Skip to content
afterbuild.dev
$ git log --oneline | wc -l

Sometimes the answer is: start again.

Most apps built with an AI builder are worth repairing. The failures that stop them reaching production are usually configuration rather than code, and configuration is cheap to fix. But not always — and a repair sold on a project that should be rebuilt is the most expensive mistake available to either of us.

Six signals. Three of them mean repair; three mean stop. Where we have a measured frequency it is quoted with its sample; where we do not, the page says so.

SIGNAL 01still repairable

One file carries most of the application

Measured: the median largest source file is 21 KB in Lovable projects and 28 KB in Replit ones — roughly 600 and 800 lines. Half of all projects have at least one file over 20 KB.

Not on its own a reason to start again. A large file is tedious, not fatal, and splitting it is ordinary work.

SIGNAL 02still repairable

There are no tests, anywhere

Measured: 30% of Lovable projects have at least one test, 12% of Replit ones, 4% of v0 ones.

Also not a reason by itself — but it changes the price of everything else, because nothing can be changed safely without first proving what currently works.

SIGNAL 03points at rebuilding

The same feature has been rewritten by the AI several times and each version is still in the repository

Measured: in a survey of 158 projects, 66% contained at least one file that nothing imports; the worst had 48% of files unreferenced.

This is where it starts to tip. Nobody, including you, can now say which version is live, and every change risks editing the wrong copy.

SIGNAL 04still repairable

The database schema exists only as clicks in a panel

Measured: 76 of 100 Lovable projects have no migrations at all.

Not a rebuild on its own, but it means the rules protecting your data are invisible from the code and have to be reconstructed by hand before anything can be trusted.

SIGNAL 05points at rebuilding

Every fix breaks something unrelated, repeatedly

Not measured. This is judgement, and it is the strongest signal on this list.

When changes stop being local, the structure has already failed. Repairing symptom by symptom becomes more expensive than rebuilding the part that keeps breaking.

SIGNAL 06points at rebuilding

The app never actually worked, only looked finished

Not measured.

There is nothing to rescue. Repair assumes a working thing that stopped working; this is a prototype that was never wired up, and finishing it is a build, not a fix.

Why we publish this

It costs us work, and it is worth more than the work it costs.

A repair that cannot hold is worse for you than a rebuild you did not want to hear about, and worse for us than the money is worth.
If the first look says rebuild, that is what we will say, and we will say what we think it would take. Whether we are the ones who do it is a separate question you can answer later.

Not sure which side you are on? That is what the first look is for, and it is free. Send the address of the page that breaks.

Send us the app