What Is Vibe Coding — and How Is It Different From Programming?

Vibe coding is a style of building software where a person describes what they want in plain language, an AI model generates the code, and the person largely accepts the result without reading it line by line — judging it by whether it runs, not by how it’s written. It’s a real shift from AI-assisted coding, where a developer still reviews and understands every change; in vibe coding, the code itself becomes almost incidental to the outcome.

Where the term comes from

Andrej Karpathy, a computer scientist and OpenAI co-founder, coined “vibe coding” in a February 2025 post, writing that he was building a throwaway weekend project by fully giving “in to the vibes, embrac[ing] exponentials, and forget[ting] that the code even exists.” He’d say what he wanted out loud, an AI coding assistant would write it, and he’d accept changes without reading the diffs — reserving actual debugging for when something broke outright. The term caught on fast enough that Collins Dictionary named vibe coding its Word of the Year for 2025, defining it as “the use of artificial intelligence prompted by natural language to assist with the writing of computer code.”

How it actually works

In practice, vibe coding runs as a loop: describe the feature or fix in a chat prompt, let the model write or edit the code, run it, and describe what’s wrong (or what to add next) in the next prompt. The tools that made this practical are AI-native code editors and agents built on top of large language models — Cursor, Replit, Lovable, GitHub Copilot, and Anthropic’s Claude Code are among the most widely used. Many now work as autonomous agents: given a goal, they can write code across multiple files, run it, read the error output, and fix it themselves, with the human mostly steering and approving rather than typing.

This differs from earlier AI coding help — like autocomplete-style suggestions — mainly in scope and in how much a person still reads. Programmer Simon Willison drew the line simply: if you’ve reviewed, tested, and actually understood the code, that’s not vibe coding anymore — it’s just AI-assisted programming, the broader category AI coding assistants fall under. Vibe coding is the version where that review step is deliberately skipped.

What it’s good for — and where it breaks down

Karpathy’s own framing was modest: fine for a disposable weekend prototype, not a description of how to build production software. That distinction has mostly held up. Vibe coding is genuinely useful for quick prototypes, internal tools, one-off scripts, and letting non-programmers build a working app from a description — the kind of software where a bug costs little and getting something running fast matters more than code quality.

The risks show up once the stakes rise. A May 2025 study of apps built on the vibe-coding platform Lovable found that roughly 1 in 10 of the 1,645 apps tested had security flaws that exposed user data — the kind of mistake a code review would normally catch. A December 2025 analysis by the code-review platform CodeRabbit found that AI-generated code carried nearly twice as many serious issues, and close to three times as many security vulnerabilities, as human-written code in the same repositories. Separately, a July 2025 study by the research group METR found that experienced developers using AI coding tools on familiar codebases actually finished tasks 19% slower than without them — even though they’d predicted a 24% speedup going in, a reminder that the productivity gains are far from automatic.

None of this makes vibe coding useless — it makes it a tool with a clear domain. The line to watch is the one Willison drew: skipping the review is fine for a weekend project you’ll throw away, and risky for anything that will hold real data or run in production.

How to try it

Getting started doesn’t require much setup. Cursor, one of the dominant vibe-coding editors, offers a free “Hobby” tier with limited use, and paid plans start at $20/month for its Individual tier (as of July 2026, per Cursor’s pricing page). The workflow is the same across most tools: open the editor, describe the app or feature in a prompt, and iterate from there — treating the AI’s first draft as a starting point to refine rather than a finished product to trust outright.