Coinage and definition

On 2 February 2025, Andrej Karpathy introduced the term “vibe coding” in a public post on X (formerly Twitter), defining it as a programming workflow where developers rely primarily on large language models (LLMs) to write and evolve software through conversational requests rather than manual coding. The post described a shift away from traditional code comprehension and toward a development style guided by real-time interaction with AI systems.

Karpathy characterised the method as “see stuff, say stuff, run stuff,” implying that the human role becomes one of supervision, intention-setting, and feedback rather than explicit implementation. Instead of reading diffs, planning architecture, or typing line-by-line logic, practitioners steer the model through natural-language instructions, iterative corrections, and runtime observation.

The phrase spread rapidly across developer communities, social media, and AI-focused forums, where it was interpreted both as a cultural meme and a legitimate redefinition of programming practice in the era of highly capable LLMs. The term was soon referenced in technology journalism, startup documentation, and industry commentary as part of an emerging shift toward conversational development environments.

LLM-driven workflow

In vibe coding, natural-language prompting replaces much of the explicit manual implementation traditionally required in software development. Instead of writing functions, modules, or architectural scaffolding by hand, the practitioner describes the expected behaviour in plain language, executes the generated code, and then iterates based on what happens at runtime.

The workflow depends on rapid feedback loops: the developer runs the program, observes outcomes, and provides error messages, missing logic, or new requirements back to the model. Rather than design documentation or fully reviewed pull requests, the conversation itself becomes the primary mechanism of refinement.

Tools commonly linked to this workflow include Cursor, Claude Code, Replit Agents, GitHub Copilot, Bolt, v0 and other conversational or real-time AI coding environments. These platforms support code generation, refactoring, and debugging directly through chat-based or voice-based interfaces, reducing the need to navigate files manually.

By March 2025, the concept had gained enough traction to appear in Merriam-Webster’s “Slang & Trending” feature, marking its transition from niche online jargon to a recognised term describing a distinct style of AI-assisted programming.

The approach builds on Karpathy’s earlier idea that “the hottest new programming language is English,” positioning natural-language intent as the primary interface layer between the human developer and the machine.

Media amplification and adoption

Following its introduction, the term “vibe coding” spread quickly through online discussions, developer forums, and AI-centred communities. Technology journalists and commentators began examining whether the workflow represented a temporary trend or a structural shift in how software would be built.

Mainstream media soon conducted public experiments to test the approach. The New York Times technology columnist Kevin Roose created “software-for-one” prototypes using conversational prompting alone, demonstrating that non-specialists could build functional tools without writing traditional code. These experiments contributed to broader public awareness and debate.

Adoption also emerged within the startup ecosystem. Y Combinator reported that approximately 25% of companies in its Winter 2025 batch operated codebases that were at least 95% AI-generated, signalling institutional willingness to treat AI-authored systems as production-viable rather than experimental. Internal tooling, automations, and early-stage prototypes increasingly reflected conversational rather than manual development patterns.

Coverage expanded across outlets including Wired, TechCrunch, and The Wall Street Journal, which documented teams using vibe-style workflows in real engineering environments. While many organisations supplemented the process with traditional review and debugging practices, the trend indicated that vibe coding had moved beyond an online joke or meme and into emerging practice.

Influencers & Thought Leaders—Narrative Control Points

A small but influential group of technologists and journalists has shaped the perception and trajectory of vibe coding. Their endorsements and critiques have created a powerful narrative flywheel, and quoting them can lend authority to any analysis of the space.

Proponents: Karpathy, Nat Friedman, Swyx

As the originator of the term, Karpathy is the central proponent. His posts on X and conceptual work on "Software 2.0" laid the groundwork for the movement . He champions a hands-off, AI-guided loop, though he acknowledges that current models struggle with novel code "that has never been written before".

Nat Friedman & Swyx As prominent investors and thinkers in the AI space, their advocacy for AI-assisted development has amplified the trend, framing it as a massive productivity unlock for the entire industry.

Media Echo Cycle Timeline

The narrative around vibe coding has followed a predictable media cycle. It began with Karpathy's initial tweet, was amplified by explanatory articles in tech-focused outlets like Ars Technica, gained mainstream legitimacy through features in *The New York Times* and *The Wall Street Journal*, and was finally tempered by cautionary tales in publications like *Fast Company*.

Human-in-the-loop review cadence

Automated tools cannot catch all flaws. Mandating regular, security-focused code reviews for all AI-generated code is crucial for identifying nuanced logical errors, architectural issues, and insecure business logic that tools might miss. This human oversight is the last line of defense against a vibe coding hangover.

Critiques and concerns

Sceptics caution that vibe coding can obscure ownership, intent, and accountability within the software development process. Because much of the code is generated rather than written manually, teams may struggle to determine who is responsible for decisions embedded in the system. This lack of transparency has raised concerns in both academic and industry discussions.

Cognitive scientist Gary Marcus argued that many early demonstrations of vibe coding relied on remixing existing code patterns rather than producing genuinely novel algorithms or reasoning. His critique positioned vibe coding not as a breakthrough in computational creativity, but as an acceleration of pattern-matching development without conceptual understanding at its core.

Software engineer Simon Willison similarly warned that bypassing traditional code review introduces long-term risks. Without human oversight of architecture, security implications, and design decisions, teams may face significant challenges in onboarding new developers, maintaining systems, and performing audits. He described vibe-generated codebases as “high-entropy systems that work—until they suddenly don’t.”

By September 2025, Fast Company reported what it called a “vibe coding hangover,” as several organisations faced outages and production incidents linked to opaque AI-authored code that lacked clear structure, documentation, or traceability. One widely referenced example involved the Leo and data leak of Tea App, which accidentally leaked authentication keys generated. The event became a cautionary symbol of the risks associated with blindly trusting AI-authored configurations and secrets.

In response to these challenges, new support ecosystems began forming around the workflow. Platforms such as VibeCodeFixers.com emerged to help developers debug, secure, and stabilise AI-generated systems, offering specialised assistance for error handling, refactoring, and post-generation hardening. These services reflect a growing recognition that vibe coding may accelerate creation—but often requires structured intervention to ensure reliability and safety at scale.