The State of the Vibe
Look, I’ve been building software for 30 years and I still managed to get myself stuck in a doom loop last week where Claude kept regenerating the same broken authentication flow over and over. So when people ask me if vibe coding is “real” development, I just laugh because honestly we’re all still figuring this out together.
But here we are in 2026 and the landscape looks nothing like it did even 18 months ago. Collins Dictionary named “vibe coding” their Word of the Year for 2025. Y Combinator reported that 25% of their Winter 2025 batch had codebases that were 95% AI-generated. And somewhere along the way we went from “AI is just fancy autocomplete” to “I built a SaaS over the weekend for the cost of an API subscription.”
So let’s talk about where we actually are.
The Tools Have Gotten Ridiculous
When Andrej Karpathy coined the term back in February 2025, he described it as fully giving in to the vibes and forgetting that the code even exists. That was a tweet about his personal workflow. Now we’ve got an entire ecosystem of tools competing to let you do exactly that.
The Full App Generators
These are the tools that promise to turn your prompt into a working application, and honestly they’ve gotten scary good at certain things:
Lovable has emerged as the go-to for non-technical founders who want polished demos fast. The UI generation is genuinely impressive and you can get a functional web demo in under 10 minutes for typical apps. They integrate with Supabase for backend, Stripe for payments, and the generated codebases are clean enough that real developers can actually work with them afterward. The catch is that complexity ceiling is real and eventually you’re going to hit walls.
Bolt.new takes a different approach with a full browser-based IDE where you’re working alongside the AI. It’s more technical than Lovable but you get way more control and you can actually see what’s happening. Token costs can spike though if you’re iterating heavily.
Replit Agent (now on version 3) is probably the most ambitious of the bunch. They’re going for full end-to-end development with autonomous agents that can plan, execute, debug, and deploy. They claim 200 minutes of autonomous development on their agent tier. Zinus apparently saved $140k using it. The execution-centered approach means it’s constantly running and testing things rather than just generating static code.
v0 from Vercel is the specialist in the group. It’s laser focused on React component generation and the output is genuinely production-quality. The limitation is it’s frontend only so you’re bringing your own backend, but if you’re in the Next.js ecosystem it slots in perfectly.
The Agentic IDEs
Then there’s the other category that’s taken over, which is AI-native development environments that go way beyond autocomplete:
Cursor has basically won the “IDE for people who actually write code” category. The agent mode with Composer lets the AI work across multiple files and understand your whole project context. They’ve got something like 360,000 paying customers and hit $200 million ARR in Q1 2025. The @ symbol system for referencing context is elegant once you learn it, and having Claude 4 models available makes a real difference on complex tasks.
Windsurf (which started as Codeium’s IDE and is now owned by OpenAI apparently?) is the cheaper alternative at $15/month versus Cursor’s $20. Their Cascade feature is genuinely good at understanding large codebases and the UX is more intuitive for beginners. Developers on Reddit consistently praise it for the “just works” factor of clicking preview and having it spin up a dev server automatically.
Claude Code is the terminal-first option for developers who don’t want to leave their command line. It’s got the deepest codebase understanding because it literally maps your entire project and can reason about dependencies across files. The 200K context window means it can hold a lot in its head at once. If you’re doing heavy refactoring or working with legacy codebases, this is probably the move. There is also a vscode plugin.
And then there’s Cline which runs as a VS Code extension and acts more like a true agent than the others, it can run commands, check results, fix its own mistakes, and keep going. Worth checking out if you want agentic capabilities without switching IDEs entirely.
The Hangover Is Real
Here’s the part nobody wants to talk about at conferences but everyone’s dealing with in practice.
Fast Company ran a piece in September calling it “the vibe coding hangover” and they weren’t wrong. Senior engineers at places like PayPal are calling AI-generated code “development hell” when they have to maintain it. A study by METR found that apps built purely through vibes were 40% more likely to contain critical security vulnerabilities like unencrypted databases.
The Lovable security thing was particularly rough. They found 170 out of 1,645 Lovable-created apps had vulnerabilities that would let anyone access personal user data. Not because Lovable is bad, but because the people building didn’t know enough to catch the problems. You can’t see what you don’t understand.
And then there’s the infamous SaaStr incident where Jason Lemkin was building something in Replit and the AI agent just deleted his entire database despite explicit instructions not to. Then lied about it. Lost work on a non-production app is one thing but imagine that happening with real users.
This is the thing that keeps me up at night honestly. The speed of vibe coding comes from skipping the review process. But that’s also exactly where the risks hide.
Where We’re Actually Headed
MIT Tech Review had a good framing on this. They’re talking about the shift from “vibe coding” to “context engineering.” The idea being that pure vibes were never sustainable and the industry is now figuring out how to actually manage the context and ground truth that AI systems need to produce reliable output.
Simon Willison (whose opinion I trust on this stuff) makes a distinction between vibe coding where you skip understanding entirely, and what he calls “vibe engineering” where you’re using AI to accelerate work you actually understand. The AI amplifies existing expertise rather than replacing it.
I think that’s right. The tools are going to keep getting better. We’ll probably see self-healing software that monitors production and fixes bugs without human intervention. The “Product Architect” role that some people are predicting, where you’re basically a creative director and security auditor rather than a traditional developer, makes a lot of sense.
But here’s my honest take after building with these tools for over a year now: the prompting skill matters way more than people realize. Some days I can solve a problem in 2 prompts. Other days I’m 20 prompts deep and going backwards. The difference isn’t the AI, it’s whether I gave it the right context and constraints upfront.
That’s actually why I built VibeQ. Because 73% of developers use AI daily now but there’s no structured way to practice the actual skill of prompting effectively. We’re all just kind of winging it and hoping for the best.
The Bottom Line
Vibe coding is real. It’s not going away. The tools have gotten good enough that non-technical people can ship functional products and technical people can move significantly faster on certain types of work.
But it’s also not magic. The doom loops are real. The security vulnerabilities are real. The “vibe coding hangover” where you’re stuck with unmaintainable code you don’t understand is real.
The developers who thrive in 2026 aren’t the ones who can type the fastest or memorize the most syntax. They’re the ones who can clearly articulate what they want, provide the right context, and know enough to verify the output actually works. That’s the new skill stack.
Whether you’re a 30-year veteran like me or someone just getting started, we’re all learning this together. The vibes are strong but the fundamentals still matter.
Now if you’ll excuse me, I need to go debug another auth flow that Claude swears is working perfectly.
Want to actually measure and improve your prompting skills? Check out VibeQ where we’ve built the first competitive platform for practicing AI-assisted development. Daily challenges, VQ scoring, structured training. No more guessing whether you’re good at this or just lucky.