What Is Vibe Coding? Origin and Definitions

Vibe coding is a term Andrej Karpathy introduced in a post on X dated February 2, 2025. It started as a description of prompting an AI and not reading the code it writes. Later sources use it more broadly for generating code from natural-language prompts. The definitions differ on one point: whether you review the code.

This page collects dated definitions from primary sources. We read them on 2026-10-03. We built no example projects and tested no tool, so the page contains no examples of our own and no claims about outcomes.

Who coined "vibe coding" and when?

What the sources say

• Collins Dictionary credits Andrej Karpathy as the person who coined the term. In a post dated November 6, 2025, Collins named "vibe coding" its Word of the Year 2025 (Collins Dictionary Language Blog). The post does not give the date of the original coinage.
• Simon Willison's post of March 19, 2025 says the term "was coined by Andrej Karpathy just a few weeks ago" (simonwillison.net). The post links Karpathy's original on X.
• IBM's explainer says Karpathy introduced the term in February 2025 (IBM Think).

Karpathy's post on X is dated February 2, 2025. We read the date and the opening lines through X's oEmbed endpoint, which returns the post text and date; the post page itself did not load for us. Willison's post gives a different day: it says the term was coined "on February 6th". We use the X date and note that the sources disagree.

How do primary sources define vibe coding?

What the sources say

Willison's post quotes Karpathy's original post in full. In it Karpathy says to "forget that the code even exists," writes "I don't read the diffs anymore," and calls the result "not too bad for throwaway weekend projects." We take the full text from Willison's quotation. The X embed data confirmed the opening sentences.

The definitions we could read are below. Each cell comes from the cited page.

Source	Date	Definition, as the source words it	Role of code review in the definition
Andrej Karpathy (X post; full text as quoted by Willison)	2025-02-02	"forget that the code even exists"	Stated in the post: "I don't read the diffs anymore"
Simon Willison	2025-03-19	"building software with an LLM without reviewing the code it writes"	Central: no review is the defining trait
Collins Dictionary	2025-11-06	"the use of artificial intelligence prompted by natural language to write computer code"	Not part of the definition
IBM Think	Published 2025-04-08; updated 2026-07-24	"the practice of prompting AI tools to generate code rather than writing code manually"	Not part of the definition; review gaps are listed as a risk

Willison draws a line against ordinary AI-assisted work. He writes that if an LLM wrote the code, you then reviewed and tested it thoroughly, and you could explain how it works to someone else, then "that's not vibe coding, it's software development." He says vibe coding is not the same as writing code with the help of LLMs.

What limits do the sources describe?

What the sources say

IBM's explainer lists five challenges: technical complexity with novel or complex real-world applications; code quality issues; difficulty debugging generated code; difficulty maintaining code whose logic the developer does not understand; and security vulnerabilities, because generated code can escape review. IBM describes "security debt" such as hardcoded credentials, SQL injection risks and inadequate authentication.

Karpathy's own framing, in the post as Willison quotes it, points to throwaway projects. That is a statement about intended use, not a measurement.

What to consider

These are our own observations. We tested nothing.

• The IBM list is a vendor explainer, not an audit. Treat it as a list of questions to ask, not as measured rates.
• The word "throwaway" matters. A project that real customers use is no longer throwaway, whatever you call the way you built it.

What do people mean when they say "vibe coding" today?

What to consider

This is our reading of the definitions above. The term covers three situations, and mixing them causes confusion.

1. Prompting without reviewing. This is Willison's narrow meaning and the closest to Karpathy's description. You accept what the AI produces and judge by running the app.
2. Prompting to produce an app. This is the broad dictionary meaning. A person with no programming background describes an app and receives a working version. Tools built for this are in our coding and app builders category, and the difference between them and editor-based assistants is in our guide to AI app builders versus AI coding assistants.
3. AI-assisted programming with review. A developer uses AI and reads, tests and owns the result. By Willison's definition this is not vibe coding. If this is what you want, see how to choose an AI coding assistant.

When you read an article, check which of the three it means. The risks and the right tools differ.

What do the sources leave unsaid?

What to consider

The sources above are definitions and an explainer. They leave several questions open, and this page does not answer them.

• How many people vibe code. None of the pages we read gives a usage figure we could verify, so we state none.
• How well it works. The definitions describe a method. They do not report how often the resulting apps work, stay secure or can be maintained.
• Where "vibe coding" ends. Willison draws the line at review. Collins and IBM do not. There is no single agreed boundary, so state which meaning you use when you write or buy.
• Which tools count. None of the definitions names a required tool. The term describes a way of working, not a product category.

How do you decide whether to vibe code a project?

These are decision rules from our assessment, derived from the definitions above.

• A throwaway prototype, a personal script or a demo. This matches the "throwaway weekend projects" framing. The cost of a mistake is low.
• Something other people will log in to, or something that stores personal data. The sources above say review gaps and security debt are the main risks. Read our page on vibe coding security risks and a pre-launch checklist before you publish.
• A product you expect to maintain for years. IBM lists maintenance as a challenge when you do not understand the code. Plan for someone to read it.
• A choice between a prompt-based builder and a visual no-code platform. See our comparison of AI app builders and no-code builders.

Limitations of this page

• The full text of Karpathy's post comes from Willison's quotation. X's embed data gave us the date and the opening lines only.
• We used three definitions (Collins, Willison, IBM). Other dictionaries and vendors may define the term differently, and we did not read them.
• Definitions are short snapshots. Dictionaries and vendors may revise them. The dates in the table are the dates we found.
• We include no example projects. We have not built any, and we will not describe outcomes we did not observe.
• We make no claims about how widely people use the term or the practice.

When is this page not the right resource?

• You want to learn to program. This page defines a term. It is not a course.
• You want a tool recommendation. We rank no tool here. Our category pages list tool profiles with documented features.
• You need a verdict on a particular app. A definition does not tell you whether a given app is secure or maintainable. That needs review of the app itself.