Is it an art or a science?

What if I told you the app you've been postponing could be live this week?

That's what vibe coding made possible. And even the best developers in the world are doing it now.

Andrej Karpathy (the person who coined "vibe coding") wrote on X that he's mostly programming in English now. Karpathy co-founded OpenAI and built Tesla's Autopilot, so he does know something about this… And doesn't write code anymore.

Neither does Anthropic. Claude Code now writes about 90% of Anthropic's own code. The tool is largely building itself, along with Claude Cowork and other portfolio products.

What is vibe coding?

Vibe coding is prompting and getting code as a result. For example “build a landing page with a beautiful hero section”, and you get it in a few minutes. You literally describe what you want in plain English (either written or spoken). The AI builds it.

The term became mainstream in February 2025 and took off fast. By March 2025 it was already in Merriam-Webster as a trending term. By the end of that year, 25% of YCombinator's Winter batch had codebases that were 95% AI-generated.

It went from meme to industry standard in under twelve months.

Why should you know about this?

For most of history, building software required a team. A developer, a designer, maybe a few contractors. You didn't just need an idea, you needed a budget and a timeline around it.

The old model was: have idea, find team (developer, designer…), wait, iterate, pay. Vibe coding cuts all of that. You're in direct conversation with the thing you're building. The feedback loop shrinks from weeks to minutes. Less time to market, less time to money.

How to start: tools and LLMs

These are the tools worth knowing. They are all solid, just spend 10 mins comparing them and choose the one you feel the most confortable with. You can reach the same result with any.

Claude Code: Anthropic's terminal-based coding agent. Writes, edits, and debugs entire codebases through conversation. Best for complex, production-grade projects. The closest thing to having a senior engineer available at all times. Mid learning curve. Kinda hard at the beginning if you’ve never coded before.

Cursor: A code editor with AI built deep into the workflow. Great if you want to stay inside a familiar development environment while letting AI handle most of the heavy lifting. Recommended for engineers.

Lovable: Goes from idea to deployed web app without opening a code editor. Best for non-technical founders who need something live fast.

Replit: Similar to Lovable but stronger for quick prototypes and experimentation. No local setup needed.

Now the LLMs. As of this writing, there are 3 clear winners: Opus 4.6, GPT 5.4 and Gemini 3.1. This will undoubtedly change in the following weeks. Here’s a snapshot from Artificial Analysis latest benchmarks:

Rules of thumb we use:

  • Gemini 3.1 best for beautiful designs

  • Opus 4.6 and GPT 5.4 for heavy, multi file work

  • Sonnet 4.6 for daily use. Similar to Opus but 40% cheaper.

But be careful…

Vibe coding builds fast. That's the point. But fast has a blind spot: security.

Veracode found on their 2025 LLM report that 45% of AI-generated code contains security flaws. Not minor bugs, authentication issues, exposed secrets, misconfigured data access. The stuff that gets companies in serious trouble.

The most documented case involves Lovable. A tech entrepreneur found 16 vulnerabilities in a single Lovable-hosted app with over 100,000 views, exposing the data of more than 18,000 users (Source: The Register).

How to not get burned

You can use Claude Code to audit the same app you just built. Claude Code has a /security-review command that scans your codebase for common vulnerabilities and suggests fixes. It takes about five minutes and catches most of the obvious stuff.

Three things to do before you ship:

  1. Run /security-review in Claude Code and fix everything it flags.

  2. Ask explicitly: "are there any hardcoded API keys or exposed secrets in this codebase?"

  3. Ask: "who has access to what data in this app, and is any of it publicly accessible?"

Treat the security review as a helpful first pass, not a final audit. If your app handles payments or sensitive user data, get a developer to review access policies before you go live. One hour of their time is cheaper than a breach or lawsuit.

What are people building?

The best way to understand what vibe coding unlocks is to look at what's getting built. A few real examples from the last few months:

  • Trading bots that monitor signals, execute orders, and send alerts, built in a weekend, no quant background required (check this)

  • Portfolio trackers that pull live data and visualize it in a custom dashboard

  • Internal ops tools that would have cost $10k to outsource six months ago

  • SaaS MVPs that got their first paying customers before hiring a single developer

  • Browser agents that automate research, scraping, and reporting workflows

The common thread: none of these required a development team. They required clarity on what to build and two hours of focused prompting.

How you can take action

Pick one thing you've been sitting on. A tool, a chatbot, an idea you've been calling "too technical." Spend 5 minutes minutes describing it to Claude Code, Lovable, or Cursor. Don't aim for perfect. Aim for something that exists.

New to Claude Code? This cheat sheet covers everything: Claude Code Commands List

You’ll be surprised by how far you can go by just typing on your keyboard and having a clear idea. 5 minutes will turn into hours because of how much fun you’ll have building it.

If you have a vibe coding win, a tool worth sharing, or a project you shipped this week hit reply. I'll feature it in the next issue.

We send these weekly. If this one was useful, consider sharing The AI Leverage with someone who should be building something.

Stefano (@stefano_martell on X)

Co-Founder @The AI Leverage by Bestapps.ai

Keep Reading