Vibe coding: How describing an app in plain words became a way to build software

Vibe coding is a way of making software by describing what you want in everyday language and letting an AI model write the code. The person steers by trying the result and asking for changes, not by reading or writing the code itself.

The phrase has become a label for a whole product category, from coding assistants to chat apps that build small tools on request.

A wedding planner built without opening a code editor

Imagine someone who needs to track RSVPs, invoices and to-dos for a wedding. A traditional route means a spreadsheet or hiring a developer. In the vibe coding route, they type: “Make me a wedding planner with a guest list, a budget and a checklist.”

The AI produces a working screen. The person clicks around, notices the budget total is missing, and types: “Add a running total at the top.” The screen updates. They repeat this until it feels right.

At no point do they need to know what language the tool is written in. That is the appeal, and also the catch.

The loop: Describe, run, react, repeat

Under the hood, a large language model (LLM) turns the description into program code. A tool then runs that code and shows the result. The person reacts to what they see, and the loop starts again.

Many current tools go one step further. An AI agent, meaning a model that can take several actions on its own, plans the pieces, writes files, runs them, reads error messages and fixes its own mistakes before showing anything. The user’s job shrinks to stating the goal and judging the outcome.

Two ingredients make this workable. Models have become much better at writing code, and tools now hold a project together across many requests, so the app is not rebuilt from scratch each time.

Where the phrase came from

AI researcher Andrej Karpathy, a co-founder of OpenAI and a former AI leader at Tesla, coined the term in a post on X in February 2025. He described a style in which you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.” He said he used Cursor Composer, a coding tool, by voice and barely touched the keyboard.

The term spread fast. According to Wikipedia, Merriam-Webster listed it as slang and trending in March 2025, and Collins Dictionary named it its Word of the Year for 2025.

What counts and what does not

Developer Simon Willison drew a useful line in March 2025. He defines vibe coding as “building software with an LLM without reviewing the code it writes.” If an AI writes code and a person reviews, tests and understands it, that is ordinary software development with a helpful tool, not vibe coding.

This distinction matters because the label is often stretched. The question is not whether AI helped. It is whether a human checked the work.

Willison’s own rule for serious work is that he will not commit code he could not explain to someone else.

Who is doing it today, including Meta

The idea now reaches beyond developers. Meta’s Muse app recently added Spaces, which let users describe an idea in chat and get a tracker, planner or project board in return. Meta calls a space “a custom experience you create inside Muse by describing your idea.” Our report on Meta’s Muse Spaces lists examples such as a trip planner and a small-business order tracker.

Meta’s own description does not say whether users can see or edit any underlying code, so it is not clear how close Spaces is to classic vibe coding. The shared idea is that a plain-language request becomes working software that the agent then maintains. Meta’s description also says connected apps can keep a space updated in the background, which goes beyond a one-off generated page.

Developer tools such as Cursor, GitHub Copilot and Replit Agent were named in coverage of the trend, as were app-building services like Lovable.

Where the vibes run out

The main risk is that nobody has checked the code. Wikipedia’s summary of the research and reporting lists several reported problems:

• Security gaps. In May 2025, 170 of 1,645 web apps generated by Lovable had a flaw that let anyone access personal information, according to the reported analysis.
• Maintainability. GitClear analyzed 211 million changed lines of code from 2020 to 2024 and found refactoring fell and duplication rose, though that study covers AI assistance in general, not only vibe coding.
• Slower, not faster. A July 2025 randomized trial by METR found developers using AI tools were 19% slower, though they believed they were faster.
• Destructive mistakes. In July 2025, a Replit agent deleted a founder’s production database despite instructions not to make changes, as reported.

Willison’s advice fits these cases: keep projects low stakes, protect secrets such as API keys, avoid usage-billed services without spending limits, and have an experienced developer review anything you plan to share publicly.

What to watch next

Three questions will shape the field. First, whether tools can add automatic checks, such as security scans, so unreviewed code becomes safer by default. Second, whether personal software, like a private tracker in a chat app, stays a hobby or becomes a normal way to use computers. Third, how companies word the line between a generated app and a saved chat, since products like Spaces blur it.

For now, the practical rule is simple. Vibe coding suits personal, low-risk tools. Anything that handles other people’s data or money still needs a person who can read the code.