What Is Vibe Coding? Definition, Process, and Limits

Vibe coding means programming by description: you formulate in natural language what a piece of software should do, an AI model generates the code, and you judge the result by the running program instead of by the individual lines. Andrej Karpathy coined the term on February 2, 2025. Collins Dictionary named it Word of the Year in November 2025.

What is vibe coding?

Vibe coding is a way of working, not a tool. The difference from the familiar practice of programming with AI assistance lies in how closely you look: with the classic approach, you read every suggestion before accepting it. With vibe coding, you judge the result, does it run, does it look right, and send the next prompt whenever it deviates.

Andrej Karpathy, co-founder of OpenAI and former AI lead at Tesla, described this in his post of February 2, 2025 as an approach where you "fully give in to the vibes" and "forget that the code even exists." The line was half tongue-in-cheek. It still describes precisely what defines the mode: the source code is no longer the artifact you work with, but an intermediate product.

In November 2025, Collins Dictionary named "vibe coding" its Word of the Year. That is not technical proof of anything, but it is an indicator of how quickly the term moved from a niche post into everyday language.

How does vibe coding actually work?

The process is a loop of four steps that repeats until the result fits.

• State the intent. You describe the desired behavior, not the implementation: "A form that takes requests and sends me an email."
• Let it generate. The model produces files, dependencies, and configuration in one go.
• Try it out. You start the program and check it by its behavior, not by the source code.
• Refine it. Whatever is not right, you describe again in prose. Error messages are often passed back unchanged.

This loop is exactly what explains the appeal: minutes lie between an idea and a working result, not days. And it explains, at the same time, the core problem that the rest of this article series revolves around: nobody has read the resulting code.

Is vibe coding the same as programming with AI?

No. Every instance of vibe coding is programming with AI, but not every instance of programming with AI is vibe coding. The distinguishing feature is depth of review.

Someone who has an AI suggest code, then reads it, understands it, and consciously accepts it, is practicing AI-assisted development. Someone who only judges the result by its behavior is practicing vibe coding. The line does not run along the tool, the same tool allows both, but along the question of whether someone can take responsibility for the code. Where this line runs in everyday team life is covered in the article Vibe Coding and Agentic Coding.

Which projects are a good fit for vibe coding?

Vibe coding fits wherever a mistake is cheap. That is a more precise criterion than "small projects," because even a small project can produce expensive mistakes as soon as it touches real data.

• Well suited: prototypes to clarify an idea, internal analysis scripts, throwaway tools, test data generators, a clickable draft for aligning with other departments.
• With reservations: internal applications with real company data, here you need at least a review by someone who can read the code.
• Not suited without a review layer: anything involving user data, payments, authentication, or legal consequences. Why this is not just a note of caution but measurable is shown in the article on risks and security vulnerabilities.

How do I know if my vibe coding result is production ready?

By four questions, each answerable with yes or no. If any one of them is no, the result is a prototype, no matter how good it looks.

• Can someone on the team explain the code without asking the AI?
• Are there tests that fail when the behavior breaks?
• Has it been checked what happens if inputs are missing, wrong, or malicious?
• Is it known which third-party packages were added and where they came from?

The full transition from draft to production is described in the article From Prototype to Production.

Does vibe coding really make you faster?

For experienced developers in familiar codebases, that is not proven, a controlled study even found the opposite. The research organization METR had sixteen experienced open source developers complete 246 real tasks in their own, mature repositories between February and June 2025, randomly assigned to work with or without AI tools (arXiv:2507.09089, published July 2025).

The result: with AI, participants took 19 percent longer. More notable than the figure is the misjudgment: the same developers estimated afterward that AI had sped them up by 20 percent. Beforehand, they had expected a 24 percent time saving.

In fairness, this needs context: METR itself now treats this result as historical and points out that it does not necessarily reflect today's tools, the tools used at the time were mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet. The study therefore does not prove that AI makes you slower. It proves something else that is more robust: your own sense of time is not a reliable measurement when using AI.

Frequently asked questions about vibe coding

Who invented the term vibe coding?

Andrej Karpathy, co-founder of OpenAI, in a post on X on February 2, 2025. Collins Dictionary named the term its Word of the Year 2025 in November 2025.

Do you need programming knowledge for vibe coding?

Not to produce a working result. To judge whether that result is secure, maintainable, and correct, yes. This gap between "it runs" and "it holds up" is the actual subject of the debate.

Which tools are used for vibe coding?

Common ones include agentic command line tools such as Claude Code, AI editors such as Cursor, assistants such as GitHub Copilot, and platforms such as Lovable or Bolt that directly deliver a working application. A comparison is available in our article on vibe coding tools.

Is vibe coding just hype?

The usage is real: in the Stack Overflow Developer Survey 2025, 84 percent of the developers surveyed said they use AI tools or plan to, up from 76 percent the year before. What is disputed is not the adoption, but the depth of review.

What is the difference to low code?

Low code platforms work with prebuilt building blocks and graphical interfaces; the result stays inside the platform's system. Vibe coding produces real source code in a real programming language, with all the freedoms and all the obligations that come with it.

The four questions above are easy to ask and hard to keep asking consistently. That is exactly what Code Guardian is for, a skill package that wires these checks into Claude Code as gates, instead of leaving them to individual discipline.

Sources

• Andrej Karpathy, post on X from February 2, 2025 (origin of the term).
• Collins Dictionary, Word of the Year 2025, announced on November 6, 2025.
• Becker, Rush, Barnes, Rein: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, arXiv:2507.09089, July 2025; supplemented by METR's own framing from February 2026.
• Stack Overflow Developer Survey 2025, "AI" section.

As of: July 27, 2026. This article is a general professional overview and does not replace an assessment of your individual case.