What Is Vibe Coding?
Quick answer
Vibe coding is building software by describing what you want to an AI model in plain language, reviewing what it produces, and steering it until the result works. You decide what to build and check the output; the AI writes most of the code. Andrej Karpathy coined the term on February 2, 2025.
Where the term came from
Andrej Karpathy, a co-founder of OpenAI and former head of AI at Tesla, coined “vibe coding” in a post on February 2, 2025. He described a way of programming where you talk to an AI tool, accept what it writes, and barely touch the code yourself. The term spread fast: Collins Dictionary named “vibe coding” its Word of the Year on November 6, 2025.
How vibe coding works
Vibe coding is a loop. You stay in charge of what gets built and whether it's right; the AI handles the typing.
- Describe. Tell the AI what you want, with context: the stack, the feature, and the constraints. “Make an accessible button with a loading state and three sizes” beats “make a button.”
- Generate. The AI writes the code, often across several files.
- Review and test. Run it. Click through it. Read the parts that touch data, money, or logins.
- Iterate. Say exactly what's wrong (“the padding is too tight on mobile”) instead of starting over.
- Commit. Save a working version in git so you can always go back.
Try it now: run the loop once
You need Claude Code and a paid Claude plan (Pro, Max, Team, or Enterprise) or a Claude Console account. Install steps are in Anthropic's quickstart.
- Open a terminal. On a Mac, press Cmd+Space, type Terminal, and press Return. On Windows, open PowerShell from the Start menu.
- Type these three lines, pressing Return after each:
mkdir my-first-app
cd my-first-app
claudeDescribe: paste this prompt.
Build a to-do list app in a single file called index.html, using plain HTML, CSS, and JavaScript. I should be able to add a task, check it off, and delete it. Then tell me exactly how to open it in my browser.Review and iterate: open the app, try every button, and describe one specific fix, for example:
Checked-off tasks should move to the bottom of the list and turn gray.Commit: when it works, save a version:
Commit my changes as a save point.What you should see: A working to-do app you built by describing it, reviewing it, and improving it: the full vibe coding loop, start to finish.
Vibe coding vs. traditional coding
| Traditional coding | Vibe coding | |
|---|---|---|
| Who writes the code | You | Mostly the AI |
| Your main skill | Writing correct code | Describing clearly and reviewing well |
| Where time goes | Typing and debugging | Planning, reviewing, and testing |
| Main risk | Moving slowly | Shipping code nobody fully understands |
How common is it?
- About 25% of the startups in Y Combinator's Winter 2025 batch had codebases that were roughly 95% AI-generated, according to YC managing partner Jared Friedman (TechCrunch, March 6, 2025).
- At Google, “well over 30%” of code checked in comes from accepted AI suggestions, CEO Sundar Pichai said on Alphabet's earnings call on April 24, 2025.
The two kinds of vibe coding tools
| Type | Examples | Best for | Trade-off |
|---|---|---|---|
| App builders | Lovable, v0, Bolt | A working prototype from a prompt, with no setup | Harder to change the foundations later |
| AI editors and agents | Cursor, Claude Code, ChatGPT Codex | Real products you own and keep building | You need to learn the basics of a codebase |
Many people start with a builder to test an idea, then move to an editor once the idea is worth keeping.
Where vibe coding breaks
- Security. AI-written code often runs fine and still has holes. In Veracode's 2025 benchmark of more than 100 AI models, the generated code introduced a security flaw in 45% of tasks. See the vibe coding security checklist.
- Growing codebases. Past a certain size, new changes start breaking old features. This guide calls that point the Vibe Wall.
- No memory between sessions. Each new AI session starts from zero unless you give it a context file like CLAUDE.md.
How to start vibe coding in five steps
- Install an AI editor (Cursor) or a terminal agent (Claude Code).
- Add a CLAUDE.md or AGENTS.md file with your stack and rules.
- Pick a small first project you can finish in a weekend.
- Commit to git after every feature that works.
- Before you share it, run a review pass with a second model. The ARC Method shows how.
Go deeper: Chapter 1: The Vibe Coding Mindset in the full guide.
Frequently asked questions
Do I need to know how to code to start vibe coding?
No, but you need to learn enough to review what the AI writes: what a component, an API, and a commit are, and how to test that something works. The AI predicts code; it does not know your users or constraints. Your review is what turns a working demo into software you can ship.
Is vibe coding the same as AI-assisted coding?
Not quite. In Karpathy's original description, vibe coding meant accepting AI-written code without reading it closely. AI-assisted coding usually means a developer stays in control and reviews every change. In everyday use the terms now overlap, and this guide teaches the version where you review what the AI writes.
What can you build with vibe coding?
Prototypes, internal tools, landing pages, and first versions of web and mobile apps are the strongest fit. Large, long-lived codebases are possible too, but they need written project context, small tasks, and code review, or changes start breaking unrelated features.
Is vibe coding safe?
It can be, if you review AI-written code like any other code. In Veracode's 2025 benchmark, AI models introduced a security flaw in 45% of coding tasks. Check authentication, input validation, secrets, and database access rules before launch, and never give an AI agent production credentials.