Vibe coding is programming by description. You tell the AI what you want, it writes the code, and you test the output without reading a line. That's the working definition. Here's what it actually means for people building things in 2026.
Why This Phrase Caught On
Andrej Karpathy coined the term in February 2025. His description was specific: fully give in to the vibes, embrace exponentials, forget that the code even exists. In practice, he accepted every AI suggestion without reviewing the diff, pasted error messages directly back into the chat without explanation, and kept going until something worked.
That sounds careless. It also describes how a significant number of people are building things now.
The tools made it viable. Cursor, Claude, and GitHub Copilot improved fast enough that non-developers can produce working software through conversation. What used to require learning a language, understanding a framework, and spending hours debugging now starts with describing what you want in plain English.
The specific change: the cost of going from an idea to a working prototype dropped from days to minutes. That's what made the approach stick.
How Vibe Coding Works
The loop:
- Open a tool (Cursor, Claude, ChatGPT, or similar)
- Describe what you want to build in plain English
- The AI generates code
- Run it or paste it somewhere and see if it works
- Describe what's wrong or what needs to change
- Repeat until done
You don't need to know any syntax. The output is real, runnable code: HTML that renders in a browser, JavaScript that handles interactions, CSS that controls layout. You didn't write it and may not understand how it works line by line.
The reason this works is that evaluating results is much easier than writing code. Looking at a page and knowing whether it looks right takes seconds. Writing the code that makes those things happen requires syntax knowledge most people don't have. Vibe coding keeps the evaluation with you and offloads the writing to the AI.
What a Session Looks Like
Say you need a landing page for a consulting practice. You open Claude and type: "Build a one-page site for a freelance design studio. Dark background, clean typography, three services, contact email at the bottom."
Claude returns complete HTML with inline CSS and enough JavaScript to handle a contact form. You paste it into an AI website builder and it's live in under two minutes.
If something doesn't look right, you describe it: "The services section needs more breathing room and the font is too small on mobile." Claude updates the code. You paste the new version.
If something isn't working at all, you describe that too: "The form submission isn't doing anything." You paste the error message if there is one, or describe what you're seeing. Claude fixes it.
The feedback loop is: describe the outcome you want, generate, test, describe what's off, update. No context-switching between building and debugging syntax. No documentation to look up. You describe what you want until you have it.
How This Compares to Traditional Development
Traditional development: you write code, you understand how each piece works, and you debug by reading what's there.
Vibe coding: you describe outcomes, the AI writes the code, and you iterate by describing what's wrong rather than reading what broke.
The tradeoff is real. For landing pages, personal sites, prototypes, and simple tools, vibe coding is faster by a large margin. For production systems handling payments, user authentication, or sensitive data, building on code you haven't reviewed creates genuine exposure. Security problems don't care how fast you shipped.
Karpathy acknowledged this himself. When building a personal project after popularizing the term, he wrote the security-sensitive parts by hand. The vibe coding approach worked for most of the project, not all of it.
Who's Actually Building This Way
The early adopters were developers using AI to move faster inside their existing workflow. The faster-growing group is people who aren't developers at all: designers building their own portfolio sites, marketers launching campaign pages without waiting for engineering, founders validating ideas before investing in product development.
The common thread is speed and directness. You want to build a thing. The AI builds it. You put it on the internet and see if it gets any traction. The whole cycle takes an afternoon instead of a week.
Where Vibe Coding Is Going
The term is already evolving. Karpathy's more recent framing is "agentic engineering": the AI handles the code, and you orchestrate, setting direction and reviewing what comes back. The role is less "developer who uses AI" and more "person who knows what done looks like."
For people building without a development background, the direction is clear. The constraint is knowing what to ask for, not how to write what you asked for.
If you're vibe coding websites and need the publish step, see How to Use an AI Web Builder for the full workflow from HTML to live URL. For a broader look at how AI changed what builders expect from tools, see AI Website Builder: How It Actually Works.