What’s wrong with vibe coding? Answered by the COAX team

Remember when coding involved long hours at the keyboard, fixing syntax errors, and debugging problems? Vibe coding is trying to rid of this complexity. The vibe coding trend has exploded across tech circles since Andrej Karpathy's famous tweet describing how he simply "gives in to the vibes" while building software. The promise sounds magical: describe what you want, and watch AI build it for you.

But despite the excitement, this approach comes with drawbacks worth understanding before jumping in. This article pulls back the curtain on vibe coding to reveal what happens when you hand the technical reins entirely to AI — and why business owners should approach this trend with healthy skepticism.

‍

Vibe coding definition 

Vibe coding concept is a new way to build software. Instead of writing code manually, developers describe what they want in plain language, and AI generates the code for them. It’s about expressing intent — even loosely — and letting the system figure out the technical details. 

The vibe coding meaning first appeared in February 2025. As Karpathy described in his influential tweet: "I just talk to Composer with SuperWhisper so I barely even touch the keyboard... It's not really coding — I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works." So, how does it work?

• Natural language input. Users articulate what they want — e.g., "Make the button blue and add a fade-in animation" — instead of writing CSS/JavaScript manually.
• AI interpretation. Large language models (LLMs) analyze the request, infer context, and generate (mostly) syntactically correct code.
• Iterative refinement. The developer reviews, tweaks the prompt if needed, and reruns until the output matches their vision.

This approach relies heavily on large language models that understand natural language and programming principles enough to bridge the gap between human intention and machine execution. 

How does vibe coding compare to traditional coding methods?

Traditional software development follows structured methodologies with clear phases — planning, coding, testing, and deployment. Vibe coding throws much of this established workflow out the window.

• Development process

With vibe coding, you're essentially conversing with AI about what you want built. This approach requires clear communication skills rather than programming expertise. You'll need to explain what you want effectively, but you won't need to memorize syntax rules or understand complex programming concepts. 

• Skills required

Traditional development demands a different skill set. Unlike vibe coders, developers master programming languages and understand memory management, data structures, and debugging techniques. They write each line of code deliberately, following technical requirements and best practices. 

• Time spent

The vibe boom changed how quickly someone can start building software. A business owner with zero coding experience could create a functional app in days rather than months. 

• Learning curve

Traditional development typically requires months or years of study before creating anything substantial. Learning programming fundamentals, language syntax, frameworks, and tools represents a significant time investment. 

• Code quality

Vibe coding produces functional code that often works for immediate needs but may lack optimization and clean structure. The code might contain unnecessary complexities or inefficient approaches that only become apparent when scaling. 

• Maintenance

Traditional development typically results in more maintainable code because the developer understands exactly what each component does. Problems can be identified and fixed precisely without guesswork. The codebase generally follows consistent patterns and practices since human developers apply their experience across the entire project.

• Project control & understanding

The difference between workflow and process builder becomes evident in who controls the implementation. In vibe coding, the AI makes many decisions about how to structure the code, which libraries to use, and how to implement specific features. This creates a knowledge gap where the requester may not understand why certain approaches were taken.

‍

Will vibe coding change software engineering?

The vibe coding trend has sparked heated debates about the future of software development. Looking at the explosive growth of vibe coding tools, it's easy to see why some predict a complete transformation of the industry. 

"On the one hand it's a gamechanger, because a lot of people are vibe coding, and over the course of a few prompting cycles you can get something that's amazing and something that – for people who can't program – it's better than anything they could do on their own," notes Matt Wood at Northumbria University.

However, claims that vibe based coding will replace professional developers are greatly exaggerated. AI-generated code often contains subtle flaws, security vulnerabilities, and maintenance challenges that become apparent as projects scale. As Simon Willison points out, "I feel like the job of a software engineer is to produce software that works... a huge amount of the work that we do with software engineers has nothing to do with typing the code."

The real transformation likely lies in how software engineering evolves rather than disappears. Just as spreadsheet software didn't eliminate accountants but changed how they work, vibe coding is the future of development — not as a replacement but as a powerful augmentation. 

