Here’s What Happened When I Gave ‘Vibe Coding’ a Try

Generative AI is changing the way we live and work in a multitude of ways, and that includes coding. The best AI bots of today can debug, refine, and create code, from a simple text prompt. You can put together a small app or website in the same way you can generate an image or write an essay with these tools.

Many professional programmers are now leaning heavily on AI to help get their work done, but the door has also been opened for those without any coding experience to get involved. This is something that’s been termed “vibe coding”—where you’re coding on vibes, basically—and all you need to get started is an idea.

But is it really as simple as simply typing out what you want and letting an AI bot do the coding on your behalf? According to a lot of AI enthusiasts I see on X, vibe coding means “it’s over” (whatever “it” is), and people are “one-shotting” all kinds of projects (as in, using a single prompt to create something).

To investigate further, I tried to create a little project of my own using ChatGPT and Gemini. Here’s how it went.

What is vibe coding?

There’s no official, fixed definition of vibe coding, but in general it refers to using natural language prompts to get an AI to code apps and websites. In the same way as you might ask ChatGPT for an image of a cabin in a wood, you can also ask it for a landing page for your company, or a tool to log and analyze income and expenditure.

Usually, despite all the talk of one-shotting, the idea is not to try and get everything included in your very first prompt, but to start small and then use follow-up prompts to refine. There are several variables involved when it comes to building websites and apps, and so you’re going to have to specify layouts, images, interactions, colors, fonts, and plenty more.

You can always get the AI to make some of these decisions for you, but to really get something close to what you originally imagined you’re going to want to be specific. AI bots are also able to debug code and fix problems for you—again, you just describe what’s not working. It’s a bit like having a conversation with an actual programmer.

A common way that people get started with vibe coding is to build simple games, and you’ll find plenty of examples of this on the web. You describe what you want to happen, and the AI writes something to match—then you chat through how you’d like the gameplay to flow, and how the visuals should appear on screen.

All of the major AI chatbot tools now have coding components included in them, and will present you with both the raw code (which you can edit manually, if you want), and a preview of how the executed code works—your app will run or your website will be displayed right inside the chatbot interface. The AI can even select a coding language for you, if needed.

Vibe coding has its limitations, not least because of the unpredictability of AI, and for larger, professional software programs and games these AI bots will just be one tool among many that coders use—there is the potential for disaster if AI is not used carefully. However, for fun, small projects, this is something now available to anyone with access to an AI with coding capabilities.

Trying out vibe coding

I decided to build a simple HTML-based elevator sim I could run in my browser—I’ve always been slightly confused about how elevators work, and it seemed like a suitable project for testing vibe coding.

ChatGPT was the first AI bot I tried to get to help me, and I quickly realized that quite a lot of prompting is needed—more than you might think at first. I couldn’t just say “build me an elevator sim”: I had to specify the screen layout, the number of floors, the speed of the elevator, how the passengers should be displayed, and how to determine the floors they’re heading to, just for starters.

After a few minutes of the AI thinking, I had my sim. And it worked, up to a point. However, it didn’t all work, all at the same time. Some of the issues the app had were freezing at certain floors, not picking people up in the right order, forgetting how many passengers were actually in the elevator, and generally not following elevator logic—each time I would point out the mistake, ChatGPT would apologize and try again.

ChatGPT’s elevator sim—which never properly worked.
Credit: Lifehacker

Most bugs could be fixed within a prompt or two, but then new issues would appear. One problem that was particularly difficult to solve was getting the elevator code to go back and pick people up when the cab was full the first time. Because I didn’t understand the code, I couldn’t really see where the problem was. With each revision the AI was apologetic, but didn’t seem to grasp what was going on.

At one point the elevator would zoom around the floors picking everyone up who called for a ride, without dropping off the existing passengers first. Then a weird graphical glitch developed on the passenger sprites. By this stage, even vibe coding was starting to feel like hard work: After about 45 minutes, and still without a fully working elevator sim, I decided my time and effort would be better spent elsewhere.

I gave Gemini a go with this task too, and to Google’s credit, it did do better. There were fewer issues, but there were still issues: Passengers still got picked up in the wrong order, and it didn’t follow my instructions entirely. I spent less time on this with Gemini, but it got closer to what I wanted, even if it still had glitches I wasn’t happy with.

Gemini was better at the task, but it still wasn’t perfect.
Credit: Lifehacker

Overall, vibe coding was a frustrating experience. Perhaps the issue is with AI understanding how an elevator works, rather than with its actual coding capabilities, but I was disappointed not to be able to get something that worked properly. Maybe when I’m over the annoyances that I had this time around I’ll go back and try something different—without the logical complexities of an elevator system.

My experience did highlight some of the limitations of vibe coding: You’re often going to need a lot of prompting to get the AI to understand what you want, and there will be bugs to fix along the way, even if the chatbots are very accommodating and polite when it comes to fixing those bugs.

There are also two recognizable hallmarks of generative AI here, too: An air of confidence and authority in responses, even when those responses are wrong, and unpredictability in the results. These AI models are designed to give varied answers to the same prompts, which is fine when you’re churning out 10 AI pictures of a waterfall but not so helpful when you’re trying to wrangle some code into shape.