account, sign out, sign in again and try a wrong password.It is often the first thing a real visitor does, so it has to work first time.
Who can see whatAs an ordinary member, try to open admin pages and another member’s data.Hiding a button isn’t a permission. The security basics guide explains why.
PaymentsPay in test mode with test cards, including one that is declined.Stripe’s testing docs explain that test transactions don’t move funds.
Phones and tabletsCheck small screen sizes, then open the site on a real phone.Chrome’s device mode docs call it an approximation and advise running your page on a real mobile device.
Empty and error statesVisit pages before there is any data, search for something that doesn’t exist, and try an address that doesn’t exist.Your first visitors see an empty app, so it should still make sense.
The production buildBuild the production version of the app and fix every error it reports.The Next.js production checklist suggests running next build locally to catch build errors.
Speed and accessibilityRun a Lighthouse report, then the W3C Easy Checks: keyboard access, page titles and image alt text.Lighthouse audits performance, accessibility and SEO; the Easy Checks are a quick first look at accessibility.

The 10-minute self-check

You can do this in any normal browser, in about ten minutes. Each row says what to do and what you should see. If you see anything else, that is your next fix.

Try thisWhat you should see
Sign up, log in, log out. Create a new account, log out, then log back in.Each step works first time. After logging out, private pages ask you to sign in again.
The main action, three times. Do the one thing your app is for (book, post, order) three times in a row.Three separate results, each saved once: no duplicates, no errors.
Refresh in the middle. Part-way through a form or checkout, refresh the page.Nothing breaks. You keep your place or start that step again cleanly, with nothing half-saved.
Double-click buy or submit. Click the button twice, quickly.One order or one entry, not two. Payment services plan for this: Stripe’s idempotent requests let a retried request run only once.
Go back after paying. After a test payment, press the browser’s back button.No second payment. You see the order, or a clear message.
A wrong or empty form. Send a form empty, then with a wrong email.A clear message beside each problem field, and nothing saved.
An expired session. Sign out in a second tab, then try an action in the first.You are asked to sign in again, and nothing is changed while signed out. OWASP advises that sessions time out and ask you to sign in again.
Phone size, then a real phone. Use your browser’s phone-size view, then open the site on a real phone.Nothing spills off the side, buttons are easy to tap and text is easy to read.
A slow network. In Chrome DevTools, open Network and throttle the connection with a slower preset such as 3G, then reload.Pages still load, buttons show they are busy, and a slow click doesn’t send anything twice.

The two test users check

One check deserves its own place: can one person see another person’s data? It takes five minutes and two email addresses.

  1. Create two accounts with different emails: person A and person B. Use a private window for B, so both stay signed in.
  2. As A, create something private, such as an order, a note or a booking.
  3. Open that item and copy its address. Look for the ID in it, often a number or a long code.
  4. As B, paste A’s address. Then open one of B’s own items and change the ID in its address to A’s.
  5. Still as B, try to edit or delete A’s item.
  • What you should see: “not found” or “no access” every time, and nothing of A’s changes.
  • Why it matters: changing an ID to reach someone else’s data is a known weakness, an insecure direct object reference. It is a common and serious gap: OWASP ranks broken access control first in its 2025 Top 10 of web application risks. The security basics guide explains the fix.

Test like a visitor, not like the builder

You know where everything is, so you click the right buttons in the right order. Visitors don’t. A few habits help you see your app through their eyes:

  • Use a private browser window. You arrive signed out, like a new visitor, with nothing remembered from your own sessions.
  • Write down what you tested. A short list of what passed makes it quick to test again after the next change.
  • Retest after each big change. A fix in one place can affect another, so rerun the parts of the checklist that the change touched.
  • Ask someone new to try it. A friend who has never seen the app finds what you no longer notice. Watch, and don’t help.
  • Report problems precisely. Give your agent the exact steps, what you expected and the exact error. If it keeps failing on the same problem, the guide to agents stuck in a loop helps.

The ready-to-launch rule

Twice, gracefully, separately, on a phone, with a way back. Launch when your core flows pass more than once, errors fail gracefully with a clear message, two test users can’t see each other’s data, you have tried it on a real phone, and you know how to go back to an earlier version.

