If you have spent any time in software circles over the past year, you have almost certainly encountered the term vibecoding (also written as "vibe coding"). What started as a throwaway tweet has become one of the defining concepts in modern software development, a shorthand for a new way of building software with AI.
This guide covers what vibecoding actually means, the tools that make it possible, who it is for, and where it is heading.
The Origin: Andrej Karpathy Coins "Vibe Coding"
In February 2025, AI researcher and former Tesla AI director Andrej Karpathy posted on X (formerly Twitter) describing a new approach to programming. He called it "vibe coding": a style where you describe what you want in plain language, let an AI model generate the code, and guide the result through conversation rather than manual editing.
Karpathy described accepting all AI-generated changes without reading diffs, pasting error messages back into the chat to fix issues, and essentially treating the codebase as something the AI manages while you steer the direction. The post resonated widely because it named something many developers were already beginning to experience: a shift from writing code line-by-line to directing an AI collaborator.
Within weeks, "vibe coding" had entered the mainstream vocabulary of software engineering. A year on, it has evolved from a provocative idea into an everyday practice at companies around the world.
From Meme to Method: The Book That Defined Vibecoding
Karpathy's tweet captured a feeling, but it was Vibe Coding: Building Production-Grade Software With GenAI, Chat, Agents, and Beyond by Gene Kim and Steve Yegge (published October 2025, with a foreword by Dario Amodei, CEO of Anthropic) that gave vibecoding its current, more serious meaning.
Where Karpathy described accepting AI output without reading it, Kim and Yegge reframed the concept as a rigorous engineering practice: directing AI agents to build, test, and ship production-grade software. Their book treats vibecoding not as a shortcut but as a discipline, one that demands clear intent, architectural thinking, and systematic verification.
The authors bring considerable weight to the subject. Gene Kim wrote The Phoenix Project and The DevOps Handbook, two of the most influential books in modern software operations. Steve Yegge is a veteran engineer and technical leader known for his years at Google, Amazon, and Sourcegraph. Together, they make the case that vibecoding is the next major shift in how professional software gets built.
This site uses vibecoding in the sense Kim and Yegge define it: software engineering with agentic AI, applied with the same rigour you would bring to any production system.
How Vibecoding Works in Practice
At its core, vibecoding replaces the traditional cycle of think → type → compile → debug with something closer to a conversation. The typical workflow looks like this:
- Describe your intent. You tell the AI what you want to build or change, using natural language. This can be as high-level as "add a user settings page with dark mode toggle" or as granular as "fix the off-by-one error in the pagination logic."
- Review the generated code. The AI produces code across one or more files. Depending on your comfort level, you might review every line, skim the structure, or accept changes wholesale.
- Iterate through conversation. If the result is not quite right, you refine it by describing what to change. Error messages, screenshots, and test output can all be fed back to the AI as context.
- Test and ship. You run the application, verify behaviour, and deploy. The AI can also help write tests, fix CI failures, and handle code review feedback.
The key insight is that the developer's role shifts from author to director. You are still making all the important decisions (architecture, user experience, business logic), but the mechanical act of translating those decisions into syntax is increasingly handled by AI.
Key Tools Powering the Vibecoding Era
Several tools have made vibecoding practical for everyday development. Here are the most widely used as of 2026:
- Cursor. A VS Code fork built around AI-first editing. Its Composer feature lets you describe multi-file changes in natural language and apply them across your project. Cursor is often credited as the tool that made vibecoding accessible to professional developers.
- Claude Code. Anthropic's CLI-based coding agent. It operates directly in your terminal, reads your codebase, runs commands, and makes changes autonomously. Particularly strong for complex refactoring and multi-step tasks.
- GitHub Copilot. The original AI coding assistant, now extended with chat and agent capabilities. Tightly integrated into VS Code and the broader GitHub ecosystem.
- Windsurf. An AI-native IDE (formerly Codeium) that combines code generation with deep codebase understanding. Its Cascade feature handles multi-file edits with awareness of your project's full context.
- Replit Agent. A browser-based development environment with an AI agent that can build, deploy, and iterate on full applications from a text description. Especially popular with non-technical builders and for rapid prototyping.
