AI Atlas
Beginner · ~2 min read #vibe-coding #coding-agent #ai-assisted-programming

Vibe Coding

Building software with AI by looking only at the result, not the code

A way of programming, named by Andrej Karpathy in February 2025, in which you tell an AI in plain language what you want, accept the code it writes without reading it, and move forward only by checking whether it works.

VIBE CODING ≠ REVIEWED AGENTIC CODINGVIBE CODING · KARPATHY, FEB 2025"forget that the code even exists"ASKvoice / plain wordsACCEPT ALLdiffs unreadRUNerror → pasteworks?ask againfits: a throwaway weekend projectREVIEWED AGENTIC CODINGPLANDIFFTESTREVIEW- if ($user) {+ if ($user?->can('update', $card)) { tests: 42 passed # can you explain every line?HUMAN APPROVAL → MERGEneeded: production, data, payments, auththe tool can be the same; the difference is whether the output gets reviewed
Definition

Andrej Karpathy coined the term in a post on X on 2 February 2025. It opens: "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."

In the same post he describes the method: he talks to Cursor Composer by voice, always hits "Accept All", no longer reads the diffs, pastes error messages in with no comment, and when the LLM can't fix a bug he works around it or asks for random changes until it goes away. He also draws the boundary himself: "It's not too bad for throwaway weekend projects."

The term spread fast; Collins Dictionary named "vibe coding" its Word of the Year for 2025. As it spread, its meaning widened, and many people now use it for any code written with AI. But what distinguishes the original definition is the absence of review. As Simon Willison puts it: if an LLM wrote the code and you reviewed it, tested it and can explain how it works to someone else, that is not vibe coding, it is software development.

So the difference between vibe coding and reviewed agentic coding is not the tool but the process. Both can use the same coding agent. In the reviewed flow the agent produces a plan and a diff, tests and linters run, a person reads the diff and only then merges. In vibe coding those gates are skipped on purpose.

Analogy

Vibe coding is like getting into a taxi in a foreign city and saying "take me somewhere nice" without ever looking at a map. On holiday, when getting lost doesn't matter, it's fun, and you usually end up somewhere nice.

But if you have a meeting to make, you check the route, watch the meter and say "stop" at the wrong turn. That is reviewed agentic coding: the driver is the same, but the route goes through your approval.

Real-world example

On a Saturday you want a small page showing the training schedule for your child's football team. You tell an agent "make a single-file page that looks like a calendar and works on a phone". You look at the result, say "change the colours to the team colours", and paste the error message when something breaks. You never open the code. Half an hour later it's done. That is pure vibe coding, and it fits here: no data, no logins, nobody gets hurt if it breaks.

The same week at work you are on a page that updates customers' payment details. You use the same agent, but this time you ask for a plan first, run the tests, read the diff line by line and get a teammate to review the PR as well. Same tool; different process.

Code examples
From vibe coding to a reviewed flow · same agent, different gates bash
# Vibe coding: ask, accept, run, look
claude -p "Make a single-file page for the training schedule"

# Reviewed flow: separate branch, tests, read the diff, then merge
git switch -c feature/payment-form
claude "Add IBAN validation to the payment form and write tests for it"
php artisan test
git diff main --stat && git diff main
# Commit and open a PR if you understand the diff; ask if you don't
When to use
  • Throwaway prototypes, demos and weekend projects
  • When you want to find out within a few hours whether an idea works at all
  • Small personal tools: scripts and pages that hold nobody's data and aren't exposed to the internet
  • For someone new to programming to see what is possible (without replacing actually learning it)
When not to use
  • Any code involving user data, payments, authentication or authorization
  • Production code you will maintain for months, together with other people
  • In an environment with access to secrets (API keys, database passwords) or a live server
  • Work whose failures don't stay with you: systems used by customers, a team or the public
Common pitfalls

Shipping the prototype to production

If a vibe-coded prototype goes live because it works, code nobody has read is now in production. Treat the prototype as a sketch; write the production version with review, or have it rewritten.

Missing security holes

Someone who doesn't read the code can't notice that the agent put an API key on the client side, skipped input validation or left out an authorization check. The app appears to work but is wide open.

Papering over bugs with random changes

The 'ask for random changes until it goes away' approach Karpathy describes hides bugs whose cause nobody understood. They can come back somewhere else, at a worse moment.

Calling all AI-assisted code vibe coding

Using the term for every kind of AI-assisted programming makes reviewed, tested agentic work look careless too. What defines vibe coding is not that a model was used but that its output was accepted without review.