// GLOSSARY · VIBE CODING

Vibe coding

Building software by describing what you want to an AI in plain language and accepting the code it generates — a term coined by Andrej Karpathy in 2025.

Last verified · 2026-09-24 · by Moe Ameen

What it is

Vibe coding is a way of building software in which you describe what you want in natural language and let an AI model write the code, rather than authoring it line by line yourself. In its purest form you barely read the output — you prompt, run what comes back, describe the next change or paste in the error, and repeat, steering by results and feel instead of by inspecting the source. The AI holds the syntax; you hold the intent.

The term was coined by Andrej Karpathy, an OpenAI co-founder and former Tesla AI lead, in a February 2025 post on X. His description set the tone that stuck: a style where "you fully give in to the vibes, embrace exponentials, and forget that the code even exists." Karpathy framed it as talking to the model, accepting diffs without always reading them, and fixing problems by asking again — coding as conversation rather than authorship.

A distinction matters here, because the phrase gets stretched. Karpathy's original sense is the hands-off, don't-read-the-code version, which he explicitly tied to low-stakes, throwaway projects. The broader everyday usage — "I built this with AI" — often means AI-assisted development where a developer still reviews, tests, and understands what ships. The industry increasingly separates the two: true vibe coding (trust the vibes) versus responsible AI-assisted engineering (the AI drafts, a human vets). Conflating them is where most of the argument about whether vibe coding is "good" or "bad" actually comes from.

The practice is enabled by a wave of tools built around the loop: agentic coding assistants like Cursor and GitHub Copilot that edit inside an existing codebase, and prompt-to-app platforms like Lovable and Replit that spin an idea into a running, hosted application. It sits alongside — and is often confused with — a broader shift toward describing outcomes to AI systems instead of specifying every step, the same instinct that drives natural-language creation in domains well beyond software.

The history

Karpathy's February 2025 post did not invent AI code generation — assistants had been suggesting code for years — but it named a behavior that was already emerging and gave it a memorable identity, and the label spread almost immediately. Within weeks it moved from a single tweet into product copy, pitch decks, and everyday developer conversation. Merriam-Webster flagged "vibe coding" as a slang-and-trending term in March 2025, and Collins English Dictionary named it Word of the Year for 2025, an unusually fast trip from coinage to dictionary recognition.

The adoption data arrived quickly too. In its Winter 2025 batch, Y Combinator reported that a quarter of the startups had codebases that were roughly 95% AI-generated — a signal that, at the earliest and most experimental end of software, describing-to-an-AI had already become a default rather than a novelty. Prompt-to-app companies rode the same wave: by September 2026 Lovable reported annualized revenue crossing $600 million with around a million new projects created weekly, one of the clearest commercial markers that natural-language software building had become a real market rather than a meme.

Alongside the enthusiasm came a sharper, more sober read. Through late 2025 and into 2026, studies and post-mortems questioned the quality and security of code shipped without review, and practitioners — Karpathy included — kept re-drawing the line between vibing on a weekend prototype and running an AI-drafted system in production with real users and real data. By 2026 the mature position was less "is vibe coding real" and more "where is it appropriate," with the honest answer being: great for exploration and throwaways, risky when unread AI code becomes the thing customers depend on.

How it behaves across platforms

PlatformBehavior
CursorAn AI-first code editor whose agent edits across an existing codebase from natural-language instructions, commonly paired with Claude models. It sits at the "AI-assisted engineering" end — the developer still has the files open — but its multi-file agent mode is where much day-to-day vibe coding actually happens.
GitHub CopilotStarted as inline autocomplete and grew into an agentic assistant that can take a described task and make changes across a repo. Because it lives inside a normal developer workflow, it is the mainstream on-ramp to AI-assisted coding for people who still review every diff.
LovableA prompt-to-app platform: describe the product and it generates a working, hosted web app with deployment built in. This is the purest consumer expression of vibe coding — the output is a running product, not a code file — and its scale (annualized revenue past $600M by September 2026) is a marker of how far the idea reached beyond developers.
ReplitA browser-based development environment whose agent can build and run an app from a natural-language brief, with hosting attached. It blurs the line between prototype and product, which is exactly why the "is this a throwaway or a real system" question follows it around.

Concrete examples

  • A founder with no engineering team describes a simple booking tool to a prompt-to-app platform, gets a working hosted version in an afternoon, and uses it to validate demand before hiring anyone to build the real thing — the canonical, legitimate use of vibe coding.
  • A developer opens an AI editor, types "add rate limiting to the upload endpoint and write a test for it," accepts the diff, runs the test, and moves on — AI-assisted engineering that is often called vibe coding but keeps a human in the review loop.
  • A weekend hobbyist builds a small browser game entirely by prompting, never reading the source, fixing bugs by pasting the error back to the model — Karpathy's original "give in to the vibes" sense, appropriate precisely because nothing important depends on it.
  • A team ships an AI-generated feature to production without review, then discovers a security hole a human read would have caught — the failure mode that separates responsible AI-assisted work from vibing on code that real users depend on.

