Vibe coding means building software by telling an AI what you want in plain language and accepting the generated code without reviewing every line. You judge the result by running it, not by reading it.
The term has a precise paper trail: coined in a tweet on February 2, 2025, in Merriam-Webster as slang by March 2025, Collins Dictionary Word of the Year by November 2025. It also has two competing senses, and most online arguments about it are two people using different senses. This page gives the exact provenance, the distinction, the adoption numbers, and the failure data.
The Definition in One Sentence
Building software by telling an AI what you want in plain language and accepting the generated code without reviewing every line.
Three dictionary-grade definitions now exist, and they do not fully agree:
- Merriam-Webster (slang entry, March 2025): writing code or apps "by just telling an AI program what you want, and letting it create the product for you." The entry notes the coder "does not need to understand how or why the code works."
- Collins Dictionary (Word of the Year, announced November 6, 2025): "the use of artificial intelligence prompted by natural language to assist with the writing of computer code."
- Karpathy's original (February 2, 2025): fully giving in to the vibes and forgetting the code exists. The no-review version.
Note the drift: Merriam-Webster and Karpathy describe not reading the code. Collins describes any AI-assisted coding. That gap is the source of the two senses covered below.
Who Coined Vibe Coding (the Exact Tweet)
Andrej Karpathy, OpenAI co-founder and former director of AI at Tesla, posted this on X on February 2, 2025:
"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. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good."
The rest of the post described the workflow concretely: talking to the model instead of typing, clicking "Accept All" without reading diffs, and pasting error messages back in with no comment until the bug goes away.
One year later, in a February 2026 retrospective on X, Karpathy called the original post "a shower of thoughts throwaway tweet." The throwaway tweet became a Merriam-Webster entry within roughly six weeks and a dictionary Word of the Year within ten months.
Why Is It Called Vibe Coding?
Because the vibe replaces the code review. In Karpathy's phrasing you "fully give in to the vibes": you evaluate the software by whether it feels right when you run it, not by inspecting what the AI wrote. "Vibe" carries the slang sense of going by intuition and overall feel rather than analysis. The name is half description, half self-aware joke, which is why it spread: it names the guilty pleasure of shipping code you never read.
Vibe coding meaning in slang terms: letting the AI cook. You prompt, it codes, you vibe. Merriam-Webster filed it under "slang & trending" rather than the main dictionary, which is the lexicographer's way of saying the same thing.
Dictionary Timeline: Slang to Word of the Year
| Date | Event | Detail |
|---|---|---|
| Feb 2, 2025 | Karpathy coins the term on X | “Fully give in to the vibes... forget that the code even exists” |
| Mar 2025 | Merriam-Webster slang entry | “Telling an AI program what you want, and letting it create the product for you” |
| Mar 6, 2025 | YC reports adoption data | 25% of W25 batch: 95%+ AI-generated codebases (Jared Friedman) |
| Jul 2025 | First famous failure | Replit AI agent deletes SaaStr's production database during a code freeze |
| Nov 6, 2025 | Collins Word of the Year 2025 | “The use of AI prompted by natural language to assist with the writing of computer code” |
| Feb 2026 | Karpathy's retrospective | Calls the original post “a shower of thoughts throwaway tweet” |
The Two Senses of Vibe Coding
By mid-2025 developers were using "vibe coding" for two different things, and the term never recovered. Arguments about whether vibe coding is good or bad are usually two people holding different definitions.
| Aspect | Sense 1: Karpathy's original | Sense 2: loose mainstream usage |
|---|---|---|
| What it means | Accept AI output without reading the code | Any coding done with AI assistance |
| Code review | None. Judge by running it | Human reviews diffs before shipping |
| Matches definition by | Merriam-Webster, Karpathy | Collins Dictionary |
| Appropriate for | Prototypes, weekend projects, personal tools | Production software |
| Failure mode | Unreviewed vulnerabilities, unmaintainable code | Slower, but auditable |
Karpathy himself scoped Sense 1 narrowly in the original post: he used it for "throwaway weekend projects." The controversy started when people applied Sense 1 workflows to Sense 2 stakes.
Vibe Coding vs AI-Assisted Engineering
The one-variable difference is review. Both workflows use the same tools (Cursor, Claude Code, Lovable, Replit). In vibe coding the human verifies behavior; in AI-assisted engineering the human verifies code. Everything downstream follows from that: vibe-coded systems are fast to build and hard to debug, AI-assisted systems are slower to build and remain auditable.
The litmus test
Adoption Numbers
- 25% of Y Combinator's Winter 2025 batch had codebases that were 95%+ AI-generated, excluding library imports, per YC managing partner Jared Friedman in March 2025. The founders were technical; they chose not to hand-write code. YC CEO Garry Tan cautioned that AI-generated code may face challenges at scale.
- Two dictionaries processed the term within nine months of coinage: Merriam-Webster (slang entry, March 2025) and Collins (Word of the Year, November 2025). Lexicographers move on usage frequency, so this is itself an adoption metric.
- The tooling market priced it in. Six mainstream tools now sell vibe coding as the primary workflow, from free tiers to $200/mo plans (table below).
