What Is Vibe Coding? Building Software Without Reading It

In brief

Vibe coding is building software by prompting an AI and accepting its code without reading it. Andrej Karpathy coined the term in an X post on 2 February 2025. Collins made it Word of the Year on 6 November 2025, but its definition leaves out the part that defines it: not reading the code.

What is vibe coding?

Vibe coding is a loop of four moves, and each one comes from Karpathy's own post:

  1. Ask in plain language, even for something as small as "decrease the padding on the sidebar by half".
  2. Click "Accept All" and skip the diff.
  3. When an error appears, paste it back in with no comment.
  4. If a bug won't go, "ask for random changes until it goes away".
A silver OpenAI knot, a terracotta Claude starburst and a rainbow glass Gemini star around a small toy house made of loose glowing blocks, with one block floating out of place.

Karpathy posted it on X on 2 February 2025. Here is the sentence that started the term: "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." He said it was possible because models like Cursor Composer with Sonnet were "getting too good". He talked to the tool through SuperWhisper and barely touched the keyboard. He also called the result "not too bad for throwaway weekend projects" and said that "it's not really coding".

One X post to Word of the Year in nine months
2026CoinedKarpathy's post on XDefinedWillison draws the lineFowler separates agentic engineeringDictionaryCollins Word of the Year

The definition hangs on one clause, and it's a clause about you, not the AI. Martin Fowler, who wrote a short entry on the term in May 2026, calls "forget that the code even exists" the key point. Simon Willison, a developer who writes about AI tools, put it as a test in March 2025: "When I talk about vibe coding I mean building software with an LLM without reviewing the code it writes." Ask yourself one question. Did you read the code? If you did, you weren't vibe coding.

Is vibe coding the same as AI-assisted coding?

No. Vibe coding is the subset of AI-assisted coding where nobody reads the output. Willison drew the line in his March 2025 post: "If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone else that's not vibe coding, it's software development." Fowler holds the same line and gives the careful version a name, agentic engineering.

The dictionary is where the line blurs. Collins defines vibe coding as "the use of artificial intelligence prompted by natural language to assist with the writing of computer code." That covers a developer who reviews every line. Of the four primary definitions below, only Collins drops the no-review condition.

SourceDateSays you skip reading the code?Its wording
Andrej Karpathy, X post2 Feb 2025Yes"I don't read the diffs anymore"
Simon Willison, blog19 Mar 2025Yes"without reviewing the code it writes"
Collins Dictionary, Word of the Year6 Nov 2025No"to assist with the writing of computer code"
Martin Fowler, bliki21 May 2026Yes"without looking at any of the code that the LLM generates"

Fowler notes the drift himself: "Sadly the term "vibe coding" really caught on, so many people use it to mean agentic engineering." So when you read "vibe coding" in a headline, check which meaning it carries. A story about a developer who ships reviewed AI code is about something else. The AI agent explainer covers the tools that make that second practice possible.

What is vibe coding good for, and when is it risky?

Vibe coding is good for software only you will use. Cursor, a maker of vibe coding tools, says so on its own help page: "Vibe coding is strong for prototypes, experiments, and internal tools, where speed matters and the stakes are low." Fowler agrees that it suits "disposable software written for a limited audience", and that people with no programming knowledge can use it to build applications for themselves.

It's risky the moment someone else depends on the result. Cursor's page is blunt: "We don't recommend shipping software you don't understand to production." Fowler names security as the most serious risk, because vibed applications can expose sensitive information or credentials. Veracode's 2025 GenAI Code Security report tested more than 100 models across Java, JavaScript, Python and C# and found "risky security flaws in 45% of tests". That's the vendor's own figure, and it measures what models write, not what a reviewer would catch.

Willison's checklist for people starting out is the most usable one:

  1. Keep the project low stakes. Ask who could be harmed if it has a bug.
  2. Watch for secrets. An API key in code you haven't read can leak.
  3. Think about private data before pasting it into anything you built.
  4. Put a billing limit on anything that charges by use. He has heard of people who ran up thousands of dollars in charges.
  5. If other people will use it, ask someone more experienced for a check first.

There's a second cost that shows up later. Fowler warns that code nobody reads piles up into low quality that is hard for a person or a model to change, and that well-structured code makes life easier for LLMs too. A prototype that survives a month tends to get rewritten, and the rewrite is where reading the code starts.

