What is Vibe Coding?
AI is changing how we write software. Tools like Cursor, Claude Code and GitHub Copilot make it possible to build in minutes what used to take days. But there's a new term that keeps popping up: vibe coding. What does it actually mean? And is it enough to build real software?
What is vibe coding?
Vibe coding is programming on feel with AI. You describe what you want in plain language, an AI tool generates the code, and you accept the result; often without fully understanding it. It's about speed and intuition, not deep understanding of what's happening under the hood.
In practice it looks like this: you type a description, the AI generates code, you run it, and if it works you move on. If it doesn't work, you paste the error message back and let the AI try again.
Who coined vibe coding?
The term was coined by Andrej Karpathy, former AI director at Tesla and co-founder of OpenAI. In February 2025 he described a new way of coding where you "fully give in to the vibes"1: accepting code you don't fully understand, and trusting that the AI gets it right.
Karpathy is no amateur. The fact that someone with his background describes this approach shows how powerful (and tempting) it is.
How does vibe coding work in practice?
Say you want to build a login page. With vibe coding the workflow looks something like this:
- Describe your feature: "Build a login page with email and password"
- AI generates code: you get HTML, CSS and JavaScript back
- Test it: you open the page in your browser
- Error? You copy the error message
- Paste it back: the AI fixes the error
- Repeat until it works
For a simple login page this works fine. Within fifteen minutes you have something that looks good and functions. The problem starts when you want to build more.
Why vibe coding becomes a problem
The big problem with vibe coding is what you might call the "scar tissue" problem. AI builds perfectly fine code 9 out of 10 times. But that one time it's not quite right, you don't notice on the surface. It looks good, it works, so you move on.
Under the hood it's a different story. The AI may have built a feature with an ugly workaround, or applied a fix that technically works but is structurally flawed. You don't see that as a user; only in the code. Such a bad solution is a scar on your project.
The problem: those scars accumulate. After a few weeks of building you have code that:
- Is hard to maintain; nobody understands why certain things were built that way
- Gets harder to extend; adding new features takes longer because they conflict with earlier workarounds
- Is fragile; small changes unexpectedly break other parts
- Isn't production-ready; it works on your laptop but fails as soon as you run it somewhere else
This is the point where many vibe coders get stuck. The project that started so fast becomes a codebase full of scars where every new change is a risk.
The rise of the "vibe code fixer"
This problem is so widespread that an entire new market has emerged: developers who get paid to fix vibe-coded projects.
On freelance platforms like Fiverr you'll find hundreds of listings from programmers who specifically focus on fixing vibe code2. There's even a dedicated platform, VibeCodeFixers.com, where more than 300 experienced developers have signed up to clean up vibe-coded projects3. Established software companies have launched dedicated "vibe code cleanup" services.
The typical scenario: someone has spent thousands on AI credits, has a half-working prototype, and now needs a real developer to make it production-ready. Developer Timothy Bramlett confirmed the trend on X: "The worst job in 2025: Vibe coding cleanup specialist. I can confirm it's real."4
Ironically, even Andrej Karpathy himself, the person who coined the term, wrote his latest project entirely by hand. He had tried AI agents, but they were "just not well enough at all"5.
The fact that an entire industry has emerged around cleaning up vibe code says enough about the limitations of this approach.
Is there a better way?
Yes. You don't have to give up the power of AI tools; you just need to use them differently. Instead of letting AI decide everything, you can collaborate with AI as a team.
With vibe coding you let AI make the decisions. With agentic engineering you make the decisions and use AI to build faster. The difference:
- Vibe coding: prompt → code → error → prompt → hope it works
- Agentic engineering: plan → prompt → validate → iterate → understand what you're building
With agentic engineering you think about architecture first, work in controlled steps, and understand what's happening in your codebase. AI becomes your partner, not your replacement.
Want to know exactly what the difference is? Read Vibe Coding vs Agentic Engineering: What's the difference?
Conclusion
Vibe coding is powerful for prototypes and quick experiments. It massively lowers the barrier to building something. But for production software, code that needs to be scalable, maintainable and reliable, you need more.
The good news: you don't have to choose between speed and quality. With the right approach you can combine the speed of AI tools with the structure that production code requires.
Want to learn more?
Read our deep dive on the difference between both approaches: Vibe Coding vs Agentic Engineering.
Or do you want to learn it hands-on?