Summarize this article with:
Andrej Karpathy posted 185 words on X in February 2025 and accidentally named an entire movement. Vibe coding is the practice of building software by describing what you want in plain language and letting AI generate the code, without reading or fully understanding what it produces.
It sounds reckless. And sometimes it is. But tools like Cursor, GitHub Copilot, and Replit have made it practical enough that 63% of vibe coding users today aren’t even developers.
This guide covers how vibe coding works, which AI-powered tools people actually use, where the approach breaks down, and how it’s changing who gets to build software. Real stats, real limitations, no hype.
What Is Vibe Coding
Vibe coding is a software development approach where you describe what you want in plain language and let AI generate the code. You don’t write syntax. You don’t memorize functions. You talk, and the machine builds.
Andrej Karpathy, co-founder of OpenAI and former AI director at Tesla, coined the term on February 2, 2025, in a post on X that racked up over 4.5 million views. His exact description: you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
That’s the part that separates vibe coding from regular AI-assisted programming. You accept the output without reading every line. You skip the diffs. When errors appear, you paste them back into the chat and move on.
Simon Willison, a well-known developer, drew a clear line here. If you review and understand every line the AI wrote, that’s just using an LLM as a typing assistant. Vibe coding means you let go of full comprehension.
Merriam-Webster added the term as “slang & trending” by March 2025. Collins English Dictionary named it their Word of the Year for 2025. So yeah, it caught on.
The practice runs on large language models (GPT-4, Claude, and others) plugged into code editors and development environments. Cursor, GitHub Copilot, Replit Agent, and Bolt.new are among the platforms people use most. The developer types a prompt, or even speaks it through voice transcription tools like SuperWhisper, and the AI handles implementation.
Karpathy was blunt about the limits. He admitted AI sometimes can’t fix certain bugs, forcing him to “work around it or ask for random changes until it goes away.” He called it best for throwaway weekend projects. But what started as an amusing experiment quickly became something much bigger.
How Vibe Coding Actually Works
The workflow is simple on the surface. You open an AI-powered editor, describe what you want in everyday language, and the model generates working code. But there’s a rhythm to it that takes a few sessions to feel out.
The Prompt-Generate-Test Loop
Everything starts with a prompt. Not code. Not pseudocode. Just a sentence like “build me a dashboard that shows daily sales with a line chart and a dropdown filter for regions.”
The AI returns full files, sometimes entire project scaffolds. You run it. If something breaks, you copy the error message straight into the chat. No commentary needed, the model usually figures it out.
This loop (prompt, generate, run, fix) repeats until the thing works. A GitHub study found developers using AI tools completed tasks 55% faster than those working without them. That speed comes from eliminating the manual translation step between idea and implementation.
The Role of the Developer in Vibe Coding
You’re not writing code. But you’re not doing nothing, either.
Direction: You decide what gets built, in what order, and what “done” looks like. The AI doesn’t know your product requirements.
Judgment: You test the output. Click through the interface. Check if the button actually does what you asked. According to Stack Overflow’s 2025 survey, only 33% of developers trust AI output accuracy, down from 43% the year before. People are using these tools more while trusting them less.
Intervention: Sometimes the AI gets stuck in a loop. It keeps generating the same broken fix. That’s when you step in, manually adjust something, or restructure the prompt entirely.
The skill here isn’t typing. It’s knowing when to keep prompting and when to take over. Took me a while to realize that the quality of your prompt engineering matters more than anything else in this process.
Tools and Platforms Used for Vibe Coding
The tool you pick shapes everything about how vibe coding feels. Some are built from the ground up for this workflow. Others bolt AI features onto existing editors. The difference matters more than most people expect.
| Tool | Type | Best For |
|---|---|---|
| Cursor | AI-native IDE | Professional developers and full-stack projects |
| GitHub Copilot | Editor extension | Teams already working in VS Code and GitHub |
| Replit Agent | Browser-based IDE | Non-developers and quick prototypes |
| Bolt.new | Prompt-to-app builder | Zero-setup apps with instant deployment |
| Lovable | Agentic app builder | Non-technical founders building MVPs |
| v0 by Vercel | UI generator | Front-end scaffolding and component generation |
Cursor dominates the conversation right now. Built on Visual Studio Code by four MIT graduates at Anysphere, it went from $1M to $100M in annual recurring revenue within 12 months. By March 2026, TechCrunch reported it had crossed $2 billion in annualized revenue with a $29.3 billion valuation. No developer tool has ever grown this fast.
GitHub Copilot has a wider install base, with over 20 million users as of mid-2025 and 90% adoption among Fortune 100 companies. It generates roughly 46% of all code written by active users. Java developers see that number climb to 61%.
Replit pivoted hard from browser-based collaborative IDE to full agentic coding platform. After launching Replit Agent in late 2024, it closed a $250 million round in 2025. Bolt.new from StackBlitz hit $1 million ARR within a week of launch and reached $40 million ARR by early 2025.
Integrated Editors vs. Chat-Based Tools
Integrated tools like Cursor and GitHub Copilot maintain file context automatically. They see your entire codebase, understand project structure, and make changes across multiple files at once. This is where complex projects live.
Chat-based tools (ChatGPT, Claude in a browser window) require more manual copy-paste. You feed code in, get code back, and move it yourself. More flexible for quick one-off tasks, but tricky when the project grows past a few hundred lines.
Cursor holds about 18% of the AI coding assistant market, second only to GitHub Copilot at 42%. The rest splits between Amazon Q Developer, Windsurf (formerly Codeium), and smaller players.
What Vibe Coding Is Good For
Prototyping. That’s where this approach absolutely shines.
Over 30% of early-stage startups globally reported using vibe coding to build MVPs in under a week during 2024, according to market research from Congruence. The Klarna CEO, who has never formally coded, mentioned receiving a working prototype in 20 minutes for concepts that previously took his engineering team weeks.
Rapid Prototyping and MVPs
Y Combinator reported that 25% of its Winter 2025 batch had codebases that were 95% AI-generated. That’s not a fringe experiment. That’s a quarter of one of the most competitive startup accelerators on the planet.
The pattern is clear: solo founders describe features in plain language, AI spits out working apps, and investors see functional demos instead of pitch decks. Rapid application development went from weeks to hours for the right kind of project.
Internal Tools and One-Off Scripts
Around 25% of small and medium businesses in North America used vibe coding for internal operations apps in 2024. Dashboards, admin panels, data import scripts, the kind of stuff that’s too boring for an engineering team to prioritize but too specific for off-the-shelf tools.
These projects work because nobody cares about long-term code quality. If the internal reporting tool breaks in six months, someone just prompts a new one.
Non-Developers Building Real Software
This is where things get interesting. 63% of vibe coding users are non-developers, according to a 2025 industry survey. They’re generating user interfaces (44%), full-stack apps (20%), and personal software tools (11%).
Designers prototyping interactive components. Product managers building proof-of-concept demos. People who used to spec features in documents are now shipping them. Kevin Roose at The New York Times experimented with vibe coding and built several small apps he called “software for one,” custom tools tailored to his specific needs.
Where Vibe Coding Breaks Down
The problems show up the moment you try to go beyond a prototype. And they’re not small problems.
Technical Debt and Code Quality
GitClear analyzed 211 million lines of code changes from 2020 to 2024. Code refactoring dropped from 25% of changed lines in 2021 to under 10% by 2024. Code duplication increased roughly four times. Copy-pasted code exceeded moved code for the first time in two decades.
Market Clarity data from 2025 shows AI-generated code has 41% higher code churn rates, meaning it gets rewritten shortly after being merged. Only 30% of AI suggestions get accepted by developers who actually review them.
Security Vulnerabilities
In May 2025, Lovable, one of the most popular vibe coding platforms, was found to have security issues in 170 out of 1,645 web applications it generated. Personal information was accessible to anyone who looked.
A December 2025 CodeRabbit analysis of 470 open-source GitHub pull requests found that AI co-authored code contained 2.74x more security vulnerabilities than human-written code. Logic errors, incorrect dependencies, flawed control flow, misconfigurations (75% more common). The list goes on.
Around 29% of Python code generated by GitHub Copilot contains potential security weaknesses. For throwaway prototypes, that’s acceptable. For anything touching user data or payments, it’s a deal-breaker without serious human review.
Complexity Ceilings
AI handles simple tasks well. Basic algorithms, CRUD operations, standard UI components. But projects involving multiple files, poorly documented libraries, or safety-critical systems? That’s where things fall apart.
In July 2025, the SaaStr founder documented how Replit’s AI agent deleted a database despite explicit instructions not to make changes. By September 2025, Fast Company reported on what senior engineers called a “vibe coding hangover,” describing development hell when working with AI-generated code on complex applications.
The “it works but I don’t know why” problem isn’t theoretical. Maintainability drops fast when nobody on the team actually understands what the code does.
Vibe Coding vs. Traditional Software Development
These aren’t competing approaches in the way most people frame them. They’re different tools for different situations, and the tradeoffs are real.
| Dimension | Vibe Coding | Traditional Development |
|---|---|---|
| Speed (initial build) | Hours to days | Weeks to months |
| Code understanding | Partial or minimal understanding | Full comprehension |
| Maintenance cost | Higher (code may be unclear or inconsistent) | Lower (documented and reviewed) |
| Skill required | Prompt clarity and problem description | Programming knowledge and fluency |
| Best for | Prototypes, MVPs, and internal tools | Production systems and regulated software |
Speed vs. Long-Term Maintainability
The initial speed advantage is massive. GitHub’s research shows 55% faster task completion with AI tools. Pull request times dropped from 9.6 days to 2.4 days in enterprise settings.
But here’s the catch. That speed creates code nobody fully reviewed. GitClear’s data shows short-term code churn nearly doubled between 2020 and 2024. You build fast, then spend more time fixing what you built. For a weekend prototype, fine. For something you need to maintain over years, that’s a problem the traditional development process simply handles better.
Where Teams Land in Practice
Most professional teams are landing somewhere in between. They use AI for boilerplate, test generation, and first drafts. Then humans review, refactor, and own the architecture.
DORA’s 2025 report found that AI acts as a “mirror and multiplier.” In organizations with solid foundations, strong source control, clear code review processes, good internal platforms, AI makes things faster. In fragmented teams with inconsistent processes and existing technical debt, it makes things worse.
Andrew Ng pushed back on the term itself, saying it misleads people into thinking professional engineers just “go with the vibes.” Karpathy eventually agreed. In February 2026, he proposed a new term, agentic engineering, to describe the more disciplined version of AI-assisted development that professionals actually practice.
The difference between vibe coding and traditional coding isn’t really about which is better. It’s about matching the approach to the stakes of the project.
Who Is Using Vibe Coding Right Now
The user base looks nothing like what you’d expect from a programming trend. A 2025 industry survey found that 63% of vibe coding users are non-developers. The people writing code without writing code aren’t just tinkering. They’re shipping products.
Solo Founders and Startups
Menlo Ventures reported that 44% of non-technical founders now build their initial prototypes using AI coding assistants rather than hiring developers. The startup accelerator Y Combinator confirmed that 25% of its Winter 2025 batch ran codebases that were 95% AI-generated.
Sabrine Matos, a growth marketer with no engineering background, built Plinq (a women’s safety app) entirely with Lovable. The app reached 10,000 users in three months and $456,000 in annual recurring revenue.
Designers and Product Managers
Product teams at AppDirect used Lovable to build 11 different internal projects without engineering involvement. The result: over $120,000 in software cost savings and a website rebuild that took one month instead of six.
Designers use tools like v0 by Vercel to generate front-end components directly from natural language descriptions. No more waiting three sprints for engineering to pick up a wireframe.
Professional Developers (Yes, Them Too)
Stack Overflow’s 2025 survey shows 84% of developers now use or plan to use AI tools in their workflow. The Wall Street Journal reported in July 2025 that vibe coding had crossed into commercial use among professional software engineers.
