This Is What Happened When I Tried “vibration Coding”
Generative AI is changing the way we live and work in many ways, including programming . The best modern AI bots can debug, improve, and create code simply by entering text. Using these tools, you can create a small app or website just like you can generate an image or write an essay .
Many professional programmers are actively using AI in their work, but now there’s an opportunity for those without programming experience to join in. It’s called “vibrocoding”—writing code based on vibrations—and all you need to get started is an idea.
But is it really as simple as just typing and letting an AI bot write code for you? According to many AI enthusiasts I see at X, vibe coding means “everything” (whatever that means), and people “run” any project “in one go” (i.e., create something with a single prompt).
To further explore, I tried creating my own small project using ChatGPT and Gemini. Here’s how it went.
What is vibration coding?
There’s no official, fixed definition of vibration coding, but it generally involves using natural language prompts to enable artificial intelligence to code apps and websites. Just as you might ask ChatGPT to send you a picture of a cabin in the woods, you can ask it to create a landing page for your company or a tool for accounting and analyzing income and expenses.
Typically, despite all the talk about doing it all at once, the idea isn’t to try to include everything in the first request, but to start small and then use subsequent requests to refine it. Websites and apps involve many variables, so you’ll need to consider layouts, images, interactivity, colors, fonts, and much more.
You can always delegate some of these decisions to AI, but to get something truly close to your original vision, you’ll need specifics. AI bots can also debug code and troubleshoot problems for you—again, all you have to do is describe what’s wrong. It’s a bit like talking to a real programmer.
People usually start working with vibe code by creating simple games, and you’ll find plenty of examples online. You describe what you want to happen, the AI writes something accordingly, and then you discuss how the gameplay should flow and what the visual effects on the screen should look like.
All major AI chatbot development tools now include coding components and provide both raw code (which can be edited manually if desired) and a live preview of how the executable code will run—your app or website will be launched directly within the chatbot interface. The AI can even select the programming language if needed.
Vibe coding has its limitations, not least due to the unpredictability of AI . For large professional programs and games, these AI bots will be just one of many tools used by programmers—unwise use of AI can lead to disaster. However, for smaller entertainment projects, it is now within the reach of anyone with access to an AI with programming capabilities.
Let’s try vibration coding
I decided to create a simple HTML elevator simulator that could run in a browser. I’ve always been a bit confused by how elevators work, and this project seemed like a good fit for testing vibration coding.
ChatGPT was the first AI bot I tried to use, and I quickly realized it required quite a bit of guidance—more than might initially appear. I couldn’t simply say, “Get me an elevator simulator”: I needed to specify the screen layout, the number of floors, the elevator speed, how passengers should be displayed, and how to determine the floors they were headed to—and that was just the beginning.
After a few minutes of AI deliberation, I had a simulation. And it worked, up to a point. However, not everything worked at once. Among the app’s problems were freezing on certain floors, incorrect passenger boarding order, a lack of information about the actual number of passengers in the elevator, and, in general, a lack of consistency with the elevator’s logic—every time I pointed out an error, ChatGPT apologized and tried again.
Most of the errors could be fixed with one or two hints, but then new ones would appear. Getting the elevator code to return and pick up people the first time the car was full was especially difficult. Since I didn’t understand the code, I couldn’t figure out what the problem was. With each correction, the AI apologized, but it seemed unaware of what was going on.
At some point, the elevator started circling the floors, picking up everyone who called it without discharging existing passengers. Then a strange graphical glitch appeared in the passenger sprites. By this point, even coding the atmosphere began to feel like a chore: after about 45 minutes, still without a fully functioning elevator simulator, I decided it would be better to spend my time and effort on something else.
I tried Gemini with this task, and to Google’s credit, it did a better job. There were fewer issues, but they still existed: passengers still disembarked in the wrong order, and the app didn’t fully follow my instructions. With Gemini, I spent less time on it, but it got closer to what I wanted, even if there were some issues I wasn’t happy with.
Overall, programming the vibrations proved to be a rather tedious experience. Perhaps the problem lies with the AI’s lack of understanding of how the elevator works, rather than its actual programming capabilities, but I was disappointed that I couldn’t get it to work correctly. Perhaps, once I’ve worked through all these issues, I’ll come back and try something else—one without the logical complexities associated with the elevator system.
My experience has revealed some limitations of vibration coding: you’ll often need a lot of prompting to get the AI to understand what you want, and along the way, errors will arise that will need to be corrected, even if chatbots are very accommodating and polite when it comes to correcting these errors.
Here, too, are two recognizable hallmarks of generative AI: confidence and authority in its answers, even when they’re incorrect, and unpredictability in its results. These AI models are designed to give different answers to the same prompts, which is fine when you’re spitting out 10 images of a waterfall, but not so helpful when you’re trying to clean up some code.