Vibe coding

Building software by describing what you want to an AI and accepting the generated code without fully reading or understanding it. You guide the direction; the AI writes the code; you trust the vibes.

Andrej Karpathy coined the term in February 2025 to describe the experience of surrendering to the flow of AI-generated code. You tell the AI "build me a to-do app with drag-and-drop and dark mode," it produces hundreds of lines of working code, and you run it without reading the implementation. If it works, you move on. If it breaks, you describe the problem to the AI and let it fix its own code. Simon Willison later drew a useful line: if you read, test, and understand the output, that is AI-assisted development. Vibe coding specifically means you did not inspect what the AI changed.

Builder example

Vibe coding is genuinely powerful for prototypes, personal experiments, and throwaway tools where the stakes are low. It becomes dangerous when the result handles data, money, identity, security, or other people depending on it. The failure mode is subtle: vibe-coded software often works for the happy path and breaks on the first edge case, security probe, or unexpected input.

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

Before sharing it publicly, review auth, validation, storage, deployment secrets, spam protection, and deletion behavior.

Common confusion: People often use "vibe coding" to describe any AI-assisted development. The meaningful distinction is whether the human reviewed the output. Using AI to write code you then read, test, and understand is AI pair programming, not vibe coding.