What this checklist doesn’t cover

These checks catch the common problems an ordinary visitor would hit. Some things need more than a browser and ten minutes:

  • Business-logic flaws: rules that work but are wrong, such as a discount that can be used again and again.
  • A professional penetration test: a specialist trying to break in on purpose.
  • Load testing: what happens when many visitors arrive at once.
  • Legal and compliance checks: privacy, consumer and accessibility rules for where you sell.
  • When to ask a professional: before you launch an app that handles health data, payments at scale or children’s data, or where a fault could seriously harm someone, get an expert review.

How 3Edit helps you test

3Edit checks your work along the way, so you start your own test run from a better place. It doesn’t replace that test run: the checklist above is still yours to do.

  • A code check after every change. After an agent changes your site, a TypeScript code check runs, whichever engine you use: the 3Edit Agent, Claude Code, Codex or 3Edit Node. The activity feed shows labels such as “Checking code quality…”, “Checking TypeScript” and “Code check passed”.
  • The preview is checked too. The preview is updated, and the changed pages are opened and checked (“Checked the preview”). The agent can read the rendered page and take a screenshot of the preview (“Preview screenshot”).
  • Repairs by the engine you chose. When a check fails, the same engine tries to repair it; it doesn’t switch engine by itself. If it can’t finish, the work is kept and labelled “Saved — needs repair” or “Needs continuation”, with the note “Ask me to fix the type errors, or use History to go back.” You can ask it to continue, or go back.
  • Screen sizes in the preview. The preview toolbar has “Desktop”, “Tablet (768 px)” and “Mobile (390 px)”, with “Open in browser” and “Reload preview”. These are screen sizes inside the preview, not real phones, so test your site on a real phone as well.
  • Show the agent what you see. In the chat, “Attach an image” lets you send a screenshot of a problem and ask the agent to look at it.
  • Project History. 3Edit keeps the last 25 versions of your site’s source, including images, on your own computer, labelled “Initial build”, “AI edit”, “Repaired” or “Restored”, with their “Changed files” and the “Current” version marked. Press “Restore” to go back to one; 3Edit first saves an “Undo point before restore”. A restore brings back your site’s files, not your database contents or keys. With GitHub connected, saved versions are pushed to your own GitHub too.
  • Access rules tested as different people. On your own Supabase, 3Edit turns on row level security and tests its access rules as a visitor, a signed-in user, the owner, another user and an admin, then rolls the test data back. Your two test users check confirms the same thing from the outside, in the finished app.
  • Next, security. Before you launch, read security basics for vibe-coded apps, and see how building works on the Web Builder page.

About the author.

Written by Humayun Rashed, founder of 3Edit. He also runs 3Edit’s free in-person AI website building and vibe coding course in London.

Questions people ask.

Use it as a new visitor would: every page, link and form, sign-up and sign-in, payments in test mode and small screens, then a real phone. Finish by building the production version and fixing anything it reports. The checklist on this page covers each step.

No. Most of the checks are using your app the way a visitor would. When something fails, give your AI agent the exact steps, what you expected and the exact error, and ask it to fix that one thing.

Use Stripe’s test mode, with its test keys and test card numbers: test transactions don’t move funds. Switch to live keys only when you are ready to take real payments.

It’s a good first check, but a screen-size preview only approximates a phone. Chrome’s own docs advise running your page on a real mobile device, so open your site on a real phone before launch.

Use two test accounts. As person A, create something private and copy its address. As person B, open that address, and change the ID in one of your own addresses to A’s. You should see “not found” or “no access” every time.

After every agent change, 3Edit runs a TypeScript code check on every engine and checks the changed pages in the preview. It doesn’t test every page and every interaction for you, so run the checklist yourself before launch.

Go back to a version that worked. In 3Edit, Project History keeps the last 25 versions of your site’s source on your computer; press Restore, and 3Edit saves an undo point first. A restore brings back your site’s files, not your database contents or keys.

Start vibe coding today.

Download 3Edit for Mac or Windows and build your first professional website in minutes.

Download now
macOSWindows