Why I don’t just trust AI-vibes: Vibe Coding vs. AI-assisted engineering

3 April 2025 | 3 minuten leestijd

Recently, a curious term has been making the rounds: vibe coding. This term was introduced last February (2025) by Andrej Karpathy in a Tweet on X. It describes the act of using AI tools to write code based purely on what feels right. You throw some prompts into the void, copy-paste the output, and call it a day. No validation, no testing, no understanding. Just vibes, because the code looks and feels right.

As someone with years of experience in data engineering, I’ve leaned heavily into using AI tools such as OpenAI’s ChatGPT and Anthropic’s Claude Sonnet. These tools have become an important instrument in my work. But let me be clear: this isn’t vibe coding.

Using AI in my day-to-day work is a bit like being a carpenter with access to the best power tools on the market.

These tools don’t design the foundation of the house I want to build. They don’t decide which joints to use, how to balance form and function, or which materials are best. That’s the carpenter’s job, my job. The tools help me amplify my skills, speed up the process, and allow for more precise execution. But, they still require a trained hand and a clear plan.

That’s how I use AI.

AI helps me write, validate, debug, migrate, and refactor code faster. It can scan for security flaws, suggest improvements, and even explain why something might not work. But I’m still the one measuring twice and cutting once. I test the code. I run it. I tweak it. I make sure it fits the larger structure of the solution or architecture.

Vibe coding, on the other hand, is like picking up a power saw without knowing how constructions work, how wood behaves or which way the grain runs. You might accidentally build something that stands up… briefly. But the risk of it collapsing later is high.

Just as you wouldn’t trust a random set of planks and nails thrown together as a house, you shouldn’t trust code blindly just because it came out of an AI.

So, AI is a powerful assistant and not a replacement for skill and expertise.

Real AI-assisted development is a skill, not a vibe

With vibe coding you treat AI like a vending machine: insert prompt, receive answer, move on. It encourages a surface-level understanding of the code being generated. This might be ideal for quick prototyping or making your first steps in programming, but it rarely scales, and there’s a big chance that it hides technical debt or subtle bugs that only emerge in real-world production environments.

You wouldn’t trust and pay for a human developer who said, “I didn’t run the code, but I feel like it’s right.” Why should it be different with AI then?

That’s why AI-assisted development is a skill. A skill that effectively requires:

• Clear prompting to describe the problem and context accurately
  
• Critical thinking when evaluating the output as you would do from a junior dev
  
• Domain expertise so you know when the output is nonsense (and it happens)
  
• Iteration to refine prompts and code until it works and actually still keep making sense

If anything, working with AI has made me a better engineer. It helps me move faster, yes, but also think more clearly about my code. It’s like pair programming with an infinitely patient, always available, assisting co-worker who doesn’t run on coffee. But the key word is assisting because I’m still in the driver’s seat.

TL;DR

If you’re using AI to write code, don’t just vibe. Validate. Test. Understand.

Good code doesn’t come from vibes, it comes from effort, iteration, and understanding. Programming using AI is a real game-changer, but only if you know how to use it in the right way.

- Jeroen