What Is Vibe Coding? What It Means for Building an App
Contents
Last updated: August 2026.
Andrej Karpathy coined "vibe coding" in a post on X on February 2, 2025, describing a workflow where you describe what you want in plain language, accept the AI's output without reading every diff, and, in his words, largely forget the code exists underneath. The post reached over 4.5 million views, Collins Dictionary named it the 2025 Word of the Year, and search interest for the term reportedly jumped more than 6,700 percent that spring. What most people mean by "vibe coding" today has drifted a fair distance from what Karpathy originally described.
Where the term came from
Karpathy's original framing was specific: working in Cursor with Claude Sonnet, using voice input to minimize typing, accepting AI-generated changes without inspecting them, and pasting error messages back without comment. He was explicit that this was for throwaway weekend projects and demos, not production software, at a point where model capability was good enough to make that fun but not yet reliable enough to trust with anything serious. The term spread because it named something a lot of developers were already doing, even outside the narrow, prototype-only context Karpathy meant it for.
The definition already split in two directions
One reading stays close to Karpathy's original sense: never reading the generated code, accepting everything, reserved for low-stakes experimentation. The other, more common usage today, is looser, any workflow where you describe intent in natural language and an AI generates the code, including professional, production-facing work with real review and oversight attached. Karpathy himself has since moved past his own term: in February 2026, a year after the original post, he described vibe coding as already outdated, introducing "agentic engineering" for the more scrutinized, professional version of AI-assisted development that emerged as models got noticeably more capable over that year.
The ambiguity that matters if you're building an app
"I vibe coded an app" doesn't tell you what got built. Lovable, Bolt.new, and v0 all let you vibe code your way to a working web app, real code, a real working product, but not something you can submit to the App Store. Huxly, Rork, and FlutterFlow let you vibe code your way to a real native mobile app instead, Expo, Flutter, or Swift, compiled and signed the same way an app built by hand would be. Both are legitimately "vibe coding" by the term's current, looser usage. Only one of them ends with something installable from the App Store or Google Play. For the fuller breakdown of which tools land in which category, see our best AI app builders for mobile apps in 2026 roundup.
Karpathy's own caution still applies past the weekend-project stage
Skipping code review works fine for a throwaway prototype, which is exactly the context Karpathy built the term around. It's a real risk once an app handles actual user data, payment information, or production traffic, the exact point where security and correctness stop being optional. That's not an argument against vibe coding an app, most of this blog exists because building that way works now, it's an argument for knowing when you've moved past the context the term was originally scoped for. For a starting point on getting from a first prompt to something you can actually ship, see how to build a mobile app without knowing how to code.
FAQ
Who actually coined the term "vibe coding"?
Andrej Karpathy, former Tesla AI director and an OpenAI founding engineer, in a post on X on February 2, 2025.
Did Karpathy mean for vibe coding to be used for real, production apps?
No, not originally. His framing was explicitly for throwaway prototypes and weekend projects, at a point where he judged AI-generated code good enough to be fun but not reliable enough to trust with production software.
Is vibe coding the same thing as using any AI coding tool?
Not in Karpathy's original sense, which specifically meant not reading the generated code at all. The term's popular usage today is broader, covering any natural-language-to-code workflow, including professional work with real review attached.
Does vibe coding always produce a native mobile app?
No, and this is the ambiguity that matters most for building an app specifically. Some vibe coding tools output web apps that can't reach the App Store; others output real native code that can. The term itself doesn't tell you which.
Has Karpathy said anything more recently about the term?
Yes. In February 2026, he described vibe coding as already outdated for professional use, introducing "agentic engineering" as the term for AI-assisted development with the scrutiny and oversight that production work requires.
Is it risky to vibe code an app that handles real user data or payments?
It can be, if code review and testing get skipped the way Karpathy's original framing describes. The risk isn't the AI-generated code itself, it's treating production software with the same "accept everything" approach meant for a throwaway prototype.
Why did Collins Dictionary name it Word of the Year?
Because it captured a real, widespread shift in how people were building software in 2025, popular enough that search interest for the term reportedly jumped more than 6,700 percent in a single season.
