Vibe Coding: Fun, fast, but not always good vibes

Alright, let’s talk about vibe coding. It’s one of those buzzwords that has been making the rounds on LinkedIn, X, and in tech circles, with people claiming they built entire startups, full-stack SaaS platforms, or even multiplayer games in mere hours, all by embracing the so-called 'vibes' of coding. Sounds a bit ridiculous, right? Well, it kind of is, but at the same time, it also makes a lot of sense. When leveraged correctly, vibe coding can be a game-changer, but (emphasis on but), it’s not a magic spell that instantly turns ideas into perfectly scalable products. 

The origin of the vibes

The term "vibe coding" was coined by Andrej Karpathy, AI researcher and former Director of AI at Tesla, in an X post on February 2nd, where he basically described coding in a flow state, powered by AI assistance, without overthinking structure or scalability. His exact words:

"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."

It’s the ultimate form of fast iteration. Don’t worry about the implementation details, just get the thing working, ship it, and move on. Sounds great, but like all things, it has its limits.

The explosion on social media

Since Karpathy dropped the term, vibe coding has spread like wildfire. People with no prior technical expertise are posting beautiful projects they spun up in hours, claiming AI tools like Windsurf, Cursor, and Claude basically built their software while they sat back and nodded approvingly.

Case in point: Pieter Levels, the serial indie hacker that challenged himself by creating 12 startups in a year, created an in-browser multiplayer flight simulator in just three hours using AI-generated code. Three hours. For a multiplayer game. And then he monetized the whole thing with in-game billboards and ended up making well-over $50,000 a month. 

That’s peak vibe coding energy and probably the main reason behind vibe codings virality. How can you resist the FOMO of potentially making enough money to pay for your kids college tuition in under 3 hours? As far as tech influencers are concerned, the only limit is your imagination. So they say.. 

My experiences with vibe coding

I’ve been doing this for a while now, long before I even knew it had a name. For the past year, I’ve been using AI tools to speedrun side projects, whether it’s prototyping, experimenting with new tech, or just seeing how far I can push these AI copilots. And yeah, it amps up what you can do, sort of like a magnifying glass for your technical skills. If you have a solid grasp of what you're building, AI makes the process faster and smoother. I’ve been working on a GraphRAG implementation and built a 3D graph visualizer in record time, using Cursor and Claude 3.7. Would it have taken me days or even weeks without them? Probably. Did I vibe code the hell out of it? Absolutely. And the results are stunning. In a short time I managed to learn about 3D graph visualization, render optimization techniques like instanced meshes and frustum culling, and graph theory.  All in an applied way, not just theory!

In one of my old vibe coding projects from last year I had the opportunity to play around with diffusion models to turn drawings into 3d meshes. Again, none of this would have been possible without some vibe coding to set up the canvases and UI to support the technologies I wanted to learn about.

Unfortunately, both of these projects have succumbed to the what I call: The Vibe Coding Event Horizon (more below)

The reality check: Vibe Coding has a limit

Vibe coding is a lot like traveling through space, exciting, full of potential, and seemingly limitless. At first, everything was smooth. You're cruising along, features are being developed effortlessly, code compiles on the first try, and the vibes are immaculate. You’re listening to your favorite tracks, AI is doing, half, if not all the work, and you feel unstoppable. But what you don’t realize is that technical debt is creeping in like dark matter… Unseen but dangerously present.

And then, you see it, the black hole of code entropy. At first, it’s manageable. A few bugs here and there, nothing a quick AI-assisted fix can’t solve. You keep going, you start adjusting things manually on the fly, but then the problems start compounding. The AI Agent is confused since you had to intervene. What are you doing, human?! Aren’t you vibing? The code is failing at an exponential rate. You’re being pulled toward the event horizon. Debugging gets harder. Strange things happen… five functions with the same name but slightly different parameters. You delete all but one, and somehow, an error is fixed. Weird, but okay.

Then, you hit it, the point of no return. The codebase collapses under its own weight. It refuses to build, refuses to run. You decide it’s time for an audit. And as you dive into the chaos, you realize the true horror: no structure, files with over 10,000 lines of tangled logic, duplicated functions littered across the project. Your once-beautiful vibe-coded masterpiece is now lost in the depths of an unmanageable black hole of technical debt.

At this moment, you have only two choices: manually untangle the mess (a task as grueling as trying to escape a singularity) or git revert and start fresh, promising yourself that next time, you’ll vibe code with a little more structure before inevitably getting sucked into another black hole.

Here’s the thing: vibe coding is not a replacement for actual engineering. It’s like a high-speed shortcut through the proof of concept stages of development, helping you get a rough version of something up and running quickly. But if you’re aiming to build a scalable, maintainable, production-ready system, you still need well-thought-out architecture, performance considerations, and a proper engineering approach.

If you blindly rely on AI-generated code without a solid understanding of what’s under the hood, you’re essentially playing with fire. The minute you need to debug, refactor, or optimize something, that ‘magical’ AI-generated code might start to look more like a Frankenstein monster. That’s why vibe coding is best used as an accelerant rather than a standalone strategy, it can drastically speed up development, but it works best when paired with proper engineering expertise and good coding practices.. If you’re building a throwaway prototype or a fun side project, it’s amazing. If you’re building something meant to scale, survive production, or integrate into a serious business, you still need structure, planning, and real expertise.

Structured Vibe Coding: the smarter approach

Vibe coding doesn’t have to be chaotic, and in fact, the best results come when there’s a structured approach behind it. If you want to get the most out of vibe coding while ensuring the results don’t fall apart, here are some ways to bring order to the chaos:

• Use AI for project planning, not just code generation. AI can assist in drafting software requirements, create architectural diagrams, and generate structured plans before you even write a single line of code.
• Think in terms of modularity and abstraction. Instead of letting AI generate huge, tangled blocks of code, structure it properly with reusable functions, clear separation of concerns, and meaningful abstractions that make future maintenance easier. I can guarantee you Claude isn’t thinking about the liskov substitution principle when it’s generating your cool Monkey Zoo Tycoon game. 
• Leverage linters, static analysis tools, and automated testing. AI-generated code is often rough around the edges, and enforcing proper coding standards ensures that the final product remains maintainable and scalable.
• Work in iterative cycles. Just because AI can generate a full feature in seconds doesn’t mean it’s optimized or even correct. Take the time to review, refine, and test your code in logical stages.

The evolution of Vibe Coding & what’s next

A year ago, vibe coding was mostly just devs using ChatGPT to write random scripts and functions. Today, it’s evolved into something way more structured, with cognitive architectures, efficient code search, AI copilots, and generalist agents like Manus pushing it even further. At Osedea, we’re organizing our third AI-focused hackathon this year, and every time we do this, we see the code output almost double. It’s mind blowing to witness firsthand how AI-assisted development is speeding up the entire process especially in the prototyping phase. 

Will vibe coding take our jobs? No, but the event horizon might keep pushing further away as better generalist coding agents emerge. AI will continue to enhance development speed and efficiency, but there will always be a limit when it comes to complex systems that require real engineering expertise. The need for structured thinking, solid architecture, and hands-on experience won’t disappear anytime soon.

At Osedea, we embrace vibe coding where it makes sense, rapid prototyping, accelerating workflows, and experimenting with new technologies. If you want to know more about what our expertise can do for you, reach out to us. Let’s vibe code responsibly and build something brilliant together.