Process
Vibe coding works now. But is it good for long-term products?
AI ships working prototypes in an afternoon. The risk isn't that they break, it's that they work well enough to become the foundation.
Describe what you want in plain language, let the model write the code. A year ago that produced demos. Now it produces things that run in production, and the question has stopped being whether it works.
The question is what you're accumulating while it works.
Where it's genuinely excellent
Validation. If you need something in front of users by Friday to find out whether anyone wants it, vibe coding gets you there faster than any team and cheaper than any contractor. Same for internal tools, throwaway experiments, and MVPs whose whole job is to answer a question and then be replaced.
When the artifact is meant to be disposable, none of the trade-offs below apply. Ship it.
What the demo doesn't show you
AI-generated code handles the path you described and tends to skip the ones you didn't. Error states. Malformed input. Concurrent writes. Anything an attacker would try. These don't announce themselves during a demo, which is precisely the problem.
Then there's comprehension. You end up with a working system nobody on the team can fully explain, which is fine right up until something breaks at 2am and there's no mental model to debug against.
Architecture is the slower issue. Good systems rest on decisions about structure, data models, and how the pieces will hold as things grow. A model optimises for making the current request work. Those aren't the same objective, and the gap between them is where technical debt accrues quietly, until a small change requires touching everything.
The failure mode worth naming
It's not that the prototype breaks. It's that the prototype works well enough.
Something built to answer a question quietly becomes the foundation, because rebuilding feels like going backwards when the thing already runs. Eighteen months later there's a product nobody can safely modify and a rewrite that can't be scheduled without stopping the roadmap.
Almost nobody decides to do this. It happens by not deciding.
How we use it
Heavily, and with a line. AI tooling genuinely accelerates real work and we'd be slower without it.
The discipline is knowing which side of validation you're on. Before it: move fast, throw things away, learn. After it, when people depend on the thing, someone with engineering judgement owns the architecture, the security model, and the performance characteristics, regardless of who or what typed the first draft.
So yes, vibe coding works now. For prototypes it's close to a free win. For the system your company runs on, it's a drafting tool with a human accountable for what ships. Speed early is worth a great deal. Speed early converting silently into fragility later is worth less than it cost.
