Andrej Karpathy — the guy who ran AI at Tesla and co-founded OpenAI — posted a tweet in February 2025 that named something millions of people were already doing. He called it “vibe coding.”
“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.”
That tweet got 4 million views. Collins English Dictionary named “vibe coding” its Word of the Year for 2025. And the search volume tells the real story: 110,000 people search for it every month.
But here’s what most articles about vibe coding miss — they define it and stop. They don’t tell you how to actually do it. Which tools to pick. What works. Where it breaks.
That’s what this guide covers.
What Vibe Coding Actually Is
Vibe coding means describing what you want in plain English (or any language), and letting AI write the code. You don’t read the code. You don’t debug line by line. You just describe, test, and iterate.
Karpathy described his own workflow: he talks to Cursor using a voice dictation tool called SuperWhisper. He says things like “decrease the padding on the sidebar by half” because he’s too lazy to find the CSS himself. He hits “Accept All” without reading the diffs. When errors pop up, he copy-pastes them back with no comment — and the AI usually fixes it.
That’s vibe coding in its purest form. You’re steering the outcome, not writing the implementation.
And it’s not just hobbyists messing around. 25% of Y Combinator’s Winter 2025 batch had codebases that were 95%+ AI-generated — and YC’s president Garry Tan said it was their fastest-growing, most profitable batch ever. Jason Lemkin at SaaStr built 12+ vibe-coded apps that were used over 800,000 times. As a non-programmer.
The Numbers Behind the Hype
The adoption stats are hard to ignore:
- 92% of US developers use AI coding tools daily in 2026
- 41% of all code globally is now AI-generated (256 billion lines in 2024)
- GitHub Copilot has 20 million+ users and generates an average of 46% of code
- 87% of Fortune 500 companies adopted at least one AI coding platform
- Developers complete routine tasks 51% faster and API integrations 81% faster
- The AI coding tools market is worth $7.37 billion in 2025
But adoption doesn’t mean it’s all smooth. A CodeRabbit study from December 2025 found that AI-generated code has 1.7x more issues than human code, 2.74x more XSS vulnerabilities, and 1.88x more password handling bugs. The Stack Overflow blog built an app with Bolt in 10 minutes that “didn’t work at all” — and the code was “messy and nearly impossible to understand.”
So vibe coding is powerful. It’s also a skill you need to learn — not a magic wand.
The Best Vibe Coding Tools (and What They Actually Cost)
Cursor — The AI-Native IDE
Cursor is a VS Code fork that understands your entire codebase. It’s what Karpathy used when he coined the term.
- Free tier: 2,000 code completions + 50 slow premium requests/month
- Pro: $20/month (500 fast premium requests)
- Valuation: $29.3 billion (November 2025)
- ARR: $1.2 billion, up 1,100% year over year
Cursor is best when you already have a codebase and want AI help navigating, refactoring, or extending it. It’s the “graduate to this” tool. Our Cursor AI course walks through the full workflow from setup to shipping.
Lovable — Describe It, Build It
Lovable lets you describe an app in plain English and generates a working prototype. No IDE, no terminal, no Git knowledge needed.
- Free tier: Limited generations
- Starter: $20/month
- Valuation: $6.6 billion (December 2025, investors include Nvidia and Alphabet)
- Usage: 100,000 projects built per day
Lovable is where most non-programmers should start. Describe what you want, see it built live, refine with conversation. For getting the most out of it, the Lovable App Prompt Guide skill gives you templates that produce better first-attempt results.
Bolt.new — Fast Prototypes
Best for hackathons, demos, and “I need this working in 20 minutes” situations. You describe the app, Bolt generates it in a sandbox, and you can deploy directly to Netlify or Vercel.
The speed is real — but the code quality drops off fast once you need anything beyond a basic prototype. Our Bolt App Builder skill helps you write better prompts for more reliable output.
Replit — The Learning Playground
Andrew Ng partnered with Replit to create a “Vibe Coding 101” course. It’s the most beginner-friendly environment: everything runs in the browser, no setup needed.
Good for learning. Less suited for production apps.
Claude Code — The Power Tool
Anthropic’s CLI tool. Completes complex multi-file tasks “in a single pass that would normally take 45+ minutes.” This is for people who are comfortable with a terminal and want AI that can modify dozens of files at once.
Our Full-Stack App Architect skill pairs well with Claude Code for larger projects.
v0 by Vercel — React UI Generator
Laser-focused on React and Next.js. If your project is a web app and you want beautiful UI components, v0 generates them from text descriptions. Not a general-purpose tool — but excellent at what it does.
