"Vibe coding" — describing what you want in natural language and letting an AI tool generate most or all of the code, with minimal manual review of the output — is genuinely great at what it's good at. The mistake is applying it uniformly across a codebase without noticing where the failure mode quietly changes.
Where it works well
- Throwaway prototypes — validating an idea before you've committed to it.
- Internal tools with a tiny blast radius — a script only you run, where a bug costs you a rerun, not a customer.
- Greenfield scaffolding — the first draft of a new feature, before it touches real data or real users.
Where it breaks down
- Anything touching money, auth, or customer data — the failure modes are expensive and sometimes invisible until an audit or an incident finds them.
- Long-lived systems — code nobody fully understands is code nobody can safely change six months later, no matter how it was written.
- Concurrent, stateful, or security-sensitive logic — exactly the categories where subtle bugs are hardest to catch by reading the output and hardest to catch with a quick test.
The tell that you've crossed the line
A useful gut check: if you could not explain, line by line, why the generated code is correct — not just that it appears to work — you've crossed from prototyping into something that needs actual engineering review before it goes anywhere near production. Speed of generation was never the bottleneck that mattered; understanding what shipped always was.
Vibe coding doesn't remove the need for engineering judgment. It just moves that judgment from "how do I write this" to "is this actually correct" — and the second question is harder to skip responsibly.
What we do differently
AI-assisted generation is part of how we build — it's a legitimate accelerant for the first draft of a lot of code. But every draft still passes through the same review, test, and QA gates any other change would, scaled to the actual risk of what it touches. The acceleration is in drafting, not in skipping verification.
The takeaway
Match your review rigor to the blast radius of the code, not to how it was written. Fast generation is a gift for prototypes and a liability for anything you're not willing to fully understand before it ships.