Vibe Coding Just Landed. Here Is What It Can Actually Do.

In February 2025, Andrej Karpathy, co-founder of OpenAI and former AI lead at Tesla, posted a few sentences on X that coined a term the entire tech world is now obsessing over. He described a new approach to building software where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." He called it vibe coding. Within weeks, Merriam-Webster had listed the term. Within a month, Y Combinator reported that 25% of startups in its Winter 2025 batch had codebases that were 95% AI-generated.

The promise is extraordinary: describe what you want in plain English, and AI builds it for you. No programming language required. No years of computer science education. Just you, an idea, and a chat interface. Tools like Lovable, Bolt.new, and Vercel’s v0 have turned this from a theoretical concept into something anyone can try right now, for free, in a browser tab.

We have spent the last few months building with these tools, breaking them, pushing them to their limits, and watching what comes out the other side. Here is our honest assessment of where vibe coding actually stands today.

What These Tools Can Do Is Genuinely Impressive

Let us start with the good, because it deserves to be said clearly: the current generation of vibe coding tools is remarkable. Five years ago, building a functional web application with a database, user authentication, and a polished interface required months of work from an experienced developer. Today, you can describe that application in a paragraph and have a working prototype in under an hour.

Lovable, which launched as GPT Engineer and has quickly become one of the most popular platforms, handles the full stack from a single chat interface. It generates React frontends, connects to Supabase for the database, handles authentication, and deploys to a live URL. Bolt.new, built by StackBlitz, runs entirely in the browser with zero local setup. Vercel’s v0 produces some of the cleanest UI components in the space. Each platform has its strengths, and all of them can produce something that looks and feels like a real product in a fraction of the time it would take to build manually.

For prototyping, this is transformative. For validating a business idea before investing in full development, this is a genuine breakthrough. For building internal tools, dashboards, and simple applications, these platforms deliver real value right now.

But Here Is Where It Gets Complicated

The marketing pitch for vibe coding platforms tends to stop at the demo. And the demo is always impressive. The problem is that demos and production-ready software are two very different things.

Kevin Roose, a technology columnist for the New York Times, wrote about his experiments with vibe coding in February 2025. As a non-developer, he built several small applications and described them as "software for one" – useful personal tools. But he was honest about the limitations: the results were "often limited and prone to errors." In one case, the AI-generated code fabricated fake reviews for an e-commerce application he was building. The AI did not flag that the reviews were not real. It just made them up and presented them as content.

This is not a bug in one specific tool. It is a characteristic of how large language models work. They generate plausible-looking output. Sometimes that output is correct. Sometimes it is confidently, invisibly wrong. When the output is a blog post, the stakes are low. When the output is code that handles user data, processes payments, or makes business decisions, the stakes are very different.

The 80/20 Wall

Every team and every developer who has spent meaningful time with these tools describes the same experience. The first 80% of a project is magical. You describe what you want, the AI builds it, and you feel like the future has arrived. Then you hit the last 20%.

The last 20% is edge cases. It is error handling. It is the authentication flow that works for most users but breaks for the ones with special characters in their email addresses. It is the payment integration that processes test transactions perfectly and fails silently in production. It is the responsive layout that looks great on the screen sizes the AI tested and collapses on the ones it did not.

And here is the critical part: that last 20% requires exactly the software engineering knowledge that vibe coding promised you would not need. If you do not understand the code the AI generated, you cannot debug the code the AI generated. You can paste the error message back into the AI and hope it fixes it, which is exactly what Karpathy described in his original post. Sometimes it works. Sometimes it introduces a new bug while fixing the first one. Sometimes it confidently tells you the problem is fixed when it is not.

The Quality Problem Nobody Wants to Talk About

GitClear, a code analytics company, published the results of a longitudinal analysis of 211 million lines of code changes spanning 2020 to 2024. The findings are sobering. Over the period when AI coding tools went mainstream, the volume of code refactoring – the practice of improving existing code – dropped from 25% of all changed lines to under 10%. Code duplication increased approximately fourfold. Copy-pasted code exceeded moved code for the first time in two decades. Code churn, where code is rewritten shortly after being merged, nearly doubled.

What this tells us is that AI tools are generating a lot of code, fast, but the quality of that code is declining by multiple measures. More duplication. Less refactoring. More churn. These are the metrics that predict long-term maintainability problems. Code that is easy to generate is not necessarily code that is easy to maintain, scale, or secure.

In May 2025, security researchers found that 170 out of 1,645 web applications built on Lovable had security vulnerabilities that would allow personal information to be accessed by anyone. Not sophisticated exploits. Just misconfigured databases and missing access controls – the kind of security basics that an experienced developer would catch but that a non-developer using a vibe coding tool would have no way of knowing to check.

So What Is Vibe Coding Actually Good For?

We are enthusiastic about this technology, and we want to be clear about why. Vibe coding is not a scam. It is not vapourware. It is a genuinely useful set of tools that has specific, real applications:

• •Prototyping and validation: Building a working prototype to test a business idea before committing to full development. This is where vibe coding delivers its highest value today.
• •Internal tools: Simple dashboards, admin panels, and data viewers for internal use where security requirements are lower and the user base is small and known.
• •Learning: Non-developers can use these tools to understand how software works by seeing their ideas translated into code. This is genuinely educational.
• •Accelerating experienced developers: Developers who understand the code the AI generates can use these tools to skip boilerplate and focus on architecture and business logic. This is where the productivity gains are real and substantial.

What vibe coding is not good for, at least not yet: production applications that handle sensitive user data, payment processing, healthcare information, or any use case where a security vulnerability or a silent bug has real consequences. Not because the tools are bad, but because the tools generate code that requires human review to be safe, and the entire pitch of vibe coding is that you do not need to review the code.

Where This Is Headed

The trajectory of these tools is clear: they are going to get better. The underlying language models are improving rapidly. The platforms are adding features every week. The venture capital flowing into this space, over $5 billion in AI coding tools in 2024 alone, virtually guarantees continued innovation.

We expect that within 12 to 18 months, the security gap will narrow significantly. The 80/20 wall will move to 90/10. The tools will handle more complexity, more edge cases, more of the unglamorous work that currently requires human expertise. This is not wishful thinking. It is the natural trajectory of every software tool category that has ever existed.

But today, in June 2025, the honest assessment is this: vibe coding is a remarkable prototyping tool, a useful accelerator for experienced developers, and a genuinely exciting preview of where software development is headed. It is not, yet, a replacement for the skill of understanding what you are building and why. The businesses that treat it as an accelerator will benefit enormously. The ones that treat it as a substitute for engineering will learn some expensive lessons.

The vibes are real. Just make sure someone on your team can read the code.