← Back to blog

Oct 29, 2026 · 8 min read

Vibe Coding Is Fun. It's Also Not Software Engineering.

Andrej Karpathy coined 'vibe coding' to describe letting AI write your code while you accept its suggestions without checking. It's great for prototypes. It's dangerous for production.

AIEngineeringVibe Coding

In February 2025, Andrej Karpathy posted on X about a "new kind of coding I call 'vibe coding.'" The idea is simple: you let an AI tool (like Cursor or Claude) write your code, and you accept its suggestions without checking them. When you get an error, you paste it back to the AI and let it fix itself. "Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away," Karpathy wrote.

It's a honest description of what a lot of people are actually doing with AI coding tools. And for certain things, it's fine. For other things, it's a problem.

When vibe coding works

Vibe coding is genuinely useful for low-stakes, throwaway work:

Prototypes and demos. If you need a quick landing page, a simple game, or a proof-of-concept to show an idea, vibe coding gets you there fast. The code doesn't need to be maintainable. It needs to exist.

Personal tools. A script to automate a tedious task, a small app for your own use, a one-off data transformation. If it works, it works. If it breaks, you fix it or throw it away.

Learning. For someone who's never programmed, vibe coding is a way to get something running without months of learning first. The barrier to entry drops from "learn a programming language" to "describe what you want." That's a real win for accessibility.

Exploring a new framework. An experienced developer who wants to see how a new tool works can vibe-code a quick example to understand the shape of the API, then write the real version properly.

In all these cases, the cost of bad code is low. If the AI generates something inefficient or insecure, the consequences are limited to you.

When vibe coding doesn't work

The problems start when vibe coding meets production code that other people depend on.

You don't know what the code does. Simon Willison's rule is that he won't commit code he couldn't explain to someone else. That's a good rule. If you accepted every suggestion without reading it, you can't explain it. You're shipping code you don't understand.

AI generates plausible nonsense. LLMs confabulate. They reference functions that don't exist, import libraries they invented, and call APIs with signatures that look right but aren't. In a prototype, this fails immediately and you fix it. In a larger codebase, the failure might not surface until a specific edge case hits production.

Security is invisible. A vibe-coded authentication flow might look correct and pass basic tests while having a subtle vulnerability. The AI doesn't know your threat model. It doesn't know what data is sensitive. It doesn't know who's supposed to access what. It writes code that looks like authentication code, and if you don't review it, you ship it.

Inefficiency compounds. A developer at TechTarget described vibe-coding a plugin that checked whether a line started with "//" by scanning the entire document every 150 milliseconds. That's 12,000 full-document scans in a half-hour meeting. The AI's solution worked, but it was wildly inefficient. In a small app, that doesn't matter. In a system with many such patterns, it adds up to real performance problems.

Debugging requires understanding. "Vibe coding is all fun and games until you have to vibe debug," as developer Ben South put it. When something breaks in production, you need to understand the code to fix it. If you don't understand the code because you never read it, you're stuck asking the same AI that wrote the broken code to fix it. Sometimes that works. Sometimes it makes it worse.

The line between vibe coding and AI-assisted development

Not all AI-assisted coding is vibe coding. The distinction matters.

Vibe coding: AI writes the code, you accept it without review, you ship it.

AI-assisted development: AI writes code, you review it, you test it, you understand it well enough to explain it, you ship it.

The second one is just software development with a better tool. The AI is a faster typist, not a replacement for engineering judgment. You still own the code. You still own the consequences.

Willison, who's been writing about AI tools longer than most, puts it this way: "The job of a software developer is not just to churn out code and features. We need to create code that demonstrably works, and can be understood by other humans, and that will support continued development in the future."

What this means for teams

If you're running a team, the practical question is where to draw the line. Here's what we tell our clients:

Let people vibe-code prototypes and internal tools. The speed gain is real and the risk is low. Set a clear boundary: if it's not going to production, the rules are looser.

Require review for production code, regardless of who wrote it. AI-generated code gets the same review as human-generated code. If the reviewer can't explain what a block does, it doesn't merge.

Don't let AI be the only one who understands a system. If a developer vibe-coded a feature and then leaves, the next developer has to reverse-engineer both the code and the intent. That's expensive.

Test what matters. AI is good at writing tests that pass. It's less good at writing tests that catch real bugs. Focus your testing effort on the parts where failures are expensive: security, data integrity, payment flows.

Be honest about the stakes. A marketing site that goes down for an hour is annoying. A billing system that leaks customer data is a disaster. Vibe coding is acceptable for the first and not for the second.

The honest take

Vibe coding is a real thing and it's not going away. The tools are getting better, the accessibility win is genuine, and for low-stakes work it's a legitimate shortcut. But it's not software engineering, and treating it like it is will get you in trouble.

The developers who succeed with AI tools are the ones who use them to move faster within a discipline they already understand. The ones who get burned are the ones who use them to skip the understanding entirely.

Karpathy knows what he's doing. He was head of AI at Tesla and a founding engineer at OpenAI. When he vibe-codes, he's doing it with full knowledge of the risks and on projects where the stakes are low. Most people doing vibe coding don't have that context. If that's you, or your team, it's worth being honest about it.

Need help applying this?

Schedule call →