According to CNET, the term “vibe coding” was coined by AI researcher and former Tesla director of AI Andrej Karpathy in early 2025, and it spread so fast it was named Collins Dictionary’s Word of the Year. It describes a new software development workflow where you “fully give in to the vibes,” using plain language prompts to get an AI to generate most of the code. The shift is happening rapidly, with data from Y Combinator’s Winter 2025 batch showing about 25% of startups now have codebases built almost entirely by AI. Tools like ChatGPT, Claude, and specialized platforms like Bolt and Replit are enabling this, allowing even non-coders to build apps and websites. The core promise is turning development into a conversation about intention, not implementation.
How it works and where it fails
So here’s the thing. The process sounds magical: you type “make me a skincare blog with an editor” and poof, you get code. Platforms like Bolt and Replit take it further by handling the entire project setup in their editor, so you never even see the raw text. You just chat, preview, and publish to a free URL. It’s incredibly empowering for prototypes and personal projects. But that’s where the vibe often meets a very hard wall of reality.
As Sam Dhar, a former engineering leader at Adobe and Amazon now at Galileo AI, points out, vibe coding replaces the need to write syntax but not the need for procedural knowledge. A beginner still needs to know what to do with that generated code—how to run it, where files go. More critically, Dhar describes software as a “pyramid of decisions,” from button colors to scaling for millions of users. You can’t prompt-engineer that entire pyramid in one go. Someone with experience has to evaluate, understand, and correct the AI’s output. “Only someone who has that knowledge and experience can truly effectively use AI to be able to build things that are production-ready,” he told CNET. Basically, AI can write the words, but you still need to be the editor.
The hidden costs and real limits
Look, I’ve tried this. The article’s author did too, spending hours tweaking prompts for a simple tool only to hit a wall. The free tokens run out. The AI hallucinates weird, non-functional code. You get something that looks right but breaks in subtle ways. This is the big risk: vibe coding works great for throwaway ideas, but it’s a minefield for anything that needs to be secure, stable, or maintained long-term. Errors and security holes are invisible if you don’t know how to look for them.
And then there’s the platform lock-in. These all-in-one editors are convenient, but you trade control for that ease. Want to move your vibe-coded app to a real server? Good luck untangling the AI’s generated structure. The piece also notes a funny gap: turning a web app into a home screen icon is trivial, but getting a real iOS app into the App Store still requires a Mac, Xcode, and a $99/year developer account—no amount of vibes bypasses Apple’s walled garden. Android is easier, with tools on Replit and Bolt to publish via Expo, but it’s still a process.
Why this changes the game anyway
So why is this such a big deal if it’s so flawed? Because it radically lowers the floor for creation. Programming has always been an elite skill with a steep initial cliff. Vibe coding turns that cliff into a slope. If you can describe an idea, you can now build a first draft of it. That’s profound. It’s not replacing senior engineers—in fact, it makes their review and oversight more crucial than ever. But it is changing the economics. If 25% of new startups are already leaning on AI-built codebases, that’s a massive shift in how teams operate and what they can build with small resources.
The tools are evolving fast, too. GitHub Copilot Workspace and others are baking this conversational style right into professional environments. The future isn’t a world where everyone is a software engineer. It’s a world where software engineering becomes less about typing syntax and more about directing, reviewing, and securing AI-generated systems. The vibes are powerful, but they’re not autonomous. And maybe that’s the point.