How do you start vibe coding?

A first vibe coded project takes five moves, and the first is picking something nobody else depends on:

  1. Choose a low-stakes idea, such as a mile-to-kilometre converter or a personal checklist page. Willison's first rule for beginners is that "Projects should be low stakes".
  2. Describe it in plain language, one small request at a time. A first prompt could be: Make one page with a box where I type miles and see kilometres, with a clean layout.
  3. Run the result and judge what it does, not what the code says.
  4. When something breaks, paste the error back in with no comment, as Karpathy did.
  5. Before you share it or feed it real data, go through the checklist above, or hand the code to someone who will read it.

The loop only works while mistakes are cheap. A converter that gets a number wrong costs you a minute. A tool that stores other people's details and gets security wrong is a different matter, and security is the risk Fowler calls the most serious, so that is the point to start reading.

Which tools do people use to vibe code?

Four tools cover most first attempts, and each one reaches a different kind of beginner. This list is short on purpose. The full comparison is in vibe coding tools, and the plan facts below were read from each vendor's own page in November 2026.

ToolWhat it buildsPlan fact
Claude ArtifactsPages and small interactive tools that open beside the chatIncluded on Free, Pro, Max, Ultra, Team and Enterprise, per Anthropic's help center
CursorCode in your own project, through an AI agent in the editorIts help page lists Design Mode, MCP integrations and Bugbot review; see Cursor pricing
LovableFull web apps from a promptFree plan gives 5 build credits a day, up to 30 a month
Claude CodeChanges to a real codebase, run from the terminalNot on Claude Free; included on Claude Pro, $20 a month billed monthly or $17 on the annual plan, per Anthropic's pricing page

Pick by how much you want to see. Artifacts is the safest start, because Willison points out that its code runs in a locked-down sandbox that can't make network requests to other sites. That also limits what it can build. Cursor and Claude Code have fewer guardrails, which is why they suit people who will review the output. Best AI for coding ranks the coding agents by score and cost if you want the numbers.

What do you do when the AI can't fix its own bug?

Karpathy hit this wall in his first post: "Sometimes the LLMs can't fix a bug so I just work around it". The usual next move is to ask a different model the same question. A second model reads the same error without the first one's earlier wrong turns in its history, and that often breaks a loop.

Whizi Pro puts Claude, GPT and Gemini in one app, and its first week is $1.99. On Pro you can run two models side by side on the same prompt, which suits a stuck bug: paste the error once and compare both fixes before you accept either. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku 5.5, with no card. Whizi is a chat app. It doesn't build or host your app, so keep your vibe coding tool for that. Coding with multiple models shows the workflow.

Workflow checklist
  • Ask one question to tell if it is vibe coding: did you read the code?
  • Keep vibe coded projects to prototypes, experiments and tools only you use
  • Never paste an API key into code you have not read
  • Set a billing limit before pointing a vibe coded app at a paid API
  • Ask someone experienced to check anything other people will use
Common questions

Frequently asked questions

Who coined the term vibe coding?

Andrej Karpathy coined it in an X post on 2 February 2025. He wrote that there is "a new kind of coding I call "vibe coding"" where you "forget that the code even exists". Collins Dictionary later named the term its 2025 Word of the Year, announced on 6 November 2025.

What does vibe coding mean?

Vibe coding means building software by describing what you want to an AI, running the result and asking for changes, without reading the code it writes. That's Karpathy's meaning and the one Willison and Fowler keep. Collins' dictionary entry is broader and covers any use of AI to help write code.

What is the difference between vibe coding and agentic engineering?

In Martin Fowler's usage, a vibe coder does not look at the code the AI writes, while someone doing agentic engineering has the AI write it but still reviews it and pays attention to its structure. He keeps the two terms separate, even though he notes that many people now use "vibe coding" for both.

Can you vibe code with no coding experience?

Yes. Martin Fowler writes that because a vibe coder doesn't look at the code, "they don't need programming skills", so it suits building applications for your own use. The limit is trust. Keep it to low-stakes projects, because you can't judge security or quality in code you can't read.

Is vibe coding safe?

It is safe for throwaway projects and unsafe for anything other people rely on. Veracode's 2025 report found risky security flaws in 45% of AI code tests, and Fowler calls security the most serious risk. Cursor says it does not recommend shipping software you don't understand to production.