AI Vibe-Coding for a Price
Created 2026/06 – Alfred Reibenschuh – parts rewritten or translated from german into english using CoPilot and ChatGPT.
Trying Out
I just tried out "vibe coding," and I'm having an absolute blast. The experience is genuinely impressive. That said, I suspect the outcome would have been a complete mess if I hadn't been able to provide the AI with highly specific, detailed, and technically informed instructions—something that only comes from years of experience in the field.
The AI is remarkably capable at turning intent into implementation, but it still depends heavily on the quality of the intent it receives. Knowing what to ask for, what constraints to impose, which trade-offs to prioritize, and what architectural pitfalls to avoid remains a distinctly human skill. In many cases, that knowledge is the product of having already made—and learned from—a great many mistakes over the years.
Observations
A junior developer could very well produce something that looks nearly identical from the outside. The UI might be polished, the features might appear complete, and the application could even function correctly for common use cases. The difference often lies beneath the surface. Without a solid understanding of architecture, maintainability, security, testing, operational concerns, and long-term evolution, it's easy to accumulate enormous technical debt. The result is software that works today but becomes increasingly difficult to modify, debug, scale, or operate tomorrow.
In that sense, AI can accelerate implementation, but it does not automatically replace engineering judgment. If anything, it amplifies the importance of that judgment. A senior engineer can use AI to rapidly generate code while continuously steering it toward sound architectural decisions. A less experienced engineer may generate the same visible result while unknowingly embedding design flaws that will eventually make maintenance a major undertaking—or, in the worst cases, render the system nearly unmaintainable.
Other Fields
I think a similar dynamic exists outside of software development. Content creation, writing, design, analysis, and many other disciplines can now be accelerated dramatically by AI. A novice can often produce something that appears finished at first glance, and they can do so much faster than before. However, the subtle details—the nuances, edge cases, consistency, depth, accuracy, and overall craftsmanship—often remain rough around the edges.
AI is exceptionally good at producing a plausible first draft. What it cannot reliably replace is the domain expertise required to recognize where that draft falls short. The difference between "good enough" and genuinely excellent work is often invisible to non-experts, but immediately obvious to practitioners with deep experience in the field.
Conclusion
To me, that's the real lesson of AI-assisted work: the bottleneck is shifting. The value is moving away from the mechanical act of producing an artifact and toward the ability to direct, evaluate, refine, and validate it. The tools are making execution faster; expertise remains what determines whether the result is merely convincing or actually good.