Vibe Coding

Programs consist of instructions that a computer executes step by step. Until now, these instructions were written by humans in an artificial language, such as Python or JavaScript. With vibe coding, one no longer writes them oneself. Instead, one describes to an AI program, in German or English, what the end result should be: “Build me a website where I can quiz myself on vocabulary.” The AI generates the text of the instructions, one runs the result, checks whether it fits, and simply describes afterward what still needs to be different if there are errors. The decisive point is: the human does not check the generated instructions line by line, but only judges whether the result feels right. That is precisely what the English word “vibe” refers to—mood or feeling.

Who can suddenly build software

Learning to program takes years. Vibe coding drastically lowers this barrier. Anyone with an idea can build a working app in an afternoon without mastering a programming language. For small tools, prototypes, and hobby projects, this is a real change. Start-ups are now showing products whose first version was created almost entirely this way.

Experienced developers also work faster. Routine work that used to take hours is now done by the AI in minutes. That is precisely why the term is interesting to investors. If a team of five can achieve what used to require twenty, that changes the cost calculations of entire industries. Companies like Cursor, Lovable, or Replit are being valued at billions of dollars, even though they are only a few years old.

But there is a downside that is being hotly debated in expert circles. Anyone who does not understand what their program does cannot find its errors. Security vulnerabilities remain undiscovered because no one has read the code. For a game, that is harmless. For an app that processes passwords or payment data, it is dangerous. Experts therefore recommend vibe coding for experiments, not for systems that many people rely on.

The cycle of describing and trying out

The technical basis is large language models. These are AI systems that have learned from huge amounts of text which words and characters typically follow one another. Because millions of programs are publicly viewable on the internet, these models have also learned what code looks like. They generate it by the same principle as an essay: character by character, always the most probable continuation.

The workflow is a cycle with four steps. One describes the wish, the AI generates code, one runs it, one describes the deviation. More modern tools shorten this cycle. They run the program themselves, read the error message, and immediately attempt a correction. Such systems are called coding agents, because they work through several steps without asking back.

A common misconception is that vibe coding is simply a synonym for AI-assisted programming. That is not correct. Anyone who reads an AI’s suggestions, checks them, and consciously adopts them is not doing vibe coding. The defining feature is precisely the letting go of control over the code. Karpathy described it roughly like this: you forget that there is even any code there at all.

From school project to quarterly earnings

In everyday life, one encounters the principle in tools like Cursor, GitHub Copilot, Claude Code, or Lovable. Some are full-fledged development environments, others are websites where you just type in a sentence and get a finished application back. For simple projects, such as a schedule planner for a school newspaper, that is often enough.

In business news, the term appears from two directions. First, in the valuations of the companies mentioned, which are among the fastest-growing software companies of all. Second, in the debate about jobs: major tech corporations are now openly citing AI tools as the reason for cautious hiring of junior developers. Whether that holds true in the long run is disputed, since the maintenance and review of code does not decrease as a result, but rather increases.