Last year, vibe coding felt like a cheat code. Describe what you want, accept the AI's output, ship the feature. This year, the bills are coming due.
A developer's firsthand account of their vibe coding disaster - published on their personal blog and widely circulated this week - illustrates what happens when AI-assisted development goes wrong in ways that are hard to diagnose and harder to reverse.
For context: vibe coding is a term coined by AI researcher Andrej Karpathy in February 2025. The idea is to treat the AI as the programmer and yourself as the product manager. Describe what you want in plain language, let the model generate the implementation, glance at the output, move on. For prototypes and throwaway scripts, this can be fast and effective. For anything that needs to survive in production, the risks compound quickly.
How the Damage Accumulates
The failure modes cluster around a few patterns.
AI coding tools produce code that looks correct - sensible structure, proper variable names, plausible comments - but contains subtle bugs that only surface under specific conditions. A developer who hasn't read the code can't recognize when the AI has gone sideways.
AI models also tend to solve problems by pulling in libraries rather than writing minimal code. Accepting suggestions without reviewing them can result in bloated dependency trees (the full list of third-party packages a project relies on), some containing security vulnerabilities or from projects that are no longer maintained.
When AI-generated code breaks, tracing the bug is exponentially harder if you never understood the implementation. You're debugging someone else's choices - someone who doesn't exist and can't explain their reasoning.
Over a longer project, different AI sessions make different assumptions about how data is structured or how components interact. These inconsistencies rarely fail loudly. They erode reliability over weeks, until the cost of fixing them exceeds the cost of starting over.
The Judgment Call AI Still Can't Make
The irony is that this style of development exposes rather than eliminates the need for engineering judgment. You still need to know enough to recognize bad output, refactor accumulating inconsistencies, and decide when to discard an AI suggestion and write it yourself.
Developers who use Cursor or Claude Code most effectively aren't the ones who accept every suggestion. They treat the AI as a fast but fallible pair programmer - reviewing output, pushing back when something feels off, maintaining a mental model of the codebase even as the AI writes more of the code.
Vibe coding as a pure "no reading required" approach works fine for demos. For anything that needs to run reliably for months, the horror stories multiply because the underlying engineering discipline doesn't vanish just because a model wrote the code.