This shift may also fundamentally change software culture itself. As one industry analyst noted, "most code will be written by kids/students rather than software engineers," potentially transforming software development into a more youth-driven creative medium. 

What are the main benefits of using vibe coding?

Vibe coding with AI turns software development from a technical specialty into something many people can do. It breaks down barriers that have kept many people from turning their ideas into working software. These aren't just hypothetical benefits — real people are experiencing changes in how they approach creation. 

• Benefits for non-developer entrepreneurs

Tomas Janu, co-founder at Digismoothie and Shopify Partner, shares his journey as a non-developer diving into AI-powered development. Starting with basic tools like Claude and manually pasting into GitHub, he initially struggled when projects grew complex. When he discovered Lovable, a dedicated vibe coding platform, his capabilities expanded. 

"The AI/vibe coding gives me superpowers as a non-technical founder," he says, despite still hitting occasional roadblocks when projects become complicated. His experience highlights that while he can now build functional applications without programming skills, complex projects still present challenges that require patience and persistence.

• Benefits for digital artists

A digital creator who calls themselves "tipudiary" sparked debate online by embracing what they call "vibe art"—using AI to enhance their creative process. "AI didn't kill creativity — it just gave it a faster engine," they explained in a spirited thread discussion. They use tools like Adobe Firefly to convert sketches to digital forms while respecting copyrights. This story reveals how creative professionals adapt AI tools to enhance their existing skills rather than replace them.

• Benefits for professional designers

David Figueiras, a design strategist working with companies like Microsoft and Google, shares: "In just 2 hours, I developed an app that can export batches of ads in multiple languages and sizes simultaneously!" he reports. Using tools like Bolt.new, he built an adjustable Figma plugin — a task that previously would have required either contacting a developer or spending weeks learning JavaScript. 

Design professional Kshitij Agrawal adds that designers are uniquely positioned to use these tools because they "already think systematically," "understand users deeply," and "have the complete context" from research through testing. The visual nature of many vibe coding interfaces plays directly to designers' strengths, allowing them to create functional products rather than just mockups.

The no-code examples highlighted demonstrate that the real magic happens when domain experts can directly implement their ideas without technical translation layers. However, this wider access comes with its limitations.

‍

What are the limitations of vibe coding? Answered by COAX engineers

Vibe coding platforms promise a revolution — write less, build faster. But seasoned developers know the reality is messier. Here’s why relying solely on vibe coding tools can backfire, with insights from our engineering team.

Prototypes ≠ production

Vibe coding tools can quickly produce solutions that appear complete on the surface, but our team consistently finds they miss critical elements needed for production readiness. These tools often deliver:

• Components that function in ideal conditions but fail with edge cases.
• UI that renders properly but lacks accessibility compliance.
• Basic functionality without performance optimizations.
• Code that works but violates security best practices.

A React.js developer with 11+ years of commercial experience, Max Gurak, shares his thoughts on this: "We used a vibe coding platform to generate a data grid component. While it rendered data beautifully, we discovered it loaded all records at once, had no virtualization, and failed accessibility audits. Making it production-ready would require rewriting 80% of the code — I’m so glad it was just a toy project".

Scaling & performance problems

Vibe coding apps excel at creating functional prototypes quickly. However, the generated code often lacks proper architectural patterns, efficient database querying, memory management, and error-handling. AI-generated solutions might fail to implement appropriate connection pooling, lack transaction management in database operations, and not account for race conditions.

Here’s what our Node.js developer, Myroslav Stelmashchuk, says about this issue. With 8+ years of experience in software development, he dives deep into the nuances of vibe coding: "What do I think of vibe coding? I recently talked to a teammate about an AI-built API that fetches user data with individual SQL queries in a loop. It might work fine during testing with 50 records, but would it hold up with 10,000+ users in production? Probably not.” 

The fundamental issue is that AI cannot anticipate scale requirements. While the generated code might solve the immediate problem, it often ignores future growth, lacks performance optimization, lacks the flexibility needed for iterative development, and creates technical debt that compounds over time.

