One message, start to finish
There is no magic step in here. This is the whole loop, in the order it happens.
What one request actually is
You send one message, in your own words. It gets picked up, your app gets built or changed, and you watch it happen as it goes — not a progress bar, the actual work.
Each project is one long conversation. Not one per feature — one. That is deliberate: it means your twentieth message still knows what you decided in your first, so you never have to explain your business twice. It also means one request at a time, and a new one waits for the current one to finish.
Then it has to prove it works
Saying it is finished is not the same as being finished, so it does not get to decide. Every change has to get past our checks before it can go anywhere:
Checking for mistakes
Your app has to actually run. This catches the single largest category of confidently-wrong AI work.Tidying up
The standards your project already agreed to, applied to the new work the same as the old.Testing
Everything the app is supposed to do, tried in full — including the parts that were working yesterday and had better still work today.
These run somewhere sealed off, with no reach into anything but your project. That matters more than it sounds: they are running work an AI produced seconds earlier, and we treat that as untrusted rather than trusting it because it came from us.
Nothing goes live until every one of them passes.
When a check fails
It sees exactly what went wrong and tries to fix it. It gets three attempts, and if the checks still will not pass it tells you so plainly rather than putting something broken in front of your customers.
Fixing our own mistakes is free. It is not deducted from your credit and does not count toward your monthly limit. The reasoning is simple: the checks are ours, so failing them is our problem, not something to bill you for while you watch. This is enforced in the data model, not in a policy document — our own fixes are structurally excluded from both.
Then it saves your work and puts it online
“Repair tickets for my mobile shop so customers can see when the phone is ready”
Work that passes is saved as a version you can return to at any time. Your app goes live in its own space, with your information kept apart from anyone else’s, at its own address.
Saved versions are why you can experiment. Every request leaves one, including the ones that did not work, and you can go back to any of them.
Going back moves forward. Returning to an earlier version adds a new step that brings your app back to that state — it does not rewind or delete anything. The work you left is still there, so you can jump forward to it again five minutes later when you change your mind. A time machine that destroys the present is not one you would use twice.
One thing to be clear about: going back brings back your app. The information stored in it is captured at every version, and the interface tells you which versions hold it, but putting that information back is not wired up yet. We say so rather than letting you find out.
What each request costs you
Every request shows an estimate of what it cost, in plain currency, on the request itself. Nothing comes off your balance without a line explaining it.
That figure is an estimate, drawn against credits we granted you. It is not an invoice, and no money is charged against it.
How your information is kept safe, in detail — and what is not done yet