williamliu.ai
EN中文
← Blog

The Rise of 'Vibe Coding' in 2025

Vibe coding lowers the cost of prototypes and experiments, but production reliability requires a more deliberate engineering approach than accepting generated code.

2 min read
  • AI coding
  • Software quality

As an engineering manager, I don’t write code day in and day out. But in 2025 I felt like a hobbyist coder again – all thanks to a trend called “vibe coding.” The idea went mainstream after Andrej Karpathy’s viral tweet, who described a new way to program by “fully giving in to the vibes”. Within weeks, vibe coding went from an inside joke to a mainstream buzzword. Suddenly everyone from weekend coders to engineers, including myself (my hobby website for elementary school kids to practice simple math)

After watching teams (and doing a simple math practicing website myself), I’ve landed on a simple take: vibe coding is incredible at accelerating momentum—but quality-critical work needs a different operating mode.

Vibe coding best at "reduce the cost of trying"

Vibe coding wins when the goal is speed-to-something-real:

  • Demos and prototypes: getting from idea → working UI → shareable link in hours.

  • Throwaway experiments: validating a concept before you invest real engineering cycles.

  • Boilerplate and glue code: wiring APIs, scaffolding, small refactors, quick scripts.

But the same behavior that makes it fast—accepting code you didn’t fully reason about—can be a liability in production.

Why it breaks down in production

In my simple math practice website, I had a great first hour of vibe coding and made a lot of progress. In the second hour, I started going back and forth on some of the new requirements I added. The biggest problem came when I began testing on different devices—laptop, mobile phone, and iPad. The layout stopped making sense on some of them, and despite prompting many times, I couldn’t move forward.

I searched online and finding that I am not alone!

  • A CodeRabbit analysis found AI-generated pull requests had more issues than human-written ones (10.83 vs 6.45 on average), including higher critical and major issues.

  • A Veracode-reported finding: AI-generated code introduced security vulnerabilities in 45% of cases.

So if you’re building systems used at large scale like WhatsApp and Instagram —where “small” bugs become incidents and SEVs—pure vibe coding is a huge risk to the stability of the system.