What You Need to Know About Vibe Coding

Vibe coding is great – until it isn’t. 

Using AI to build software with just a prompt, without having to write any code yourself, quite simply wasn’t possible as recently as 2023. It’s amazing.

But it’s important to know its limitations. If you’re a non-technical entrepreneur or manager using ChatGPT, Claude, Cursor, Lovable or any other AI coding tool, you can cause yourself and your company a lot of trouble by trusting vibe coding tools too much.

What is vibe coding?

Before we move on, it’s important to understand what makes vibe coding different from other forms of AI-assisted programming.

In the words of Simon Willison, legendary programmer and co-creator of Django, vibe coding means “building software with an LLM without reviewing the code it writes”.

He goes on to explain that vibe coding is great when the stakes are low, your money isn’t on the line, and you have no reason to be concerned about security.

Vibe coding is not the same as professional AI-assisted programming. When you’re responsible for building software that’s stable, secure, scalable, maintainable, explainable, and operational in a real-world environment with high stakes – you’re no longer vibe coding. As Mr Willison explains:

“If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone else that’s not vibe coding, it’s software development. The usage of an LLM to support that activity is immaterial.”

Problem #1: Is vibe coding a lot of code a bad thing?

AI coding tools are notoriously bad at keeping their outputs short. At least that’s the case in June, 2025, which is important to mention considering how quickly this ecosystem is evolving.

And this might shock people that don’t have programming experience, but code is a liability. Professional software developers usually opt for less code wherever it’s possible. 

“Writing less code is generally better than writing more code”

-David J Malan, Computer Science Professor leading the most popular entry-level programming course, Harvard’s CS50

The less code you write, the less likely it is that it’ll contain mistakes. Not a big issue when you’re vibe coding a simple program for yourself. It starts being a problem when you’re building a digital product that you intend to sell, or you’re working with professional developers that have to review your code. 

Problem #2: Is vibe coding safe?

Vibe coding itself is neither safe nor risky. It can get dangerous depending on how you’re using it, what you’re building, and what you do with vibe-coded software.

When you search X (previously Twitter) for terms like “vibe coding” and “hack,” you’ll find no shortage of stories from people that published vibe-coded apps only for them to be immediately hacked.

Or you may end up in a similar situation as Jason Lemkin, founder of SaaS community SaaStr. Mr Lemkin went all in on vibe coding, spending several days and hundreds of dollars using Replit AI to build software.

After a while, things got weird. He found out that the AI was covering up bugs and issues, creating fake data and reports, and lying about unit tests. Later, the AI deleted his database, and advised Lemkin that it was impossible to restore even though that was not the case. 

The story had a somewhat positive ending. Replit’s CEO quickly responded to the issue, and they performed a security overhaul within 72 hours to fix the issues that caused this situation.

But this remains a cautionary tale for entrepreneurs and managers that believe they can build commercial, market-ready products purely with vibe coding. Right now, it’s risky and far from simple.

If the stakes are low and you’re just building something for yourself, it will be hard to do any real damage. But if you want to release a stable, secure, and scalable digital product, the best that vibe coding can currently give you is a prototype. After that, let your in-house developers take over, or outsource further development to a reliable agency.

Problem #3: Is vibe-coded software easy to scale?

Scaling vibe-coded software is hard.

Vibe coding leads to straightforward solutions instead of optimized ones, with code that uses resources inefficiently and struggles when it’s met with more data.

Performance degradation under load is not a recipe for success. At worst, it can lead to a dramatic increase in cloud computing costs, costing you tens of thousands of dollars.

One industry expert who experimented with vibe coding noted that while his vibe coded creations worked, the code was "enough to make a senior software engineer go blind."

It’s fine for personal projects, and it’s a cool way for people to get into programming without having to learn for several months. But when you aim to serve thousands of users simultaneously, those inefficiencies become a huge roadblock.

If you're planning to grow beyond a few hundred users, you'll likely need to rebuild or significantly refactor your vibe-coded foundation.

Problem #4: Is it easy to debug or maintain vibe-coded software? 

As Craig Zingerline, experienced vibe coder and seasoned founder described it in his LinkedIn post:

”What usually starts out strong with AI vibe coding products ends up in refactoring hell. Anything complex takes time.”

Vibe coding a complex product and fixing it later on might end up taking longer than just having an experienced developer create it from the ground up.

In professional programming, many problems are solved in the early stages of a project. Ideation, workshops, product design, prototyping – it all helps prevent issues before development begins.

The fundamental problem is that traditional debugging requires understanding how the code works, but vibe coding explicitly encourages you to ignore the code. 

As one seasoned developer observed: 

"Debugging AI code can prove difficult, especially if you're not familiar with it. Developers who try to debug AI-generated code, especially code they didn't write or understand, will struggle to reverse-engineer how it works."

All of this means that purely vibe-coded software is unlikely to ever reach the maintenance stage.

Problem #5: Vibe coding and technical debt 

Some industry experts are expecting a "technical debt tsunami" as the result of vibe coding.

Unlike traditional technical debt, which accumulates gradually, vibe coding can generate massive debt overnight. Without proper supervision, vibe coding generates code that ignores:

• Proper architecture patterns

• Code modularity and reusability

• Comprehensive documentation

• Robust testing frameworks

• Security best practices

The cruel irony is that fixing technical debt is often delayed because tasks like code cleanup, backend refactoring, or infrastructure upgrades don't produce visible features that users request. But they are vital for keeping products operational and sustainable long-term.

Critical warning signs that indicate AI technical debt problems include: 

• increasing time spent debugging, 

• declining system performance, 

• frequent production issues, 

• difficulty implementing new features, 

• rising cloud infrastructure costs, 

• team members who can't explain how critical systems work.

When vibe coding makes sense

One of the biggest success stories involving vibe coding was when famous solopreneur Pieter Levels spent half an hour creating a game that almost immediately started generating thousands of dollars in revenue.

Using Cursor, Levels created a Python websockets server and complete game system, showcasing how AI can assist in building complex systems when guided by someone with clear vision and technical understanding.

The thing is, it’s highly unlikely that anyone would be able to replicate this success. Levels is a well-known figure, a master of marketing, a skilled programmer and startup founder with many profitable projects under his belt. 

It’s not a good benchmark to aim for when you embark on your own vibe coding journey. But there’s still a lot you can achieve with this skill:

Rapid prototyping – vibe coding is awesome for quickly validating ideas and building proof-of-concept applications. As one expert notes, "it depends what you build." Vibe coding works best for "simple, lightweight, straightforward consumer apps."

Learning and exploration – non-technical founders report that vibe coding helps them understand what's possible and communicate more effectively with technical teams later.

Internal or personal tools – simple internal dashboards, data processing scripts, or workflow automation tools are good candidates for vibe coding.

When to avoid vibe coding entirely

Don’t vibe code when you’re dealing with enterprise security requirements. If your application handles sensitive data, requires compliance certifications, or needs enterprise-grade security, vibe coding is not appropriate.

Be very careful about using vibe coding against anything that's charged based on usage. There are horror stories of people racking up thousands of dollars in API charges from poorly implemented vibe-coded features.

Don’t expect vibe coding to output complex business logic. If your application requires sophisticated algorithms, complex data relationships, or intricate business rules, the limitations of current AI will likely frustrate you.

How to approach vibe coding

Remember Jason Lemkin’s story of AI lying to him and deleting his database? It hasn’t discouraged the SaaStr CEO from vibe coding, and he has since put out a detailed guide on how to prevent similar situations. A few of the key highlights include:

• Be prepared to transition from pure vibe coding by adopting either a hybrid approach (vibe coding for rapid prototyping and working with software developers to turn it into proper code), or a complete rebuild (use the results of vibe coding as a specification for a development team).

• Consider that custom design requires substantial work, enterprise security features have to be manually implemented, complex integrations may be impossible, and performance optimization is best done by experienced developers.

• Document your business logic, maintain clear specifications, and regularly evaluate whether you're approaching the vibe coding platform's ceiling for your use case. This documentation becomes invaluable whether you stay on the platform or eventually transition to traditional development.

• Remember that AI systems actively try to appear successful even when failing. Implement independent verification for essentials like database operations and data integrity, API integrations, authentication and security, or backup and recovery systems.

And ultimately, if you’re unable to do something now, don’t treat it as a failure, because “the platforms are evolving rapidly; what's impossible today might be straightforward in six months."

The future of vibe coding

The future of vibe coding is becoming multimodal, incorporating voice, visual, and text-based coding to enhance productivity. We're seeing the emergence of:

• Voice-driven coding – developers can speak their intentions and see code appear

• Visual programming interfaces – drag-and-drop environments generate sophisticated backend logic

Specialized tools are emerging for specific domains like game development, embedded systems or data science. These domain-specific variants incorporate best practices and patterns unique to their fields, producing more precise and optimized solutions.

Specialization addresses one of current vibe coding's biggest limitations – the generic, one-size-fits-all approach that often produces functional but non-optimal code.

Current platforms are rapidly evolving to address known limitations:

• Better code export – platforms are improving their ability to generate clean, maintainable code that professional developers can work with

• Enhanced security frameworks – built-in enterprise security features reduce the need for manual security implementation

• Improved debugging tools – better visibility into AI-generated code and its behavior

• Advanced integration capabilities – more sophisticated APIs and third-party service connections

All in all, the future of vibe coding is bright.

Will vibe coding tools replace software developers?

As things stand right now, vibe coding tools can’t replace software developers. Not a chance. 

In the future? All anyone can do now is guess, and predictions are varied:

• Anthropic CEO Dario Amodei expects that 90% of all code will be written by AI in 2026.

• Mark Zuckerberg expects AI to replace mid-level engineers by the end of 2025, freeing them up to do more creative work.

• Google reported that 25% of their code was already being written by AI in the last quarter of 2024.

• IMB CEO Arvind Krishna predicts that up to 30% of code will be written by AI, ultimately serving to make developers more productive, not replace them.

The most rational perspective for now is that vibe coding platforms are a new tool that:

• enables anyone to build basic software without having to learn about programming at all, which wasn’t possible before,

• speeds up the workflows of professional software developers by allowing them to quickly test different ideas.

So use it within reason, but don’t expect that you’ll get a market-ready software product with pure vibe coding. To achieve that, you need to work with experts.