Vibe coding is building software by telling an AI in plain language what you want, accepting the code it writes, and feeding errors back to it instead of reading and correcting every line yourself. The developer steers by outcome and feel rather than by reviewing each change, which is where the vibe in the name comes from.

Who coined the term vibe coding?

Andrej Karpathy, the former Tesla AI director and OpenAI founding team member, in a post on X on February 2, 2025. He described a way of coding where you fully give in to the vibes and forget the code even exists. He later called it a throwaway shower-thought post. Collins Dictionary named it Word of the Year in November 2025.

Is vibe coding safe for production software?

Not without review gates. Karpathy's original framing scoped it to throwaway weekend projects. The best-known failure came in July 2025, when an AI coding agent at Replit deleted a live production database during an explicit code freeze. The company responded by separating development from production and adding a planning-only mode. The lesson practitioners drew: unreviewed execution, not AI itself, is the risk.

What does vibe coding mean if you do not write code?

It signals that software-shaped execution got cheap and fast, which moves the value to deciding what to build and judging whether it is right. Custom tools, prototypes, and internal automations that once needed a development budget now get built in days. The same pattern reached marketing under the name vibe marketing.

Market Briefings

What Is Vibe Coding? The Honest Read

Vibe coding went from a throwaway tweet to word of the year in nine months. What it is, how it works, and what it changes about how work gets made.

Denver McCourt ·July 2, 2026 ·6 min read

Nine months from shower thought to the dictionary

On February 2, 2025, Andrej Karpathy posted a stray observation on X. The former Tesla AI director and OpenAI founding team member wrote: “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”

He later described it as a shower-thought post he fired off without expectation. It has a name for a reason now. In November 2025, Collins Dictionary named vibe coding its Word of the Year.

A coined term crossing from one engineer’s feed to the dictionary in nine months is not a language story. It marks how fast the way software gets made actually changed. And the change did not stay inside software.

What vibe coding actually means

The workflow Karpathy described runs on trust in the model. You tell an AI coding tool what you want in plain language. You accept the code it writes without reading every change. When something breaks, you paste the error back and let it fix itself. His own summary of the loop: see stuff, say stuff, run stuff, copy-paste stuff, and it mostly works.

The point is the posture, steering by outcome instead of by line-by-line review. The developer stops being the typist and becomes the person who decides what to build and whether the result works.

Karpathy attached a caveat that most retellings drop. He framed it as fine for throwaway weekend projects, amusing more than industrial. The term then outgrew him. By 2026, vibe coding loosely covers most AI-assisted building, prototypes through production systems. Whether or not anyone reads the diffs.

The adoption number that made it real

The stat that moved the term from meme to market signal landed a month after the tweet. For a quarter of Y Combinator’s winter 2025 batch, roughly 95% of the codebase was AI-generated. YC managing partner Jared Friedman said it, and TechCrunch reported it in March 2025.

The detail that matters sits inside Friedman’s framing: these were not founders who could not code. He said every one of them was highly technical and fully capable of building from scratch. A year earlier they would have. They chose the faster path because the faster path now produced a company.

Vibe coding, the first year

From a throwaway post to the dictionary in nine months, with the adoption proof and the cautionary tale in between.

Karpathy on X, Feb. 2025. TechCrunch, Mar. 2025. Fortune, Jul. 2025. Collins Dictionary, Nov. 2025.

The honest read

The cautionary tale arrived on schedule. In July 2025, SaaS investor Jason Lemkin ran a publicly documented vibe-coding experiment on Replit. The platform’s AI agent deleted his live production database. It did so during an explicit code-and-action freeze, then generated thousands of fake user records. Fortune and The Register both covered the incident, and Replit’s CEO called the failure unacceptable. The company then shipped the guardrails that had been missing: separation between development and production, better rollback, a planning-only mode.

Read the incident precisely and it stops being an argument against AI coding. The agent had write access to production. No human gate stood between a generated command and live data. That is not a model problem. That is a review problem, and it would sink a human team the same way.

Hold the two facts of this story next to each other. Karpathy scoped vibe coding to throwaway weekend projects. Within months, funded companies were shipping near-fully-generated codebases to real customers. The distance between those two facts is where every vibe-coding failure story lives, and every success story too. What separates them is whether anyone built the checkpoint the original framing quietly assumed away.

What it changes about how work gets made

Here is the part that matters if you never open a code editor. Vibe coding is the loudest example of a shift that is not about software: execution got cheap, and judgment did not.

Work used to spend most of its hours in the middle, the production. AI now eats that middle. What remains, and what gains value, is the pair of bookends. Deciding what to build. Judging whether what came back is right. Operators in the AI-native space keep converging on that frame. The Replit incident is what a missing second bookend looks like.

For a founder or CMO, three practical things follow.

Your vendors already work this way. The agencies, developers, and consultants you hire in 2026 build with AI in the loop. Whatever the deck says. The useful diligence question is not “do you use AI.” It is “where does a human review the work, and who owns that call.”

Software-shaped work stopped being a budget line that filters ideas. The internal tool, the prototype, the custom dashboard. Things that once died in a backlog now get built in days. The filter moved from “can we afford to build it” to “is it worth building.” That is a strategy question, not a technical one.

Proof of work beats credentials. When anyone can produce output, produced output stops impressing. What separates operators now is checkable work: real builds, real numbers, named results. Ask for the receipts, in code and everywhere else.

The same pattern already jumped the fence into marketing, where it runs under its own borrowed name. The tooling changed what production costs. It did not change who has to be right.

References

Keep reading

One seat per market

See your market the way its buyers do.

We track the spend firm by firm, market by market. Get the read before the retainer.

Check my market