What is vibe coding?
Vibe coding is building software by telling an AI model what you want and accepting the code it writes without reading it. Andrej Karpathy coined the term in February 2025 for throwaway weekend projects. It is fast for prototypes and personal tools, and risky for anything other people depend on, because nobody has reviewed what shipped.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
You describe what you want in plain language, typed or spoken, to an AI coding tool such as Cursor or Claude Code.
The model writes or changes the code. You accept the changes without reading the diff.
You run the result. If it works, you ask for the next thing. If it breaks, you paste the error back and let the model fix it.
The cycle repeats until the thing does what you wanted. Nobody has read the code, which is fine for a toy and a problem for anything shared.
The moment you start reading and testing the code, it stops being vibe coding and becomes ordinary AI-assisted programming.
Source: Simon Willison: Not all AI-assisted programming is vibe coding
Vibe coding versus AI-assisted programming
The term gets used for any coding with AI, which loses the point. Karpathy's original post was about a specific mood: speak to the tool, accept every change, never look at the diff, and when something breaks, paste the error back in. The code "mostly works" and you move on. Simon Willison pinned the definition down a month later as building software with an LLM without reviewing the code it writes.
The review is the line. A senior engineer using Claude Code to draft a migration, then reading it, running the tests and fixing two edge cases, is not vibe coding. They are programming with a fast assistant. The same engineer letting the tool rewrite a payment handler and shipping it because the checkout page still loads is vibe coding, and the title on their business card does not change that.
Example: a weekend prototype
You want a page that turns a CSV of your running times into a chart. You open Cursor, describe the page, and accept what comes back. The chart is upside down; you say so and it flips. You ask for a dark theme, a filter by month, a share button. Two hours later you have a page you would never have bothered to build by hand, and you could not say which library draws the chart. That is the ideal case: one user, no secrets, nothing to lose if it breaks.
Where the output is a page like this, a shared board that teammates' agents can keep editing is one fair place for it to live, because the next person's agent can pick up the file without anyone having to understand it first.
Where it works and where it fails
It works when the cost of a bug is low and the cost of your time is the constraint. Prototypes, demos, internal dashboards, one-off scripts, personal tools, and anything you can throw away. It also works as a way to learn what current models can and cannot do, which is useful even for people who would never ship the result.
It fails when the software touches other people's data or money, when it has to be maintained, or when someone will later need to change it. Generated code tends to work before it is safe: it handles the path you tried and trusts whatever arrives on the paths you did not. Karpathy himself moved on in February 2026, proposing "agentic engineering" as the name for the professional version, where you orchestrate agents and act as oversight. Vibe coding kept the meaning he gave it: no oversight, by choice.
Common mistakes
- Shipping the prototype. The demo works, so it goes live. Nobody checked what it does with a malformed upload or a second user. Treat a vibe-coded prototype as a spec for the real thing, not as the thing.
- Pasting errors forever. If the fourth fix for the same error is still not working, the model does not understand the problem. Read the code at that point, or start again with a clearer brief.
- Hiding the method. If a teammate will touch the code, say it was generated and unreviewed. They will read it differently, and they should.
FAQ
- Who coined the term vibe coding?
- Andrej Karpathy, in a post on X on 2 February 2025. He described giving in to the vibes, forgetting the code exists, clicking Accept All without reading diffs, and pasting errors back in. He called it fine for throwaway weekend projects. Collins Dictionary named it word of the year for 2025.
- Is vibe coding the same as using AI to write code?
- No. The difference is review. If you read, test and understand the code an AI writes, that is AI-assisted programming, whatever the tool. Vibe coding is specifically the version where you skip that step and judge the result only by whether it seems to work.
- Is vibe coding safe for production software?
- Not on its own. Code nobody has read can leak data, trust the wrong input or delete the wrong thing while looking fine from the outside. Use it for prototypes, personal tools and low-stakes pages. Anything other people rely on needs a review, tests and someone who can explain the code.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.