Vibe coding: how AI writes code for you — and why it matters

Andrej Karpathy — one of the founders of OpenAI — posted a tweet in February 2025 that shook the developer community. He admitted he no longer writes most of his code himself. He leaves it to AI and just checks the result. He called it vibe coding — programming by feel, without knowing the syntax.

Within a few months, vibe coding stopped being a developer curiosity. It became a phenomenon. Startup founders with not a single line of code on their CV are launching working apps. Marketers build internal tools themselves. Accountants build their own dashboards.

What exactly vibe coding is

Classic programming looks like this: you learn a language's syntax (Python, JavaScript, SQL…), understand data structures, read documentation, search Stack Overflow, debug errors.

Vibe coding looks different: you describe to the AI what you want. The AI writes the code. You check whether it works. If it doesn't, you tell the AI what's broken. You repeat.

The key tools today are so-called AI IDEs — development environments where the AI has access to the entire project at once. The most popular are Cursor, Windsurf and GitHub Copilot Workspace. Unlike a simple chatbot, they see the whole project structure and can edit dozens of files at once.

Why it exploded right now

Technically it was possible earlier. But it worked badly — models lost context, mixed up variables, generated code that couldn't be plugged into the rest of the project.

In 2025–2026 three things changed:

The context window. Today's models work with hundreds of thousands of tokens — they can hold an entire project in memory, not just one file. That's the difference between an assistant who remembers only the last sentence and one who knows your whole company.

Reasoning models. The latest generation — Claude Opus 4.6, GPT-5.5 and Gemini 3.1 — don't just produce fast answers, they have slowed-down "thinking". Before the output they run through options, spot logic errors and fix them. The result is dramatically more stable than a year ago.

Integration tools. The AI connects directly to git, the terminal, the browser, the database. It doesn't just produce code — it runs it, sees the error, fixes it itself. The loop closed.

Concrete examples from practice

Internal tools without an IT department

The owner of an e-shop with 12 employees needed a dashboard: today's orders, average cart value, stock below minimum, alerts for complaints older than 48 hours. An agency quote: €3,200 and 6 weeks.

With Cursor and Claude, he did it himself over a weekend. No prior programming experience. The dashboard runs on his computer, connected to his e-shop data.

Reporting automation

The CFO of a mid-sized company spent 3 days every month assembling reports from Excel into PowerPoint. She described to the AI what the data looks like and how the result should look. By the afternoon she had a Python script that does it automatically. Now the report takes 20 minutes.

Prototypes before investment

A startup with an idea for a SaaS tool needed a working prototype for investors. A developer would cost €6,000 and 2 months. The founder learned the basics of vibe coding, used Windsurf + Claude Opus 4.6 and had a clickable prototype in 3 weeks. The demo landed, the investment came.

What vibe coding won't replace

It's important to be honest — vibe coding isn't a cure-all and has real limits.

Production security. AI-generated code is functional, but not always secure. Without an experienced developer to review it, vulnerabilities easily creep in — SQL injection, poorly handled inputs, incorrect authentication. For publicly accessible applications, code review is a must.

Scalability. A weekend project for 10 users and an app for 10,000 concurrent users are two different things. Architecture, database indexes, caching — these are areas where AI advises, but experience decides.

Integration with legacy systems. If your company runs on a 20-year-old ERP or custom software with no API, vibe coding won't help. AI can't work with what it can't see.

Maintenance. Written code needs maintenance. Dependency updates, bug fixes, new features. Without a basic understanding of what the code does, a vibe coder can get lost fast.

How to do it practically

If vibe coding interests you and you have no technical background, here's a realistic path:

1. Start with a small problem. Not an app — a script. Something that repeatedly costs you time: renaming files, formatting data, downloading reports. One specific problem, a clear output.

2. Use the right tool. For beginners the most accessible are Claude.ai (via the web interface) or Cursor for working on a project. You don't need to install anything complicated.

3. Learn to write good prompts. That's the real skill of vibe coding — not Python syntax, but the ability to describe precisely what you want. Context, examples, expected output. The more specific, the better the code.

4. Test every step. Don't try to have the AI write a whole app at once. Iterate — one function, test, next function. Problems surface earlier and more easily.

5. Keep a backup. Before every larger change, back up your code (or teach the AI to work with git). Losing work over a bad instruction is needless frustration.

What it means for businesses

Abroad (mainly in the US) vibe coding spread like wildfire. Elsewhere we're a few months behind — but the trend is coming, and fast.

Companies that adopt vibe coding early gain concrete advantages: faster prototyping of ideas, cheaper internal tools, less dependence on external vendors for routine automation.

Companies that ignore it will, in two years, find themselves in a world where their competitors automate processes that take them weeks — over a weekend.

It doesn't mean firing the IT department. It means giving every employee the ability to solve a small problem themselves — and freeing IT for the things where experience genuinely decides.

The practical question: do I need help?

It depends on what you want to build. For a simple script or an internal tool for your own use — you'll manage yourself, with a bit of patience and the right tools.

For a system that will process customer payments, store sensitive data or connect to key company systems — I recommend having someone alongside you to review the code and close security gaps. Vibe coding is a powerful tool, but like any tool, the result depends on who's holding it.