Last week a paying customer wrote to me about the app she had built with JustCopy. Two lines: the photo upload doesn’t work, and the things I type don’t show up.
I read her whole conversation with the agent. It had written “Done” five times. Every one of those was about a page it had never opened. It had generated the page, checked that the code compiled, and declared victory. She was the first person to actually use it.
That is how every AI app builder works today, including mine until this week. Generating an app and checking that an app works are different jobs, and we all do only the first. The person who finds out whether it works is the owner, or worse, the owner’s first customer.
The rule change
An app now gets used before the agent is allowed to call it done.
Here is what happens after every build in the studio:
Build. You type one sentence. The app is on screen, usually well under a minute.
Plan. A short list of checks is written from the app itself, in plain words: “Add an entry”, “Attach a photo”, “Delete it”, “Reload and check it is still there”.
Use. The app pane gets a breathing blue border, the same signal you see when an AI takes over a browser, and a cursor starts using your app. It types, presses Save, drops in a sample photo, deletes an entry, reloads the page.
Judge. After each check, the words on the page are compared with what should be there. A tick or a cross lands in the chat with a plain reason: “The text box resets after every letter.”
Fix. A cross goes straight back to the agent with the failing step and the page errors. It fixes the problem and only the failed checks run again. This costs the owner nothing.
The border and the cursor are deliberate. Nothing happens off screen. You watch every tap land, and you can see the moment something fails.
What it found on day one
I ran it on a fresh app and on a copy of the customer’s app. Both of her complaints were found and fixed without a person reading the transcript. Three things it caught:
A text box that reset after every letter. The agent had defined the app’s parts inside the page’s own drawing routine, so the box was rebuilt on every keystroke and only the first letter survived. The test said exactly that, and the fix landed a minute later.
A delete that did not stick. Entries were created with a temporary id, so deleting one removed the wrong record. No visitor would have reported this. They would have left.
My own production bug. The first recording of the launch video failed the photo upload check. The cause was on my side: the service that stores app files had lost a permission, so every photo upload in every generated app had been failing. That is the real reason her uploads never worked. The test found it before a second customer did.
The rules it follows
It never signs in or out. The preview shares the owner’s session, so a test that signed out would sign the owner out too.
Everything it creates is removed when the run ends: the sample entries, the sample photo.
It does not pay, email or text anyone. Checks that would need a card, an inbox or a phone are listed as skipped, with a reason.
Every action is visible, and the chat says what was tried and what was seen.
What it does not do yet
Dropdowns, date pickers and checkboxes are not driven yet, only typing, clicks, uploads and reloads. The checks run in the studio, on screen; a silent rerun before publish and a daily sweep of published apps are next. And an outcome that needs eyes, a photo that appears but looks wrong, still needs yours.
Why I am writing this down
Because “Done” is the most expensive word in this product. Every time the agent said it without checking, a customer paid for the gap with their time and their trust. Making the app prove itself in front of the owner is the only fix I have found that does not depend on the model being right.
Type one sentence, watch it build, watch it prove itself, then one tap puts it live. Free to start: