In February 2025, Andrej Karpathy posted a tweet that split the software world down the middle. He called it "vibe coding": "I don't write code line by line anymore. I tell the AI what I want, it writes it, and I just go with the vibes."
Karpathy is the former AI director at Tesla and part of OpenAI's founding team. Not some random guy. And the idea set off both real excitement and real argument.
At AppDuce we've had AI tools in our daily development work for over a year now. Here's what we've actually seen.
What exactly is vibe coding?
Traditional coding: understand the problem, design a solution, write the code line by line, test, debug.
Vibe coding: describe the problem to the AI in plain language, let it generate the code, and look at what comes out. If it works, keep going. If it doesn't, tell the AI to fix it.
The difference is real. In the first approach you control every line. In the second you focus on the output and hand the process details to the AI.
The tools: GitHub Copilot, Cursor, Claude Code, Windsurf, Bolt. They're AI-powered editors and assistants, and they do more than autocomplete. They'll generate whole functions, whole pages, sometimes a full project skeleton.
When does it actually work?
It doesn't work for everything. But in the right spots it's a serious boost.
Prototyping and MVP development. When you want to turn an idea into a working demo fast, these tools are incredible accelerators. CRUD apps, form-based interfaces, simple APIs can come together in minutes. We lean on this for the first drafts of our internal projects.
Boilerplate code. The repetitive stuff every project needs: API routes, database schemas, form validation, test files. Letting the AI write it saves time and keeps things consistent.
Learning new technology. When you're poking at an unfamiliar library or framework, telling the AI "build this with X" is much faster than working through the docs. It's learning by doing, with a helper.
Small tools and scripts. A one-off data transformation, a CSV parser, a batch file rename. Vibe coding is perfect for these.
Where are the limits?
Here's the part most "vibe coding is amazing" pieces skip over.
Complex business logic. Payment systems, user authorization, real-time sync: handing critical business logic to the AI is risky. The generated code looks like it works, then breaks on the edge cases.
Performance optimization. The AI produces generic solutions. It doesn't know your app's specific bottlenecks. N+1 query problems, memory leaks, unnecessary re-renders: only someone who understands the system is going to catch those.
Security. This is the limit that matters most. Shipping AI-generated code to production without a security review is dangerous. SQL injection, XSS, weak input validation: the AI misses these constantly.
Architectural decisions. Monolith or microservice, which database, what caching strategy: those calls need context. The AI can read your existing code, but it can't really grasp your future scale, your team structure, or where the business is going.
Our approach: with AI, not despite AI
At AppDuce we treat AI tools like a junior developer. Great for a first draft. But:
- We review every piece of generated code
- We write the critical business logic ourselves
- We have the AI write tests, then check the results ourselves
- We make architectural decisions as a team
Working this way, our development speed went up by roughly 40%. We don't give up quality for it, because we keep the AI on a short leash.
Will developers lose their jobs?
Short answer: no. Longer answer: the roles will change.
Vibe coding automates the simple, repetitive tasks. That frees developers to spend their attention on the harder problems: system design, user experience, performance, security.
Think of Excel. When it launched, accountants didn't lose their jobs, but the ones who "just did addition and subtraction" lost value. Same shape here. Developers who "only write CRUD" will feel the squeeze, while the ones who can think in systems, make architectural calls, and use AI tools well become more valuable.
Getting started
If you want to try vibe coding, start with one of these:
- Cursor. A VS Code-based AI editor that reads your codebase and generates with that context in mind.
- Claude Code. A terminal-based assistant that can make changes across a whole project.
- GitHub Copilot. The most widely adopted of the bunch. Works inside VS Code and JetBrains.
Begin on a small side project. See the quality of the generated code with your own eyes before you let it anywhere near production.
Final take
Vibe coding is neither magic nor nonsense. Used well, it's a real shift in how you develop, and it delivers serious gains. But go in thinking "I don't know how to code, I'll let the AI do it," and you'll end up with software that's unmaintainable, insecure, and brittle.
The best results come from pairing the AI's speed with a developer's judgment. That's exactly how we work.
If you're interested in using AI not just in your development process but inside the product itself, take a look at our guide on how to add AI to your app.