Why your AI prototype built with Lovable or Replit isn't 90% done. Hal9 founder Javier Luraschi explains the gap between vibe-coded prototypes and real MVPs.
I see this pattern every single week.
A founder shows me their app. They built it with Lovable or Replit over a few weekends. The UI looks great. Buttons work. They tell me they're "90% done" and just need help with "the last few bits."
Then I look under the hood.
There's no real backend. Data isn't being saved properly. Half the features are hard-coded. They've built 80% of the user interface while missing 100% of the actual backend.
The difference between a prototype and a product is the backend. Until you solve that, you don't have an MVP.
These tools are genuinely amazing. Lovable, Replit, Bolt, Claude Code - they're fantastic for getting something off the ground quickly.
The problem is they're too good at the easy parts.
A founder spends a weekend and gets a working demo. It feels like real progress because it looks like a real product. But the stuff that makes an app actually work - connecting to data sources, integrating functionality, the actual AI logic - that's all missing.
A frustrated founder showed up recently after spending 3 months building an app in Lovable. Features that worked at one point stopped working. He felt like Lovable "didn't understand him."
We rebuilt it correctly in 7 days.
The frontend was solid. It just needed real engineering behind it. Within a week, he tested his MVP during New York Fashion Week and collected real customer feedback. Something that felt out of reach just days before.
We helped build Proofbound, an AI-powered book publishing platform. After working with us on the initial prototype, the founder figured he could continue building himself with Claude Code.
He was right. With practice, he could build demos on his own.
The tough lessons came after. That's when he discovered what he didn't know - and what no AI assistant could tell him.
He wasted weeks trying to make the prototype work on Netlify before discovering a fundamental limitation: Netlify is designed for short-running services, but AI book generation takes 30+ minutes. Longer than their timeout allows. An experienced developer would have known Netlify was wrong for this use case from day one.
Then there was the invisible technical debt. Early on, he wisely used simple database queries for small metadata. Everything worked. Then Claude Code stored larger objects in that same table during a feature addition. Nobody noticed - tests passed, app worked. Until Supabase notified him he'd consumed an entire month's data egress quota in a single day.
Three months of accumulated problems, invisible until they exploded.
Claude Code has another quirk that compounds this: it desperately wants to make you happy. When tests fail, it "fixes" the test rather than the underlying code. When it encounters bugs, it wraps them in workarounds. A 30-minute bug fix becomes a week-long refactoring project once you dig into the accumulated mess.
No matter how good these LLM tools become, they can't help with questions you don't know to ask.
I've seen too many founders burn months and thousands of dollars with freelancers - only to start over.
Freelancers are paid to implement, not to question your strategy. Their incentives aren't aligned - they make more money the longer a project drags on. They won't push back when your idea is off track.
What early-stage founders need is a technical counterpart aligned with their success. Someone who'll tell you Netlify is wrong before you build on it. Someone who knows which questions to ask before you know to ask them.
A client almost spent six figures acquiring a company for their MVP. I took a hard look at the product. Not impressive. Definitely not "buy-a-whole-company" worthy.
We built it in 14 days. Within 30 days, they had 3 potential customers lined up.
Clarity creates speed. Vague vision equals slow progress.
Start with validation, not code. Use LLMs to pressure-test your idea. If customers get confused at the wireframe stage, they'll definitely be confused when it's built.
Build light with no-code first. Lovable and Replit can generate working demos from prompts. It won't scale, but all you need initially is a conversation starter.
Know when you've hit the ceiling. The moment you're wrestling with the tool instead of building your product, you need real engineering.
Get a technical counterpart, not just a coder. Someone who'll push back on bad ideas and knows the traps you're walking into.
Your vibe-coded prototype is valuable. It's a great communication tool. The frontend might even be worth keeping.
But it's not 90% done. It's probably 20% of the way there.
The founders who succeed won't be those who prompted the fastest. They'll be the ones who knew when to admit what they didn't know.
Stuck wrestling with your vibe-coded prototype? Hal9 helps startup founders go from idea to working product in under 30 days. Book a strategy call .
Vibe coding uses AI-powered tools like Lovable or Replit to build applications through natural language prompts. These tools excel at creating user interfaces but typically produce apps with no real backend. Founders end up with most of a UI but none of the data handling, integrations, and AI logic needed for a functional product.
With the right technical support, founders can go from idea to working MVP with paying customers in under 30 days. The key is using vibe coding tools for quick frontend prototypes while partnering with experienced engineers for backend architecture from the start.
A prototype demonstrates an idea visually. A real MVP proves someone will pay for it. The gap is entirely backend: data connections, integrations, AI logic, and scalability. Until you solve the backend, you don't have an MVP.
Hal9 acts as a fractional technical co-founder for startup founders building AI products. We preserve frontend work founders have already done while implementing proper backend architecture and AI integrations. We've rebuilt vibe-coded apps that took founders 3 months in as little as 7 days.
When you're fighting the tool instead of building your product. If features suddenly stop working, if you can't figure out why things break, or if the AI doesn't understand what you're building - you've hit the ceiling. That's when real engineering becomes necessary.