The hidden costs of AI-generated code become clear when examining its limitations compared to professional custom web development. When you check if code was written by AI, these issues become apparent — the solutions work in demos but fail under scrutiny. What appears to be a 90% complete solution often requires 90% more work to make production-grade.

"Functional" doesn’t mean safe

Vibe coding platforms frequently generate code that passes basic functionality tests while containing dangerous security vulnerabilities. These vibe coding tools often miss proper input validation and sanitization, secure authentication patterns, the principle of least privilege implementations, and numerous data protection requirements.

These drawbacks lead to critical security gaps:

• SQL injection risks as generated queries often use string concatenation.
• XSS vulnerabilities are caused by missing output encoding in UI components.
• Potential data exposure — over-fetching and improper serialization tend to cause gaps in inside information that are visible to unauthorized users.
• Finally, auth shortcuts are a risk too, caused by weak session management implementations.

Mariana Dobrianska, a QA/QC engineer with 5+ years of experience, stresses: "We audited a feature built with this approach that exposed admin endpoints to regular users. The AI implemented the requested functionality perfectly — but it didn't understand security contexts."

Vibe coding becomes dangerous when security isn't explicitly specified. Teams must review all AI-generated code very carefully. To make this process simpler, check our web security checklist to see the areas that need to be examined and ensure the safety of your generated code.

Costly maintenance

Maintaining AI-generated code comes with its own set of challenges. One major issue is knowledge vaporization — no developer truly understands the code, let alone how to work with it. Debugging becomes a headache as stack traces often lead to generated boilerplate rather than meaningful logic. 

Upgrading the framework is another risk, as updates can easily break the AI’s "magic" implementations. On top of that, onboarding new engineers is a struggle, as they face a steep learning curve just trying to make sense of it all.

Misha Push, a Ruby on Rails developer with 7+ years of experience, also had a very nuanced experience: “I was testing a new model like I do sometimes and wrote a service component with Sonnet 3.7 — what should’ve taken an hour ended up taking three, just trying to untangle the AI’s "helpful" output.”

The UX illusion of vibe coding

If you look closely, vibe coding often generates visually appealing but non-functional components. While the designs seem to follow a system, the implementation tends to be shallow, missing states and edge cases. Performance can also take a hit, as these components aren’t always built with efficiency in mind.

Natali Deputovych, a UI/UX designer with 4+ years of experience, explains why this happens:

"An AI-generated form looked perfect in Figma but failed basic usability tests. It had no validation, poor error states, and confusing navigation."

The bottom line is that the AI-simply coding approach to UI development creates pixel-perfect but interaction-flawed experiences.

‍

How to make vibe coding work at its best: Best practices & tools

Now you're probably wondering how to build software using vibe coding, but without crashing and burning. While it’s not a silver bullet, these vibe coding best practices reveal where it delivers real value — and where it should never touch your codebase.

• Small-scale rapid prototyping

Vibe coding excels when you need to validate concepts fast. Instead of wasting days coding a feature that might get scrapped, describe your vision in plain language and generate a working prototype in minutes. It’s perfect for UI mockups with basic functionality, proof-of-concept APIs or algorithms, or visualizing data flows or architecture.

• Simple debugging

Stuck on a cryptic error message? Vibe coding tools are good at explaining complex stack traces in human terms, suggesting fixes for common bugs (like null reference exceptions — nothing too complex but time-consuming), and generating test cases to reproduce issues. Unlike traditional debugging, you’re not just patching errors — you’re learning why they occurred. It’s also a chance to learn and grow within your area of expertise.

• Automating the grunt work

Repetitive tasks are a great case for vibe coding. Use it to generate boilerplate code (like CRUD endpoints or DTOs), convert legacy code to modern frameworks, or batch-edit configurations across files. However, as evident from our previous chapter, you should always review outputs, as AI might miss project-specific nuances.

What does it look like in real life with real developers (or non-developers) projects? Let’s explore it as well.

What are some real-world applications of vibe coding?

Vibe coding is gaining traction, with developers and non-technical users experimenting with AI-assisted development. While some see promising results, real-world applications reveal both its potential and limitations. Here’s a closer look at how it’s being used — and the challenges that come with it.

