What Karpathy meant by vibe coding
Correcting the attribution in our original caption and separating experiments from production review.
Originally posted · Published here · 1 source · Resume reading
Our original caption credited Windsurf's Varun Mohan with introducing the term vibe coding. That attribution was wrong. Andrej Karpathy used the term in his February 2025 post describing an unusually hands-off approach to building software with language models [1].
Read the original description
Karpathy described asking for changes, accepting generated code without reading the differences, and feeding errors back to the model. He also described losing track of the growing code and working around bugs. He placed that experiment in the context of disposable weekend projects [1].
That is more specific than using an AI assistant while understanding and reviewing its changes. The original post helps explain why two people can use the same label for very different working habits.
Keep responsibility visible
For a prototype, write down what the experiment needs to demonstrate and what can be discarded. For software that will handle real users or data, inspect the changes and test the behavior before release.
Ask the assistant to explain unfamiliar code, then verify the explanation against the implementation. Keep credentials and production access outside an exploratory task. A working screen is useful feedback, but it does not establish that storage, permissions and failure handling are correct.
The correction here concerns the term's attribution and original context, not a claim that every AI-assisted workflow has the same risks.
Keep this guide
Filed under AI agents
Sources
Next in this stack
Keyword LOOP
Five rules for building an AI agent that works
Keyword STARTER
Write skill metadata an agent can recognize
Keyword SIMPLEFIRST
Choose an AI workflow before building an agent