DEV Community

Aakash Verma
Aakash Verma

Posted on Originally published at lovable2live.com

Vibe Coding an App: Where It Stops Working

What is vibe coding? It's building software by describing what you want to an AI and accepting what it writes without reading the code. Andrej Karpathy named it in a post on X in February 2025: "fully give in to the vibes, embrace exponentials, and forget that the code even exists." The same post said it was "not too bad for throwaway weekend projects". By November 2025 Collins Dictionary had made it its Word of the Year. A year on, people use it for apps that take card payments, which is not what the man who named it had in mind.

I don't think that's wrong, by the way. Most of the apps we work on now started as vibe-coded prototypes in Lovable, Claude Code, Cursor, Bolt or Replit, and the founders who built them got further in a month than they would have in six with a freelancer. The problem is never that the code was written by an AI. It's that nobody noticed the moment the project stopped being a weekend thing.

Which tool, and what it costs

You don't need to agonise over this. Lovable is the easiest start if you can't code at all: it hosts the app, gives you a database through Lovable Cloud, and you mostly talk to it in a chat box. Claude Code and Cursor work on a real code folder on your machine (or in the browser now, for Claude Code), which is better once there's a second person involved or you want the code in GitHub from day one. Bolt, Replit, v0 and Base44 sit somewhere in between.

On price, budget $20 to $30 a month for the tool while you're exploring and $100 or more once you're building every day. The plans moved a lot this year; our Claude Code vs Cursor post has the numbers as they stood on November 17, and Lovable pricing covers where Lovable credits go. The tool bill isn't the expensive part. The database, email, auth and hosting bills that show up after launch are, and a Lovable Cloud bill in particular can climb quietly. Our free Lovable cost check reads your repo and tells you what it would cost elsewhere.

Where it stops working

Here's the honest version. Vibe coding works for as long as you are the only person who'd get hurt by a bug. A broken button in your own demo costs you ten minutes. The same broken button with 40 paying users costs refunds, support email and, sometimes, a database you can't put back together.

A timeline of a vibe-coded app from the first demo to the first payment, showing which risks switch on at each step

The risks switch on one at a time, and they're easy to miss because the app looks exactly the same before and after each one.

The first stranger signs up. Now someone else's data is in your database, and every table needs row level security that actually checks who's asking. This isn't theoretical. In 2025 a researcher found that more than 170 apps built with Lovable let anyone read other users' data because the generated Supabase policies were missing or wrong; it was published as CVE-2025-48757. Lovable has security scans now, and they're good at spotting a table with no policy at all. They're much worse at spotting a policy that exists and is wrong.

The first payment. Stripe webhooks, refunds, a test key left in a live build. AI tools write checkout pages very happily. They rarely write the webhook handler that marks the order paid when the customer closes the tab before the redirect, and that's the one that matters.

The first change after launch. This is the one that catches people. You ask for "make the dashboard faster" and the AI rewrites a query, renames a column, or edits a migration that's already been applied to production. In a prototype that's fine. With live users it's an outage, and the AI will cheerfully tell you it's fixed.

Veracode tested code from more than 100 language models on 80 tasks last year and found security flaws in 45% of the results, a rate that hadn't improved as the models got better at writing code that runs. That number is about isolated tasks, not your whole app, but it matches what we see in reviews: code that works and leaks.

The founder who checked everything except one thing

A founder came to us in October with a vibe-coded tutoring marketplace. She'd done more than most: she'd asked the AI to "add security", run Lovable's scan, read the warnings, fixed them. Tutors could upload ID documents for verification. The storage bucket for those was public, because an early prompt had said "make the profile photos load faster" and the AI made the whole bucket public instead of just the photos folder. Every passport scan had a guessable URL. The scan didn't flag it because public buckets are a legitimate choice for photos.

The fix took us about twenty minutes. Finding it took a person who knew to look at storage policies and ask why that bucket existed.

Keep vibe coding, add a second pair of eyes

I wouldn't tell anyone to stop. The speed is real and a lot of good products will be built this way. What I'd say is: once a stranger can sign up, someone who reads code should see every change before it goes live, and the database, backups and deploys should belong to someone who isn't guessing.

That's the job we do. If your app is about to take real users or money, we take Lovable apps to production and Claude Code or Cursor apps to production from $500, usually in 5 to 7 days. If you'd rather keep building with the AI yourself, our AI build partner setup is from $600 once, then from $500 a month, and every pull request the AI writes gets reviewed within a working day. Just want someone to look for the passport-bucket kind of problem? The security review for AI-built apps is $350. Not sure it's worth paying anyone yet? Run the numbers in the app development cost calculator first.


Originally published on lovable2live.com. Aakash Verma, software developer and founder of Lovable 2 Live.

Top comments (0)