Linus Torvalds, creator of Linux, used Google’s AI tools to vibe code a component of his AudioNoise project in early 2026. Even at the top of the profession, the pull toward prompt-driven development is real.
| User Type | Primary Use Case | Common Tools |
|---|---|---|
| Solo founders | MVPs and investor demos | Lovable, Bolt.new, Replit |
| Designers | Interactive prototypes and UI components | v0, Cursor |
| Product managers | Proof-of-concept demos | Bolt.new, ChatGPT |
| Senior engineers | Boilerplate generation, tests, and refactoring | Cursor, GitHub Copilot, Claude |
The Skill Behind Good Vibe Coding
Vibe coding looks effortless in demos. In practice, it’s a skill that takes time to develop, and the gap between a good prompt and a bad one is the difference between a working app and a frustrating afternoon.
Writing Clear, Scoped Prompts
Telling an AI “make me a website” gives you garbage. Telling it “build a dashboard with a sidebar nav, a line chart showing monthly revenue using Recharts, and a dropdown filter for date range” gives you something usable.
Specificity wins. The best vibe coding prompts include architecture preferences, design patterns, edge cases, and framework choices. Karpathy himself used Cursor Composer paired with Anthropic’s Claude models, feeding detailed spoken prompts through voice transcription.
Breaking Projects Into Small Pieces
A Harvard study of 285,000 firms found that AI tools produce measurable results on scoped, well-defined tasks but struggle with complex, multi-file projects. The same principle applies to prompting.
What works: one feature per prompt, test it, move on. What doesn’t: asking the AI to build your entire mobile application in a single conversation. That’s how you end up with tangled dependencies and mysterious bugs that neither you nor the AI can trace.
Knowing When the AI Is Wrong
Stack Overflow data from 2025 shows 66% of developers cite “AI solutions that are almost right, but not quite” as their top frustration. Only 3% highly trust AI output without reviewing it first.
You don’t need to read every line. But you do need to know what “broken” looks like. Can you spot when a React component re-renders infinitely? Can you tell when an API integration is using the wrong authentication method? Some baseline knowledge of how back-end systems and UI/UX patterns work still matters, even if you never write a line yourself.
How Vibe Coding Changes Who Can Build Software
The barrier to building software used to be years of study, then years of practice. Now it’s access to a $20/month subscription and the ability to describe what you want clearly. That shift has consequences, and not all of them are comfortable.
The New Builder Class
Menlo Ventures calls it “TAM expansion in coding.” The total addressable market for development tools is growing because the definition of “developer” is expanding. 47% of AI users now apply AI to coding for work or school, and 41% use it for personal projects.
Gartner forecasts that by the end of 2026, 40% of enterprise applications will feature task-specific AI agents, up from under 5% in 2025. The low-code market is projected to grow from $37 billion in 2025 to $264 billion by 2032.
People who would never have considered building web apps or custom applications three years ago are now doing it weekly.
The Impact on Junior Developers
This is the part nobody wants to talk about, but the numbers are stark.
A Stanford Digital Economy Study found employment for software developers aged 22 to 25 declined nearly 20% from its late 2022 peak. Entry-level tech job postings dropped 67% between 2023 and 2024, according to Stanford Digital Economy Lab analysis. CS graduates face a 6.1% unemployment rate. Computer engineering grads sit at 7.5%.
A Harvard study of 285,000 U.S. firms found that when companies adopt generative AI, junior employment drops 9 to 10 percent within six quarters. Senior employment barely changes.
Salesforce announced it would hire no new engineers in 2025. Klarna froze developer hiring entirely in late 2023 (they eventually reversed this). Indian IT services companies reduced entry-level roles by 20 to 25%, according to EY. LinkedIn, Indeed, and Eures tracked a 35% decline in junior tech positions across major EU countries during 2024.
What This Means for Hiring and Team Structure
The role is shifting, not disappearing entirely. Senior developers remain critical. AI doesn’t make architectural decisions, handle compliance, or interpret vague client requirements. Microsoft CEO Satya Nadella noted that AI writes 20 to 30% of Microsoft’s internal code but emphasized that juniors still need strong fundamentals.