• A small strategy game development 

An experienced developer created a polished, turn-based strategy game in just four hours using vibe coding tools like Cursor and Grok. He started with a simple prompt: "Make a 10x10 grid-based sci-fi strategy game." The AI handled core logic, UI, and even debugging fixes. He barely touched the keyboard — just iterated via natural language.

"It feels like cheating. I’d describe a problem (‘Make warp lanes work like this’), and the AI fixed it. But my coding background helped — I knew when to override its decisions." Here we get to the core of it: Vibe coding accelerates development, but experienced developers get the best results — they spot flaws and guide the AI effectively.

• A newbie “sort of built an app”

A non-technical user attempted to build an app using Replit and Cursor, expecting a no-code-like experience. The AI generated a working prototype within minutes, but the initial excitement faded when the app broke during basic testing. Debugging consumed significant time and resources, forcing users to upgrade to paid plans. 

"I was floored when it worked… until it didn’t. Vibe coding isn’t magic — it’s more like a turbocharged intern that needs supervision."

The lesson here is clear: While vibe coding platforms lower the barrier to entry, they don’t replace problem-solving skills. Without coding knowledge, users often hit roadblocks that AI alone can’t resolve.

• Create an MVP website without coding

A developer created an AI sales coach using vibe coding in under an hour with Cursor. The tool analyzes sales calls by scoring call performance against business metrics, providing detailed feedback across evaluation criteria, and suggesting actionable improvements. While showcasing the demo as a fun example of quick AI development, they emphasize that serious applications require human oversight.

This aligns with a broader trend: vibe coding works well for mockups, MVP development, and early validation, but complex functionality still requires manual development.

Best vibe coding tools

Vibe coding is a powerful addition to the developer toolkit, but it’s not yet a standalone solution. For now, the best approach is to use it strategically — not as a crutch, but as an accelerator for routine tasks. Let’s check some of the best vibe coding tools available.

• Tempo Labs offers a sweet spot for both non-coders and intermediate programmers. It generates a Product Requirements Document and user flow diagrams alongside your code. You can build visually or with AI prompts, plus it supports Stripe payments and Supabase authentication. They recently added GitHub repo importing, though this feature needs some polishing.

• Looking for one of the best coding tools that pioneered the space? Cursor kicked off the vibe coding trend early on. It started with simple chat and code completion, then added the Composer feature so you can tell AI to make changes directly. The downside? It's complicated with rules files and context files to manage. Also, don’t overload it with too much context and remember to switch to agent mode — otherwise, it might refuse to communicate and just advise you to learn coding (not a joke, it did so for a developer).

• Bolt.new from Stackblitz lets you build full-stack apps visually with AI. One of its coolest features is importing designs straight from Figma. Thanks to web containers, you can run Node in your browser and edit code in a browser-based VS Code environment. It works with Supabase for auth and database operations but doesn't have built-in payment system integration yet.

• GitHub Copilot has become a developer favorite with inline code suggestions and multi-step tasks via chat. With security and management features plus organization codebase indexing at the Enterprise level, it's perfect for students, teachers, and individual developers. Seriously, even the non-developer editor of this article tried it to learn some HTML structures! The main challenge is ensuring that code quality aligns with developer intent.

• For non-coders, Lovable.dev might be the friendliest option around. Beyond the usual AI app building, you can select specific parts of your app and ask the AI to make targeted changes. It connects with GitHub, so you can switch between coding in your editor and using Lovable's visual UI prompts.

• Windsurf offers similar features to Cursor but with a nicer interface. The results are comparable, but Windsurf added a cool feature — you can preview your app right in the code editor, which is super convenient.

• Replit offers an all-in-one solution to make, launch, and scale apps with real-time code completion. Higher tiers unlock collaboration, security, and auto-scale deployments, though resource-intensive projects might face performance limitations.

• Trae, coming from the TikTok makers, has a great user experience and a generous free tier. It lacks MCP integration, though, so you can't use it for your entire build process. Like others, it struggles with context management for existing codebases.
• The powerful Sourcegraph is one of the best coding tools for professional developers working with large codebases and teams. It connects hundreds of repositories for searching and batch changes like refactoring or security fixes. Its VS Code extension, Cody, adds chat and code completion powered by impressive search capabilities.
• Devin works as an autonomous developer who plans, codes, debugs, and tests with minimal help. The catch? It only interfaces through Slack, which is odd for solo developers but works for teams.

