Vibe coding is an approach to software development in which a person describes what they want in natural language and lets an AI coding assistant generate, run, and iterate on the code, largely without reading or editing the generated code by hand. The term was coined by Andrej Karpathy in a February 2025 post on X, where he described "fully giving in to the vibes" and forgetting that the underlying code even exists, accepting suggestions from tools like Cursor without close inspection.
Origins and definition
Karpathy's post described a workflow enabled by capable large language models and agentic editors: state an intent, let the assistant write the implementation, run it, paste any errors back in, and repeat, treating the model's output as a black box guided by feel rather than verified logic. The term spread quickly through developer communities and was added to Merriam-Webster's word watch list later in 2025, reflecting how common the practice had become among hobbyists, prototypers, and some professional teams. It builds on tools such as GitHub Copilot and Claude Code, as well as no-code and low-code AI app builders that let non-programmers assemble working software through conversation alone.
Practice and tooling
Vibe coding typically pairs a chat-style or agentic assistant with a live preview or an automated test-and-run loop: the assistant edits files, executes the program or its test suite via tool use, observes the output or error, and revises its own code without the human inspecting each change. This differs from earlier AI-assisted coding, where a developer reviewed every suggestion before accepting it. The practice is most associated with rapid prototyping, small internal tools, hackathon projects, and solo "indie" products, where the cost of an undiscovered bug is lower than in production infrastructure.
Reception and criticism
Supporters argue vibe coding radically lowers the barrier to building software, letting people without programming training create working applications and letting experienced engineers iterate far faster on throwaway prototypes. Critics warn that code produced this way often contains security vulnerabilities, poor error handling, and architectural debt that surfaces once a project scales, since no human has verified the logic; several widely discussed 2025 incidents involved vibe-coded products leaking data or breaking in production. Some commentators distinguish "vibe coding" from "AI-assisted coding" or "agentic coding" precisely on this axis, whether a human reviews the generated code before it ships, and the term has been extended loosely to describe similarly low-scrutiny AI use in other creative fields, a habit some critics link to the wider rise of AI slop. As of 2025, vibe coding remained more common in casual and prototype contexts than in codebases with strict compliance, security, or reliability requirements, where teams still rely on code review and testing regardless of how the code was drafted.