What is vibe coding?

Vibe coding is a way of building software by describing the desired outcome in natural language and accepting the code an AI writes, largely without reading it. The builder works through prompts and results: describe a feature, run what comes back, describe what's wrong, repeat. The code itself is treated as an implementation detail the AI owns.

Andrej Karpathy, a co-founder of OpenAI and former AI director at Tesla, coined the term in a post on X on 2 February, 2025, describing "a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." The name stuck: by November 2025, Collins Dictionary had named vibe coding its Word of the Year.

How does vibe coding work?

A vibe coding session is a loop. The builder tells an AI tool what to build, the tool generates and often runs the code, and the builder reacts to what appears on screen: approving changes, describing fixes in plain language, and pasting error messages straight back into the chat. There's no upfront specification and no code review; natural language is the entire interface. Karpathy's original post captured the posture: "I 'Accept All' always, I don't read the diffs anymore."

The workflow runs on 2 families of tools. Browser-based app builders like Lovable, Bolt, v0, and Replit generate, host, and deploy full applications from a prompt, which makes them the common entry point for non-engineers. Agentic coding tools like Cursor and Claude Code work inside real codebases and suit builders who want more control over files, versions, and deployment.

How does vibe coding differ from AI-assisted engineering?

The boundary is review. Simon Willison, one of the most-cited writers on the practice, argues that vibe coding specifically means building with an LLM without reviewing the code it produces. A developer who reads, tests, and understands everything the model writes is doing AI-assisted programming. In Willison's words: "that's not vibe coding in my book — that's using an LLM as a typing assistant." Collapsing the two terms, he warns, gives a false picture of both.

The disciplined variant has since acquired its own name. In early 2026, Karpathy himself described agent-driven development with heavy oversight (plans, specifications, verification, and real code review) as the emerging professional default, a practice now widely called agentic engineering. Vibe coding trusts the output; agentic engineering verifies it.

Why does vibe coding matter for designers and product teams?

Vibe coding collapses the distance between an idea and a working artifact. Designers can turn a concept into a functional prototype with real data and real interactions instead of a static mockup, and product managers can validate demand with a working slice of product before committing engineering time. The shift is visible at scale: Y Combinator reported in March 2025 that a quarter of its Winter 2025 startup batch had codebases that were roughly 95% AI-generated.

It also feeds a quieter movement: personal, home-cooked software. People build small tools for one household, one team, or one workflow, software that was never worth a development budget but is now a weekend prompt session. For designers, the skill that matters shifts from writing code to articulating intent precisely and judging whether what came back is actually good.

What are the risks of vibe coding?

Unreviewed code carries unexamined failure modes. Vibe-coded apps tend to work on the happy path, the exact flow the builder tested, while edge cases, malformed input, and hostile users go unhandled. Security is the sharpest risk: AI-generated code routinely hardcodes API keys, skips access control, and ships vulnerabilities that nobody catches because nobody is reading. Costs bite too when an app calls paid services with no spending limit attached.

The practical mitigations are workflow habits rather than coding skill: keeping secrets in environment variables instead of the code itself, setting spending limits on connected services, asking the AI to explain and audit its own output, and treating anything vibe-coded as prototype-grade until a review happens. When a project starts handling real users, payments, or personal data, it has outgrown pure vibe coding.

How has vibe coding evolved since 2025?

The term went from a self-described throwaway post to Collins' Word of the Year within 9 months, and the backlash arrived almost as fast. As vibe-coded projects reached production, stories of leaked keys, surprise bills, and unmaintainable codebases pushed practitioners to narrow the term back toward what Karpathy originally meant: a casual mode, fit for prototypes, experiments, and personal tools.

By 2026 the practice had split into two registers. Vibe coding survives as the on-ramp, the fastest way for a non-engineer to get from idea to working software, while professional teams moved toward agentic engineering: the same AI leverage wrapped in specifications, tests, and review. The distinction is no longer about whether AI writes the code; it's about who verifies it.