Andrej Karpathy called it “vibe coding” — programming where you fully surrender to AI, embrace the vibes, forget that code even exists, and simply describe what you want. It sounds like a joke, but it’s changing the entire field.
Where the Term Came From¶
In February 2025, Andrej Karpathy — former head of AI at Tesla and co-founder of OpenAI — described on social media a new way he programs. He said “vibe coding” is a state where you simply talk to AI, describe what you want, accept the results, and if something doesn’t work, paste the error message back into the chat. You don’t read the code, you don’t understand it, and you don’t care.
The community’s reaction was polarized. Experienced developers shook their heads — how can you program without understanding the code? But beginners and product people started building functional applications at speeds they couldn’t have imagined before.
How Vibe Coding Works¶
A typical workflow looks like this: you open Cursor, Claude Code, or a similar tool, describe in natural language what you want to create, and AI generates the complete implementation. If the result doesn’t work, you copy the error message and say “fix it.” Repeat until it runs.
The key difference from traditional AI-assisted coding: you don’t write code and then have it completed. Instead, you describe intent at a high level and AI decides on implementation details — architecture, libraries, data structures.
Tools That Make It Possible¶
- Cursor: AI-native IDE with an integrated chat that sees the entire project. Cmd+K for inline edits, Composer for multi-file changes.
- Claude Code: Terminal agent that reads, writes, runs, and tests. Full autonomy over the project.
- Windsurf: VS Code fork with an AI-first approach, “Cascade” flow for progressive implementation.
- Replit Agent: Cloud IDE where you describe an application and the agent creates, deploys, and hosts it.
- bolt.new / Lovable: Web tools for generating complete applications from a single sentence.
What Vibe Coding Does Well¶
Prototyping: An MVP that would take a week is created in an afternoon. Landing pages, CRUD applications, internal tools — for these use cases, vibe coding is phenomenal. A single person can build a product in a weekend that would have previously required a team.
Exploration: Want to try a new technology? Tell the AI “build me a GraphQL API with Prisma and PostgreSQL” and in minutes you have a working foundation to learn from.
Automation: One-off scripts, data transformations, migrations — tasks that aren’t worth carefully engineering because you’ll discard them after first use.
Where Vibe Coding Fails¶
And now the important part. Vibe coding produces code that the author doesn’t understand. That’s fine for a prototype. It’s a disaster for a production system that processes payments, personal data, or controls medical devices.
Security: AI-generated code often contains subtle vulnerabilities — SQL injection, missing authorization, hardcoded tokens. If you don’t read the code, you won’t catch these bugs.
Technical debt: Vibe-coded projects tend to accumulate duplications, inconsistent patterns, and unnecessary dependencies. Each subsequent prompt adds a layer of complexity that nobody manages.
Debugging: When something breaks in code you don’t understand, you’re dependent on AI to fix it. If AI can’t diagnose the problem, you’re stuck. Without a mental model of the code, you lack the intuition for where to look.
A Spectrum, Not a Binary Choice¶
In practice, it’s not “vibe coding vs. traditional coding.” It’s a spectrum. At one end is full vibe — I don’t understand the code, I don’t care. At the other end is AI-assisted coding — I use AI as an accelerator but understand every line and manage the architecture.
Most professional developers operate in the middle. They let AI generate boilerplate and routine code, but review the output, manage the structure, and manually handle critical parts. This approach combines the speed of vibe coding with the discipline of software engineering.
Impact on the Job Market¶
Vibe coding democratizes software creation. Product managers, designers, consultants — people without a formal CS education — can now build functional applications. This doesn’t mean the end of programmers. It means programmers solve harder problems — architecture, performance, security, distributed systems — while delegating routine implementation.
A new skill is emerging: prompt engineering for code. The ability to clearly communicate intent, structure requirements, and efficiently iterate with AI. It’s not just a technical skill — it’s about communication and systems thinking.
Vibe Coding Is Here — Adapt¶
Ignoring vibe coding is as foolish as ignoring Stack Overflow in 2010. It’s a tool. Powerful for prototypes and automation, dangerous for critical systems without review.
Learn to use it where it makes sense. And teach juniors why they must understand what AI generates. The future belongs to developers who can do both.
Need help with implementation?
Our experts can help with design, implementation, and operations. From architecture to production.
Contact us