Vibe coding

Vibe coding is a way to build software by describing the desired result, running what AI generates, and steering through visible behavior while paying little attention to the underlying code.

In vibe coding, the main feedback loop is conversational and experiential. You describe a to-do app, run the generated version, report that dragging feels wrong, and ask the AI to repair it. You may never learn which files or functions produced the behavior. Andrej Karpathy introduced the phrase in February 2025 for a deliberately code-oblivious style of creation. The label now gets applied too broadly, so the useful question is concrete: does the builder steer mostly through the result, or can the workflow also inspect and verify the implementation when the stakes demand it?

Builder example

The approach lowers the cost of experiments and personal tools because it lets someone work at the level of intent and behavior. Its limits appear when unseen implementation details determine privacy, security, money, accessibility, or maintainability. Reading every line is not the only answer. Tests, permission boundaries, dependency checks, runtime monitoring, and a recoverable deployment can establish properties that visual trial and error cannot.

A founder builds a waitlist app in an afternoon. It works locally and stores email addresses.

Before sharing it publicly, define and test authentication, validation, storage, secrets, abuse controls, and deletion behavior.

Common confusion: Vibe coding is not a synonym for all AI-written code, and one manual review does not create a permanent category change. A project can begin with result-driven exploration and later add an engineering contract that makes consequential behavior observable and testable.