Vibe coding

Vibe coding, or vibing, is a [low-code programming] [methodology] that uses artificial intelligence (AI) tools to generate code based on natural language [requirements specifications].

Vibe coding is one end of a spectrum whose other end is agentic engineering. The differentiator is how much structure, verification, and human judgment surrounds the AI’s output. Without both tests and evals, the practice is always vibe coding, however sophisticated the prompts.

The term was introduced by Andrej Karpathy – a co-founder of OpenAI and former AI leader at Tesla – in a post on social media platform X in February 2025:

There’s 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. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I’m too lazy to find it. I "Accept All" always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away. It’s not too bad for throwaway weekend projects, but still quite amusing. I’m building a project or webapp, but it’s not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

– Andrej Karpathy
https://x.com/karpathy/status/1886192184808149383

The term quickly gained traction in the tech community, with many developers and AI enthusiasts sharing their own experiences and interpretations of vibe coding. A March 2025 article on Ars Technica described vibe coding as being "all about surrendering to the flow", as opposed to "being about control and precision".

The concept has been compared to pair programming, where a human programmer collaborates with an AI assistant to create software. Certainly, vibe coding shifts the role of the computer programmer from writing code to defining and iteratively refining requirements, and then testing the output of the AI-generated code – the "driver" role in the pair programming methodology.

Tools that may be used for vibe coding include Cursor, Replit, GitHub Copilot, Anthropic Claude, Google Gemini, and Sonnet. These tools use large language models (LLMs) to generate code based on natural language prompts. For example, a user might ask the AI to "create a web app that allows users to track their fitness goals" and the AI would generate the necessary code to create that application.

Vibe coding is a form of [low-code/no-code programming] that makes software development more accessible to non-programmers, allowing them to create applications without deep knowledge of any particular programming language, application framework, or runtime environment.

Vibe coding is particularly well-suited to [rapid prototyping] and development of [proof-of-concept] software. Vibe coding lacks the precision and [correctness] of traditional programming methods, raising questions about the quality of the generated code. Therefore, vibe coding is unsuitable for the programming of business-critical production software.

Vibe coding names the review stance, not the resulting code’s quality. The stance is accepting the agent’s code without human review — the diff treated as opaque; what matters is whether the program behaves. Don’t use "vibe coding" as a synonym for "low-quality AI coding." It trades inspection for speed. Reading diffs is usually the slowest step, so dropping it removes the main bottleneck. For code whose failures are cheap — prototypes, one-off scripts, internal tools — that’s a reasonable trade. The risk scales with the code’s lifespan and stakes.

The cost arrives later. Vibe-coded changes accumulate into a codebase nobody has read, and behaviour was the only thing checked, so anything behaviour doesn’t surface — a secret written to logs, a missing edge case, quietly wrong data handling — ships unseen. The first time someone debugs the system is the first time anyone reads the code. With human review gone, whatever automated verification still runs (tests, types, automated review) is the only gate. This is the opposite end of the spectrum from the software craftsmanship stance, which treats the code itself as the artifact to be honed and read.

The 80% problem

AI agents rapidly generate roughly 80% of the code for a feature, but the remaining 20% — edge cases, error handling, integration points, and subtle correctness — demands deep contextual knowledge that current models often lack. The nature of AI errors has shifted from syntax mistakes to conceptual failures: wrong assumptions about business logic, failure to seek clarification on ambiguous requirements, missing edge cases, and architectural decisions that create subtle long-term maintenance burdens. These are harder to detect precisely because the code "looks right" and may pass basic tests. See agentic engineering and automated testing.

Hidden debt

Vibe coding’s near-zero entry cost hides a compounding operational burden:

  • Token burn rate: dumping unstructured files into context and looping on unverified mistakes.
  • Maintenance tax: bugs surface months later and engineers must reverse-engineer unstructured AI-generated spaghetti.
  • Security remediation: rapid code generation produces rapid vulnerability generation, each flaw exponentially cheaper to catch in design than in production.

See large language model for the token economy, and agentic engineering for the disciplined alternative.

See also prompt engineering, agentic engineering, and spec-driven development, a reaction to vibe coding that front-loads intent into a persistent specification.

References

  • The New SDLC With Vibe Coding, Addy Osmani, Shubham Saboo & Sokratis Kartakis (2026).