Matt Garman, CEO of AWS, called replacing juniors with AI “one of the dumbest ideas” a company can have. His reasoning: junior employees are often the most proficient AI users, having grown up with these tools. Stack Overflow’s 2025 survey backs this up, with 55.5% of early-career developers using AI tools daily, a higher rate than their senior counterparts.
The path forward probably looks like fewer pure coding roles and more positions that blend prompting skills with systems thinking, software prototyping, and product judgment. Traditional development roles will evolve. Some will expand. The ones focused purely on writing boilerplate code? Those are already shrinking.
FAQ on Vibe Coding
Who coined the term vibe coding?
Andrej Karpathy, co-founder of OpenAI and former AI director at Tesla, coined the term in a post on X on February 2, 2025. The post was viewed over 4.5 million times and sparked widespread adoption of the concept.
Is vibe coding real programming?
It produces real, functional code. But the developer doesn’t write or always understand it. Simon Willison argues that if you review every line the AI generates, that’s AI-assisted coding, not vibe coding. The distinction matters.
What tools do people use for vibe coding?
The most popular platforms include Cursor, GitHub Copilot, Replit Agent, Bolt.new, Lovable, and v0 by Vercel. Each handles the prompt-to-code workflow differently depending on project complexity and user experience level.
Can non-developers use vibe coding to build apps?
Yes. A 2025 survey found 63% of vibe coding users have no development background. They’re building user interfaces, full-stack applications, and personal tools using nothing but natural language prompts and AI code generation.
Is vibe coded software safe to deploy?
Not without review. A CodeRabbit analysis found AI co-authored code contains 2.74x more security vulnerabilities than human-written code. For prototypes, the risk is low. For production systems handling user data, manual review is necessary.
How is vibe coding different from using GitHub Copilot?
Copilot suggests code completions that developers review line by line. Vibe coding means accepting AI output without fully reading it. The key difference is comprehension. Copilot assists. Vibe coding delegates.
What are the biggest limitations of vibe coding?
Complex multi-file projects, security-sensitive applications, and long-term maintainability are the main weak spots. GitClear data shows code duplication increased 4x and refactoring dropped sharply as AI-generated code grew between 2021 and 2024.
Does vibe coding replace traditional software development?
No. It handles prototypes, MVPs, and internal tools well. But production-grade systems still need human oversight, proper testing, architecture decisions, and code review. Most professional teams blend both approaches.
Will vibe coding eliminate developer jobs?
Junior roles are shrinking. Stanford data shows employment for developers aged 22 to 25 dropped nearly 20% from 2022 peaks. But senior positions remain stable. The role is evolving toward oversight and systems thinking rather than disappearing entirely.
How do I start vibe coding?
Pick a tool like Cursor or Replit. Describe a small project in plain language. Run the output. Fix errors by pasting them back into the chat. Start with throwaway projects before attempting anything production-level.
Conclusion
Understanding what is vibe coding comes down to one shift: the person describing the software no longer needs to be the person writing it. Large language models like GPT-4 and Claude handle the translation from idea to functioning code.
That’s powerful for prototyping, internal tools, and rapid app development. It falls short for production systems that demand reliability, scalability, and long-term maintenance.
The tools are getting better fast. Cursor crossed $2 billion in revenue. GitHub Copilot hit 20 million users. But AI-generated code still needs human judgment for anything with real stakes.
Vibe coding didn’t replace the software development process. It added a new entry point. Whether you’re a founder, a designer, or a senior engineer, the question isn’t whether to use it. It’s knowing when to stop prompting and start reviewing.
Many of his resources are available on various design marketplaces and for free on Codepen.
Over the years, he's worked with a range of clients and contributed to design publications like Design Your Way, Designmodo, WebDesignerDepot, WPDean, and Speckyboy, and Slider Revolution among others.
- 4 Scalable Hosting Providers for Growing Small Business Websites - April 9, 2026
- 7 Best Private Equity CRM Platforms for Middle-Market Deal Teams [2026 Comparison] - April 8, 2026
- Markdown Cheat Sheet - April 8, 2026
