Vibe coding

Vibe coding is describing the outcome you want in plain language and letting an AI model write the code, without reading most of it. Andrej Karpathy named it in February 2025 - "fully give in to the vibes... and forget that the code even exists"1 - and it entered Merriam-Webster's slang dictionary the same year.2

It is not a lower standard. It is a different division of labour: you keep the judgement about what should exist, and hand over the typing and the recall.

What it is not

Where it breaks

Vibe coding works beautifully and then stops rather suddenly, and the turning point is not the size of the thing. It is verification: the moment nothing can tell you whether the last change broke something, progress reverses.

Which is why tests extend vibe coding further than any tool choice does. A test replaces "I understand this code" with "the machine can tell me if I broke it" - exactly the signal you gave up by not reading. The tools piece goes into where each kind of tool sits relative to that line.

How to do it well

Two habits carry most of it. Verify continuously - keep something running that can tell you the thing still works, because that is the signal you traded away. And commit constantly, because your undo is git rather than your memory of what changed.

The rest is judgement about where to spend your attention, which is what the longer piece on the tools is for - including the five-step practice, the failure modes, and what to do when a session starts fighting itself.

Which tool for which job

The longer version, with the failure modes: the best vibe coding tools.

Check yourself

Three questions.

What defines the point where vibe coding stops working?

Losing the ability to verify that a change is safe Adding a database, once state stops being local Passing a certain size, measured in files or lines Reaching the context limit of the model you use

Karpathy's original framing was about forgetting the code exists. What does that trade away?

The ability to write code by hand at all, over time The option to use a different model later Any claim to authorship of the result The reader's own knowledge of what depends on what

You want to change an app that already exists. What fits?

A prompt-to-app builder that owns the scaffolding A coding agent in your editor or terminal A chatbot you paste the relevant files into A locally-run model, for speed on a familiar repo