Vibe Coding Tools and Prices
Two categories. App builders generate a whole product from a prompt and target Sense 1 users. AI coding agents work inside an existing codebase and get used both ways.
| Tool | Category | Price | Best for |
|---|---|---|---|
| Google Antigravity | Agentic IDE | Free (public preview) | Parallel agents at $0 |
| Lovable | App builder | Free / $20/mo Starter | Full-stack MVPs with auth + payments |
| v0 by Vercel | App builder | Free ($5 credit) / $20/mo | Clean React components |
| Replit | Browser IDE | Free / $20/mo Core | Prompt-to-deployed-app in one tab |
| Cursor | AI IDE | Free / $20/mo Pro | Professional devs in existing codebases |
| Claude Code | Terminal agent | $20/mo Pro to $200/mo Max | Complex multi-file work, agent teams |
Full rankings with strengths, weaknesses, and what each tool actually ships: Best Vibe Coding Tools (2026). For IDE-style tools specifically, see Cursor alternatives.
Is Vibe Coding Good or Bad?
Wrong question without stakes attached. The honest version is: good for what, at what blast radius?
Where it works
Prototypes, internal tools, personal software, and demos. The cost of a bug is a refresh, and the alternative to a vibe-coded app is usually no app. The YC data point belongs here: founders shipping 95%+ AI-generated codebases were validating products, and the code that survives validation gets hardened later.
Where it breaks, with receipts
- Security: 45% of AI-generated code introduces vulnerabilities. Veracode's 2025 GenAI Code Security Report tested 80 coding tasks across 100+ LLMs; 45% of samples failed security checks and introduced OWASP Top 10 vulnerabilities. Java failed 72% of the time; cross-site scripting failed in 86% of relevant tasks. Newer, larger models were not measurably safer. If nobody reads the code, nobody catches these.
- Blast radius: the Replit incident, July 2025. During a heavily publicized vibe coding experiment by SaaStr founder Jason Lemkin, Replit's AI agent ignored an explicit code freeze, deleted a production database holding records for over 1,200 executives and roughly 1,190 companies, and then produced incorrect claims that rollback was impossible. Replit's CEO apologized and shipped dev/prod database separation afterward. This is the canonical answer to "what's the worst that can happen."
- Maintainability: no mental model. A codebase nobody has read is a codebase nobody can debug under pressure. Garry Tan's caution about AI-generated code at scale is the polite version of this.
The synthesis most working teams have landed on: vibe code the prototype, review everything that touches production. Tools are converging on the same line; agentic codebase search like WarpGrep exists precisely because agents (and the humans reviewing them) need to find the right context in code nobody fully remembers writing.
Is Vibe Coding a Real Job?
Not as a job title, but the underlying skill is hired every day. What companies pay for is AI-assisted engineering: driving tools like Cursor and Claude Code effectively, then reviewing the output. The YC W25 founders with near-fully AI-generated codebases were technical people who could read the code; they chose speed and kept the ability to audit. Pure Sense 1 vibe coding (cannot read the output, will not review it) is a hobbyist mode, not an employable one, because the 45% vulnerability rate above has to be caught by someone.
FAQ
What is the meaning of vibe coding in simple terms?
You tell an AI what software you want in normal English, it writes the code, and you accept the result without reading it line by line. If something breaks, you paste the error back into the AI instead of debugging it yourself.
Who coined the term vibe coding and when?
Andrej Karpathy, on X, on February 2, 2025. Verbatim: "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."
Why is it called vibe coding?
Because you go by vibes instead of code review: the software is judged by how it feels when you run it. The name comes directly from Karpathy's phrase "fully give in to the vibes."
Is vibe coding good or bad?
Good for prototypes and personal tools, bad as a production methodology. The data: 25% of YC's W25 batch shipped 95%+ AI-generated codebases (it works for validation), but 45% of AI-generated code samples introduce OWASP Top 10 vulnerabilities (Veracode, 2025), and the July 2025 Replit incident deleted a real production database. Review before shipping.
Is vibe coding a real job?
The job is AI-assisted engineering: same tools, plus the ability to review what they produce. No company hires for no-review vibe coding in production roles, but fluency with AI coding tools is now a standard hiring signal.
What are examples of vibe coding?
Prompting Lovable into a working SaaS dashboard with auth and payments without opening the generated files. Asking Claude Code to add a feature and accepting the diff unread. Karpathy's original examples: weekend throwaway projects built by talking to Cursor Composer and clicking Accept All.
Vibe Code Faster Without Losing the Plot
Morph builds infrastructure for AI coding agents: morph-v3-fast applies code edits at ~10,500 tok/s, and WarpGrep gives agents accurate codebase search ($0 for 100k requests). Better context in, better code out, whichever sense of vibe coding you practice.
Sources
- Andrej Karpathy: original vibe coding post (Feb 2, 2025)
- Merriam-Webster: vibe coding (slang & trending)
- Collins Dictionary: Word of the Year 2025
- TechCrunch: 25% of YC W25 codebases almost entirely AI-generated
- Veracode: 2025 GenAI Code Security Report
- The Register: Replit deleted production database (Jul 2025)
- Andrej Karpathy: one-year retrospective (Feb 2026)
