We keep getting hired to fix vibe-coded apps
AI can produce a working-looking app in an afternoon. Months later it breaks in production, nobody in the company can explain why, and lately those companies have been calling us.
AI can produce a working-looking app in an afternoon. Months later it breaks in production, nobody in the company can explain why, and lately those companies have been calling us.
We use AI every day, so this isn't an argument against it. It's about one failure we keep seeing: an app that was vibe coded into existence and is now falling apart.
Vibe coding means describing what you want to an AI, accepting what it produces and shipping it, without anyone on the team understanding the code. For a demo or a weekend prototype, that's a reasonable shortcut.
Trouble starts when the prototype becomes something the business depends on. The AI was aiming for an app that looks like it works today. Nobody asked it to handle real data, or a change request a year later.
An AI can write the code quickly. Someone still has to understand it well enough to keep it running once real customers and real data are involved.
We've taken over several of these projects, and the problems repeat.
Nobody on the team understands the app. A model wrote it one prompt at a time, so when something breaks there's no design to reason from.
Security holes are easy to find. We see passwords and API keys sitting in the code, and customer data open to anyone who knows the right URL. The AI was asked to make a feature appear and was never asked to keep attackers out.
Performance falls apart with real volume. A screen that was fine with 10 records becomes unusable at 100,000, because the demo data was tiny.
The same rule is written five slightly different ways in five places, so a "simple" change breaks things nobody knew were connected. There are no tests and no notes on why anything was done.
Any one of these can be fixed on its own. Combined, in a system nobody understands, they make every change request risky.
The AI did what it was asked. Nobody asked it for something secure and maintainable, and that part comes from engineering judgment rather than a prompt.
An experienced engineer keeps the whole system in mind, weighs trade-offs and is accountable when something breaks. Skipping that step doesn't save time. It creates a debt that tends to come due when the app is busiest.
First we read every line and map what the system actually does, so a person understands it again. Screens and flows the AI got right, we keep. Then we fix the foundation (data structure, validation, security, speed) and add documentation and tests, so the next change is safe to make.
We use AI heavily in this work too, with an engineer reviewing every line and answering for the result. That combination is fast. AI with nobody reviewing its output is how these apps got into trouble in the first place.
If you built something with AI and it works, keep going. If changes keep breaking things, or you have bugs nobody can reproduce, have it looked at before it fails in front of a customer. Book a call and we'll tell you what's worth keeping and what needs to be rebuilt.