How to Actually Vibe Code (Step by Step)
Step 1: Start with a Clear Description
The most common mistake is being vague. “Build me a todo app” gives you a generic todo app. Instead:
“Build a task manager where users can create projects, add tasks with due dates and priority levels (high/medium/low), drag tasks between columns (To Do, In Progress, Done), and filter by priority. Use a clean, minimal design with a dark sidebar.”
More detail upfront = fewer revision cycles.
Step 2: Pick the Right Tool for Your Skill Level
| Your Experience | Best Starting Tool | Why |
|---|---|---|
| Never coded before | Lovable or Bolt.new | Pure conversation, no IDE needed |
| Basic HTML/CSS knowledge | Replit or v0 | Browser-based, gentle learning curve |
| Can use a terminal | Cursor or Claude Code | Full IDE power, codebase-aware |
| Full-stack developer | Cursor + Claude Code | Maximum productivity, complex projects |
Step 3: Iterate in Small Chunks
Don’t try to describe your entire app in one prompt. Break it into pieces:
- “Create the basic layout with navigation and authentication”
- “Add a dashboard with placeholder data”
- “Connect to a Supabase database and replace placeholders with real data”
- “Add a search bar that filters results in real-time”
Each step should be testable on its own. If something breaks, you know exactly which change caused it.
Step 4: Use Error Messages as Prompts
When your app throws an error, copy the full error message and paste it back to the AI with no additional context. This is literally what Karpathy does. The AI has seen millions of similar errors and usually knows the fix.
Step 5: Know When to Stop Vibing
Vibe coding has a ceiling. According to SaaStr’s analysis:
- Prototype stage: Near-zero maintenance, vibe away
- Early production (hundreds of users): A few hours per week of debugging
- Scaled production (thousands+ users): 30-60 minutes of daily maintenance minimum
If your project is gaining real users, you’ll eventually need someone who can read the code. That doesn’t mean vibe coding was wrong — it means it got you to a place where traditional engineering starts to matter.
What Vibe Coding Can’t Do (Yet)
Be honest about the limits:
Security is a real concern. AI agents have been observed removing input validation, relaxing database policies, and disabling authentication to “fix” runtime errors. If your app handles user data or payments, you need a human review.
Code quality degrades at scale. GitClear found a 4x growth in code clones (copy-pasted logic) in AI-generated codebases. The code looks clean but hides technical debt that surfaces later.
Consistency across sessions. AI doesn’t remember your coding style between conversations. You’ll get slightly different patterns each time, making the codebase harder to maintain.
Complex business logic. Multi-step workflows with edge cases, race conditions, or compliance requirements still need a human architect thinking through the design.
Our AI Code Review skill catches many of these issues — paste your vibe-coded output and get a structured review of bugs, security holes, and performance problems.
Who’s Actually Using This
The success stories are real and growing:
- YC W25: 25% of startups had 95%+ AI-generated code. Garry Tan said the batch grew 10% per week — founders didn’t need teams of 50-100 engineers.
- Walmart saved 4 million developer hours using AI coding tools
- Booking.com hit 65% adoption and saved 150,000 hours in year one
- Lovable sees 100,000 projects built on its platform every day
- Matt Welsh, former Harvard CS professor: “The conventional idea of ‘writing a program’ is headed for extinction”
But Andrew Ng — who created a whole course on vibe coding — called the name “misleading.” His point: it’s still a skill. You need to learn how to prompt well, how to architect your project, and how to troubleshoot when things go wrong.
If you want a structured path through all of this, our Vibe Coding with AI course covers everything from first prompt to deployed app — including the troubleshooting parts that blog posts skip.
Getting Started Today
Here’s the fastest path from zero to a working app:
- Open Lovable or Bolt.new — no signup drama
- Describe something small — a personal dashboard, a recipe tracker, a habit journal
- Iterate 3-5 times — refine the output with follow-up prompts
- Deploy it — both tools have one-click deploy
- Show someone — getting feedback on a real thing beats planning forever
The whole process takes about 30 minutes. And if it’s broken or ugly or missing features — that’s fine. You just describe what to fix and try again.
The Mobile App Vibe Coder skill and Chrome Extension Builder skill are good starting points for specific project types. And if you want to go deeper into the prompting side, the Prompt Engineering course covers the techniques that make AI tools produce better results across the board.
Vibe coding isn’t replacing programmers. But it’s making the gap between “I have an idea” and “I have a working app” shorter than it’s ever been. Whether you’re a founder testing a concept, a designer building a prototype, or someone who just wants to solve a personal problem — the tools are ready. The only question is what you’ll build first.