Common mistakes

  • Treating vibe coding and responsible AI-assisted engineering as the same thing. Karpathy's original term means not reading the code; most professional AI-assisted work does read, test, and understand it. The words matter because they carry very different risk.
  • Vibing on production systems. The approach shines for prototypes, throwaways, and exploration; shipping unread AI code to a system with real users and data is where the security and maintainability problems documented through 2025–2026 actually bite.
  • Assuming the AI understood the requirement. The model writes what your prompt implied, not what you meant. Without a test or a human check, a plausible-looking result can be quietly wrong — the same "looks right, is wrong" trap as any unverified AI output.
  • Skipping the fundamentals because "the AI handles it." When something breaks in a way the model cannot fix by re-prompting, someone still has to read the code — and if no one on the project can, the prototype has become a liability.
  • Confusing "I built this with AI" with "I vibe coded this." A lot of what gets called vibe coding is ordinary engineering accelerated by AI. Overusing the term for careful work muddies the genuinely useful distinction Karpathy drew.

The honest take

The honest take on vibe coding is that Karpathy named a real and useful behavior, and then the internet immediately over-extended it. The version that works is the version he described: low stakes, exploratory, "forget the code exists" — perfect for testing an idea before you commit real effort to it. The version that gets people in trouble is treating that same don't-read-it posture as a way to run software other people rely on. The skill in 2026 is not vibing or not vibing; it is knowing which mode a given piece of work calls for, and being willing to actually read the code the moment the stakes rise.

There is a broader pattern worth naming, and it is where this connects to what we do. Vibe coding is one instance of a bigger shift — describing an outcome to an AI in plain language instead of hand-assembling every step — and that same instinct now runs through content creation. Kompozy applies the describe-it-and-get-a-finished-result idea to content rather than code: you set a [Persona Brief](/glossary/persona-brief) and a topic, and the engine drafts and ships blog articles, carousels, short-form video, newsletters, and social posts across eight platforms plus blog and email, on brand, on schedule. The important difference is the guardrails. Where vibe coding's risk is unread code reaching production, a content engine keeps a human review pipeline and [quality gates](/glossary/quality-gates) in the loop, so the "give it the vibe, get the output" speed doesn't come at the cost of publishing something wrong. Same shift, higher stakes handled — the vibe is the input, not the excuse.

Frequently asked questions

What is vibe coding?

Vibe coding is building software by describing what you want to an AI model in natural language and letting it generate the code, rather than writing it yourself line by line. In its original sense you barely read the output — you prompt, run the result, describe the next change or paste in the error, and repeat, steering by results and feel. The term was coined by Andrej Karpathy in February 2025.

Who coined the term vibe coding?

Andrej Karpathy, a co-founder of OpenAI and former AI lead at Tesla, coined it in a February 2025 post on X, describing a style where "you fully give in to the vibes, embrace exponentials, and forget that the code even exists." The phrase spread within weeks — Merriam-Webster flagged it as trending in March 2025, and Collins English Dictionary named it Word of the Year for 2025.

Is vibe coding good or bad?

It depends entirely on the stakes. For prototypes, throwaway projects, and exploring an idea, it is fast and genuinely useful — Karpathy tied his original sense to exactly those low-stakes cases. For production systems with real users and data, shipping AI code without review raises well-documented quality, maintainability, and security risks. The skill is knowing which mode the work calls for.

What tools are used for vibe coding?

Two broad categories: agentic coding assistants that edit an existing codebase from natural-language instructions, such as Cursor and GitHub Copilot; and prompt-to-app platforms that turn a described idea into a running, hosted application, such as Lovable and Replit. The first fits developers who still review every change; the second targets people who want a working product without opening a file.

What is the difference between vibe coding and AI-assisted development?

Karpathy's original vibe coding means not reading the code — you trust the AI's output and fix issues by re-prompting. AI-assisted development means the AI drafts code but a human reviews, tests, and understands what ships. In practice the two get used interchangeably, but the distinction is important: they carry very different levels of risk, especially for anything that reaches production.

Can you vibe code a real production app?

You can start one that way, but shipping unread AI-generated code to real users is where the risks concentrate — logic errors and security vulnerabilities a human review would catch, plus maintainability problems when no one on the team understands the codebase. The realistic pattern is to vibe code the prototype to validate the idea, then bring human engineering discipline to anything that goes live.

Related terms

  • Agentic Loop — The repeating perceive-reason-act-observe cycle that turns a language model from a one-shot text generator into an agent that pursues a goal across multiple steps.
  • Context Engineering — Curating the whole set of tokens a model sees at inference — instructions, tools, references, memory, retrieved data — not just the wording of one prompt.
  • AI slop — Low-quality, generic media mass-produced by generative AI with little human oversight, and now the content audiences and platforms increasingly reject.
  • System Prompt — A system prompt is the foundational instruction block sent to an LLM before a conversation. It sets the model’s role, tone, rules, and priorities.
  • AI glossary (2026) — A plain-English reference to the AI terms creators actually run into in 2026 — LLM, token, prompt, hallucination, multimodal, agent, RAG, diffusion, fine-tuning, and inference — with what each one means for the person making content.
  • Quality gates — Four automated checks every Kompozy output passes before autopilot ships it: persona, platform-cadence, fact-anchor, brand-safety.
Related deep guides

← All terms · Get started →