Stop Coding, Start Vibe Coding

In early 2025, Andrej Karpathy, former Research Director at OpenAI, posted a tweet that defined a new concept: Vibe Coding.

He wrote: "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."

This tweet sparked discussions across the entire tech industry. Some saw it as the future of programming, others dismissed it as hype. But regardless of your opinion, one fact remains: 25% of startups in Y Combinator's Winter 2025 batch used AI to generate over 95% of their code.

What Exactly is Vibe Coding

Let me be direct. Vibe Coding is not about not writing code. It's about changing your relationship with code.

The traditional programming workflow looks like this: you learn the syntax of a programming language, understand data structures and algorithms, then write out the logic line by line. When you hit bugs, you debug them yourself. When you encounter something unfamiliar, you search Stack Overflow. The entire process requires a massive accumulation of technical knowledge.

Vibe Coding works completely differently. You describe what you want in natural language, and AI turns it into code. You don't need to know how React's useState works. You just say "I need a counter that increments when you click a button." You don't need to know SQL syntax. You just say "Help me store user registration information."

Karpathy described his workflow in his original tweet: "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."

This sounds somewhat crazy. A former OpenAI research director admitting he "accepts all changes without looking at the code." But this is the core philosophy of Vibe Coding: delegate code details to AI, while humans focus on product vision.

What GPT-5.2 Changes

In early December 2025, OpenAI released GPT-5.2. This model has significant implications for Vibe Coding.

Let me share some specific numbers. GPT-5.2 achieved a 55.6% score on the SWE-Bench Pro programming benchmark, covering four programming languages. Compared to the previous generation, hallucinations were reduced by 30%. This means AI-generated code is more accurate, with fewer instances of "confidently making things up."

More importantly, GPT-5.2's context understanding capabilities have improved qualitatively. It can remember longer conversation histories and understand more complex project structures. You can feed it your entire codebase and ask it to make cross-file modifications without forgetting previously discussed content.

Tool-calling capabilities are another key improvement. The core of Vibe Coding is making AI an autonomous agent, not just answering questions but proactively executing tasks. GPT-5.2 performs much better in this regard. It can more accurately invoke various tools and APIs, completing more complex workflows.

Real Cases: Success Stories from Non-Programmers

Let me share some real examples.

A Reddit user shared their experience: zero programming background, built a complete full-stack to-do app in less than 5 hours using Cursor AI. The app uses React for frontend, Firebase for backend, has routing navigation, animations, and complete styling. Throughout the process, they didn't write a single line of code. Everything was done through conversation with AI.

An even more impressive case is Pieter Levels. He used the AI editor Cursor Grok 3 to build a multiplayer game in 17 days, creating a Python websockets server, with annual revenue reaching $1 million. From zero to launch to profitability, the entire project took less than three weeks.

Y Combinator CEO Garry Tan said in an interview that Vibe Coding has become an "unavoidable progression" and the "dominant way to code." He warned that companies not adopting AI-assisted development will fall behind. This isn't just his opinion. Data from YC's Winter 2025 batch shows that 25% of startups used AI to generate over 95% of their code.

There's also an interesting case of an 8-year-old child. A viral tweet documented them making a simple game in Cursor AI, with the entire process being just describing what they wanted, and AI implemented it. This shows that Vibe Coding has truly lowered the barrier to programming to the level of "being able to speak."

The Vibe Coding Tech Stack

If you want to start Vibe Coding, you need to know the most popular tool combinations.

Brain Layer uses ChatGPT or Claude. This is your strategic planning tool. Before writing code, discuss product positioning, feature planning, and technology choices with AI. Let it help you brainstorm product names, design logo concepts, and prioritize MVP features. Many people skip this step and jump straight into coding, only to build something in the wrong direction.

Builder Layer uses Cursor. This is currently the most powerful Vibe Coding IDE. Based on VS Code, it integrates AI capabilities that understand your entire project context. You can describe features through voice or text, and Cursor will generate code, modify files, and run tests. Cursor's valuation reached $29 billion in 2025, with ARR exceeding $1 billion. This shows how massive the market demand is for such tools.

