My AI-built app keeps breaking. What actually fixes it
Somewhere between the demo that impressed everyone and real users logging in, AI-built apps start breaking. It does not matter which tool built it, Lovable, Bolt, Replit, v0, Base44, or a vibe coding session with Cursor, Windsurf, or a coding agent. Login stops working after a change that had nothing to do with login. The database times out under load it should handle. Something that ran fine for weeks fails silently, and you find out from a customer
If that is your week, here is what is actually happening in the code, and what fixes it
Why it breaks
These tools optimize for producing working code fast. Nobody in that loop is responsible for the code still being changeable six months later. So the same five problems show up in almost every AI-built codebase I read: the same logic written in several places, no separation between screens and business rules, state kept in several stores that drift apart, no tests, and layered patches that each fixed a symptom and left the cause in place
None of that shows up in the demo. It shows up when real users create real data and the app has to change while they are using it
Why "just ask the AI to fix it" stops working
Prompting a fix works early. Each fix patches the symptom where it appeared. The patches accumulate, and because the AI does not remember its own earlier patches, they start interacting. This is why the same feature breaks over and over, and why every new prompt feels riskier than the last one. At that point more prompting is adding to the pile, not fixing it
What actually fixes it
A person actually reads the code, not just skim it, finds what the causes are, which is usually a much shorter list than the symptoms suggest. Then changes land in small verifiable steps, with tests added around each thing before it changes. It is boring, but it works
Most broken AI-built apps do not need a rebuild. They need the 3 or 4 structural causes found and fixed, and the dozens of symptom patches cleaned up behind them. A rebuild is sometimes right, but that approach is rarely required
What you can do today, without a developer
Stop prompting new features until the breakage stops. Every feature added to a breaking app raises the cost of the fix
Get the code out of the tool and into a repository you own, whether that is Lovable’s GitHub sync, Bolt’s download, or Replit’s Git export. You cannot get help on code nobody can read
Write down the last ten things that broke. If three are the same feature, that pattern is exactly what a code reading finds the cause of
Where I fit, so you can judge my bias
Reading these codebases is what I sell: a fixed-price audit that scores the code, names every problem with the file it lives in, and gives you a fix plan in priority order with effort estimates. There is a complete sample report of a real AI-built app on the audit page, so you can see the format before deciding anything. If an audit is not the right tool for your situation, I will say so on the first call, which is free.