• Claude Code is a newer terminal-based tool that "reads and understands" your codebase and maintains this memory for future sessions. While powerful, it's one of the most expensive options, with some simple changes costing around $5 in tokens.
• For those seeking the best AI to write code with robust error handling, Perplexity AI excels at coding syntax, giving debugging tips, and explaining common coding concepts. Higher tier plans unlock advanced AI models with various deployment options, though it sometimes lacks creativity in generating solutions.
• Want to use the mainstream power of ChatGPT for coding? OpenAI's popular AI offers general capabilities for complex problem-solving and research, with API integration available for your apps. While the free version has limited access, it remains one of the best AI to use for coding for beginners looking to dip their toes into vibe coding.‍
• Cline packages powerful features as a VS Code extension rather than a complete fork. With real-time code suggestions and custom instructions, it's great for smaller, focused coding tasks where you need specific AI models. Its raw generation quality is similar to Cursor since they use the same AI.

The choice of the best AI tool depends entirely on your workflow and needs. For visual builders, check out Tempo Labs or Bolt.new. Solo coders might love Cursor or Windsurf. Teams with large codebases should consider Sourcegraph. Beginners can start with ChatGPT, Claude, or Replit's user-friendly interfaces. Whatever your level, there's a vibe coding tool for you to adapt to your needs.

However, as AI code assistance tools become more sophisticated, so do AI-based cyber threats. To future-proof your infrastructure, consider automation testing services focused on AI-generated code audits, analysis tools that detect vulnerable patterns in machine-written code, and runtime monitoring solutions that catch anomalies unique to AI-assisted development. At COAX, we’ve seen and tested it all, so contact us if you need assistance from our QA team.

‍

Is vibe coding a good thing or a bad thing?

The vibe coding trend offers clear benefits: it helps non-programmers build software, speeds up prototyping, and shifts focus from technical details to solving real problems. The vibe examples we've seen show how entrepreneurs, artists, and designers can now create digital products without deep technical knowledge.

However, vibe coding isn't perfect. AI-generated code often lacks security features, struggles with complex tasks, and can be difficult to maintain long-term. What works in a demo might not hold up in the real world.

Rather than replacing traditional development, vibe coding works best as a complement to it. It's great for quick prototypes and simple tasks, while professional coding remains necessary for building robust, secure, and scalable systems.

Vibe coding shows the power of AI to accelerate digital creation — but turning those fast prototypes into real, reliable products takes expert guidance. Our team combines the flexibility of AI-powered tools with solid engineering practices to help you move from idea to execution. Whether you’re experimenting with AI-generated code or need custom AI development from scratch, we’re here to build with you.

‍

FAQ

What is vibe coding?

Vibe coding is an approach to software development where, instead of writing code manually, developers describe what they want in plain language and AI generates the code for them. It originated in February 2025 after Andrej Karpathy's influential tweet about "giving in to the vibes" while building software.

How long does it take to learn vibe coding?

Learning vibe coding is primarily about effective communication with AI rather than mastering technical concepts. Most people can start building simple applications within days, compared to the months or years required to learn traditional programming.

Is vibe coding suitable for large enterprise applications?

Currently, vibe coding is better suited for prototypes, MVPs, and smaller applications. Large enterprise applications often require considerations for security, scalability, and maintainability that vibe coding tools may not adequately address without experienced oversight.

What skills are still valuable in a vibe coding workflow?

Understanding basic programming concepts and system design principles and being able to evaluate code quality remain valuable. Communication skills become particularly important since you need to clearly explain your requirements to the AI.

How to check if code is AI generated?

To check if code is AI generated, look for these tell-tale signs: inconsistent naming conventions, overly generic implementations, unusual commenting patterns, missing edge case handling, and code that works superficially but has hidden issues in security or scalability. AI-generated code often lacks the nuanced architecture decisions that experienced developers make.