Backend Layer uses Supabase. An open-source Firebase alternative providing database, authentication, storage, and real-time subscriptions. Its advantage is a clean API design that AI can easily understand and operate. Tell AI "I need user registration and login functionality," and it configures Supabase Auth. Tell AI "I need to store user articles," and it designs the database schema.

Deploy Layer uses Vercel. One-click deployment, automatic HTTPS, global CDN. You don't even need to understand what these concepts mean. Just push code to GitHub, and Vercel automatically deploys your site.

Payment Layer uses Stripe. If your product needs to charge money, Stripe is the easiest payment solution to integrate. AI can generate Stripe integration code, handling subscriptions, one-time payments, refunds, and various scenarios.

Practical Considerations

Vibe Coding is not a silver bullet. Before you start, there are several things you need to understand.

First is code quality. AI-generated code may contain security vulnerabilities, performance issues, or practices that don't follow best standards. Because AI learns from public code repositories, and those repositories contain many insecure examples. For weekend projects and quick prototypes, this isn't a big problem. But if you're building a product handling sensitive user data, you need professional developers to review the code.

Second is technical debt. The speed advantage of Vibe Coding comes from "accepting without looking." But this also means you might accumulate large amounts of code you don't understand. When the project grows, when you need to make complex modifications, when AI can't understand your requirements, this technical debt becomes exposed.

Third is AI limitations. AI doesn't truly understand your business logic. It's doing pattern matching. For common problems, it excels. But for your unique business requirements, you need to describe them very clearly, otherwise AI might give you "correct but useless" answers.

Finally is cost. While Vibe Coding reduces labor costs, AI model calls have fees. Frequent use of top-tier models like Claude Opus or GPT-5.2 might cost tens to hundreds of dollars monthly. This is also an expense for individual developers.

The Shift in Competitive Advantage

This is what I consider the most important insight from Vibe Coding.

In the past, if you wanted to start a software company, technical ability was the biggest barrier. You either had to write code yourself or have money to hire developers. Many great product ideas died simply because the founder couldn't code.

Now, the technical barrier has essentially disappeared. Anyone who can clearly describe what they want can build a product. This means the focus of competition has shifted from "can you build it" to "will anyone want what you built."

In other words, product insight, market understanding, and user empathy have become more important than technical ability. Previously, programmers could coast on technical skills. Now they must think about "is this feature actually valuable." Previously, product managers had to defer to programmers. Now they can turn their ideas into reality themselves.

This is a reshuffling of the entire tech industry. Those who truly understand users and truly solve problems will gain unprecedented advantages. Those who are just "able to write code," if they can't provide higher-level value, may face difficulties.

How to Start Your First Project

If this article makes you want to try Vibe Coding, here are some specific suggestions.

Step one, choose a simple project. Don't try to build a complex SaaS right away. Build a small personal tool, a static website, a simple calculator. The goal is to familiarize yourself with the tools and workflow, not to create something earth-shattering immediately.

Step two, sign up for Cursor Pro subscription. This is currently the most mainstream Vibe Coding tool, $20 per month, including sufficient AI call quota. After installation, open your project folder and start conversing with AI.

Step three, learn to ask questions. The core skill of Vibe Coding isn't writing code. It's expressing requirements. You need to learn how to turn vague ideas into clear instructions. Don't say "make a nice page." Say "make a product introduction page with three feature showcase blocks, a CTA button, using a dark theme."

Step four, embrace iteration. The first output is rarely perfect. You need to learn to continue adjusting and conversing based on AI's output. "This button is too small." "Change the color to blue." "Move the title to the left." Vibe Coding is a conversational process, not a one-time command.

Step five, maintain backups. AI sometimes makes mistakes, sometimes deletes files it shouldn't. Before letting AI make major changes, at least save a version with Git commit. That way even if something goes wrong, you can roll back.

The barrier to programming has dropped to the level of "language expression ability." The competitive barrier now isn't whether you can code, but your understanding of the product. If you have an idea, now is the best time to bring it to life.