The landscape is moving quickly, with new tools and capabilities emerging regularly. What they all share is the same core pattern: you describe intent, the AI generates code, and you iterate through conversation.
Who Is Vibecoding For?
One of the most interesting aspects of vibecoding is how it serves very different audiences:
Professional Developers
For experienced engineers, vibecoding is a productivity multiplier. Tasks that once took hours of boilerplate (setting up API routes, writing database migrations, building form validation) can be completed in minutes. This frees up time for the work that matters most: system design, performance optimisation, and solving novel problems.
Companies that embrace this approach are seeing significant gains. Some organisations are reporting that the majority of their code is now authored by LLMs, with engineers focusing on architecture, code review, and quality assurance. You can browse AI-friendly employers to see which companies are leading this shift.
Founders and Startup Teams
For founders building MVPs, vibecoding has compressed timelines by orders of magnitude. What once required hiring a development team or spending months learning to code can now be prototyped in days. This has been especially powerful for non-technical founders who can describe their product vision and iterate on a working prototype directly.
Non-Technical Builders
Perhaps the most radical change is that vibecoding has opened software creation to people who have never written a line of code. Designers, product managers, marketers, and domain experts can now build internal tools, automations, and even customer-facing applications by describing what they need. Tools like Replit Agent have made this particularly accessible.
Common Concerns (and Honest Answers)
Vibecoding is not without its critics, and many of the concerns are worth taking seriously.
Code Quality
The concern: AI-generated code may contain bugs, security vulnerabilities, or architectural problems that are not immediately obvious.
The reality: This is a legitimate risk, especially for developers who accept all changes without review. The most effective practitioners treat AI as a junior developer: capable and fast, but requiring oversight. Code review, automated testing, and established engineering practices remain essential. The quality of AI-generated code has improved significantly since 2025, but human judgement is still the critical safeguard.
Job Displacement
The concern: If AI can write code, will developers lose their jobs?
The reality: So far, the evidence points more towards role evolution than replacement. Developers who adopt AI tools tend to become more productive, not redundant. The demand for software continues to grow, and AI is expanding what is possible rather than simply automating existing work. That said, the skills that matter are shifting. System thinking, clear communication, and the ability to evaluate AI output are becoming more valuable than raw typing speed. For more on how this affects career decisions, read our latest thinking on the blog.
Vendor Lock-in
The concern: Relying on AI tools creates dependency on specific vendors and their pricing models.
The reality: The code itself is standard. It is not locked to any particular AI tool. You can switch between Cursor, Claude Code, Copilot, or any other tool and continue working on the same codebase. The competitive landscape is also keeping pricing in check, with multiple strong alternatives available across every category.
The Future of Vibecoding
Looking ahead, several trends are shaping where vibecoding is going:
- Agentic workflows. AI tools are becoming more autonomous, capable of running tests, deploying code, and iterating on feedback loops without human intervention at every step. The line between "assistant" and "agent" continues to blur.
- Organisational adoption. Early adopters treated AI coding tools as individual productivity hacks. Increasingly, companies are building their entire engineering culture around AI-assisted development, with dedicated tooling budgets, updated hiring practices, and new definitions of what "senior engineer" means. See why AI-friendly employers matter for developers navigating this shift.
- Specialised models. General-purpose LLMs are being complemented by models fine-tuned for specific languages, frameworks, and domains. This is improving quality and reducing the need for correction.
- Broader access. As tools become more capable and affordable, the barrier to building software continues to drop. This is expanding the total number of people who can create technology, beyond just making existing developers faster.
Getting Started
If you are new to vibecoding, the best way to learn is to try it. Pick up any of the tools mentioned above (Cursor and Replit are great starting points) and start with a small project. Describe what you want, see what the AI produces, and iterate from there.
For experienced developers, the key is to lean into the workflow rather than fighting it. Let the AI handle the boilerplate while you focus on the decisions that require human judgement. You might be surprised how quickly it becomes second nature.
And if you are looking for a workplace that supports this way of working, take a look at our directory of AI-friendly employers. These are companies that actively build their engineering culture around AI tools, rather than merely tolerating them.