I Tried Building 5 Apps in One Weekend with Vibe Coding
Saturday morning, coffee in hand, I decided to do something stupid: build five apps in 48 hours using only AI tools. No pre-existing code. No exceptions. Here's exactly what happened.
I've been skeptical of the "built a startup in a weekend" posts for a while. Half are marketing for the tool that built it. The other half are landing pages with nothing behind them.
So I ran the experiment honestly. One weekend. Five apps. Every vibe coding tool I'd been curious about. Starting from scratch each time.
Total spent: $47. Shipped: 3 out of 5. One of them I'm actually thinking about launching.
Here's what happened, hour by hour.
Saturday, 8:47 AM — The Setup
Coffee, laptop, a notebook for tracking what worked. I already paid $20/mo for Cursor. I subscribed to Bolt.new ($20, first month), Lovable ($19, first month), and put $8 on Claude Code API credits. I used v0's free tier and had Windsurf's free Pro trial already running.
Total damage: $47 for the weekend. Less than most people spend on dinner.
My rule: I had to actually ship each app to a live URL before moving on. No "works on my machine." No excuses. (All the tools I used are in the directory if you want the specs.)
App 1 — AgentPad (9am–1pm, Bolt.new)
The idea: a landing page generator that turns a one-line agent description into a full marketing site. Niche, but I had a use case — I wanted one for a side project.
I typed: "Build me a tool where I paste in a description of an AI agent and it generates a full landing page with hero, features, pricing, and FAQ. Use Tailwind and Next.js. Deploy-ready."
Bolt spat out a working preview in about four minutes. The first version was hilariously generic — stock "Hero Section" text, placeholder images, no real logic. The second prompt fixed the content generation. The third fixed the layout.
The one thing that bit me was the Vercel deploy hook. Bolt couldn't wire up the OAuth flow on its own, so I exported the code and ran vercel deploy from my terminal. Five minutes. Fine.
By noon I had a working tool at agentpad.vercel.app. Real input, real output, real deploy. 237 lines of code I didn't write.
Verdict: Bolt crushed this. For a self-contained landing page tool, it's faster than anything else I tried. Just don't expect it to wire up infrastructure for you.
App 2 — RevenueCub (2pm–6pm, Lovable)
An MRR tracker for indie hackers. Paste your Stripe key, see your MRR, churn, and growth in one dashboard. Lovable's Supabase integration is what I wanted to stress-test.
First prompt got me an authed dashboard with Supabase tables auto-created. Auth, users table, a dashboard route, all in under forty minutes. Impressive. Second prompt added the Stripe connector skeleton. Third prompt broke everything.
Here's the thing nobody warns you about: these tools have regression amnesia. I asked Lovable to fix the webhook signature verification, and it rewrote three files that were already working. Every "fix" round pulled one thing forward and two things back.
At 5:40pm I gave up on one-shotting the Stripe piece. I switched to Cursor, wrote a proper webhook handler in about 180 lines, pasted it back into the Lovable project, and called it. Shipped at 6:30pm with mock Stripe data and a TODO for the real sync.
Verdict: Lovable is brilliant until it isn't. For auth-plus-CRUD it's one-shot territory. The moment you need a real third-party integration with subtle logic, you're back in Cursor. Read the Bolt vs Lovable breakdown for the full side-by-side.
Get the Vibe Coding Cheat Sheet
Best tool for every use case + pricing + pro tips. One page, zero fluff. Plus weekly updates on new tools.
App 3 — The Disaster (8pm–11:45pm, Cursor)
I got cocky. Tried to build an AI support chatbot with a custom knowledge base, vector embeddings, streaming responses, and an admin dashboard to review bad replies. In Cursor. At 8pm on a Saturday.
The vector DB setup ate the first hour. Chunking and embeddings ate another. Then streaming broke because I mixed up the Anthropic streaming API with the OpenAI one Cursor initially reached for. By 10:30pm I had chat working but no history, no admin view, and the retrieval was picking wrong chunks half the time.
By 11:45pm I had 680 lines of code, three migration errors, and a database that wouldn't seed. The embeddings worked. The retrieval broke. I fixed retrieval. The UI broke. I fixed the UI. The embeddings broke again.
I killed it at 11:47pm and went to bed. Honestly, this was a skill issue — I knew four complex subsystems in one evening was too much and did it anyway.
Verdict: Cursor can build complex things, but "complex state at 10pm with tired eyes" is a combination no AI fixes. Sleep is a feature.
App 4 — AdSpotter (Sunday 9am–12pm, Cursor + Claude Code)
Chrome extension that highlights ads on any webpage so you can see exactly what's promoted vs organic. Not a blocker — an outliner. Because ad blindness is real and sometimes I want to see the system I'm swimming in.
Cursor wrote the Manifest V3 scaffolding and the content script skeleton. Then I handed the detection logic to Claude Code — looking for class names with "ad", "sponsored", "promoted", iframes from known ad networks, a few heuristic patterns. Claude Code was way better at the edge cases than Cursor was.
Settings popup took 40 minutes. Packaging and submitting to the Chrome Web Store took 30 more. By 12:03pm the .zip was uploaded and pending review.
This is the one I'd actually launch. It solves a real problem, has zero servers to maintain, and I'd use it myself. The codebase landed at 680 lines. Clean.
Verdict: Cursor + Claude Code together is the combo. Cursor for scaffolding, Claude Code for the logic that matters. Tight scope, single surface, no auth or billing — exactly the kind of thing these tools ship well.
App 5 — PingBucket (Sunday 2pm–5pm, v0 + Vercel)
A simple uptime monitor — give it three URLs and it pings them on a schedule, emails you when one dies. Nothing novel.
v0 handled the UI in about 20 minutes. The backend was a Vercel cron job on top of a Vercel Function. The whole thing took three hours including deploy.
It works. Nobody needs it. Five uptime monitors already do this better. I built it because I felt like I had to finish the experiment, not because the problem was worth solving.
Verdict: v0 is great for UI. But I shouldn't have built this. I knew before starting. Shipping the wrong thing is still shipping the wrong thing.
What I Actually Learned
Most of the takeaways aren't about tools. They're about me.
The 48-hour pressure killed more apps than the tools did. App 3 failed because I tried to build a three-day thing in four hours. If I'd picked a smaller scope, Cursor would've shipped it fine.
Combining tools beats picking one. The winner used Cursor AND Claude Code. The failures tried to stay inside one tool.
Lovable is magic until it isn't. When something breaks at the third layer of abstraction, you need to understand the stack. I'm not sure non-coders do.
Bolt's best feature isn't speed, it's the constraint. The browser-only environment forces you to scope small. That's a feature, not a bug.
Vibe tools have regression amnesia. Ask them to fix one thing, they'll quietly break five others. Commit every 30 minutes. When something works, stop touching it.
Saturday output beats Sunday output. I shipped two Saturday, one Sunday. Front-load the ambitious stuff.
$47 bought me an experience that would've cost thousands in agency time. Three apps I can show people, one I'd launch, and a mental model of which tool does what. Good trade.
The One I'd Actually Launch
AdSpotter. I used it for about an hour after I built it and found two sites I didn't realize were 80% sponsored content.
It wins because it solves a real, tiny, specific problem. Built-in distribution through the Chrome Web Store. No server costs. No auth. Small enough that one person can maintain it forever. That's the shape of a good vibe coding output.
I'm going to polish it this week and list it on the Chrome Web Store properly. If it works, I'll submit it to the claw.mobile showcase for a free backlink.
The lesson I keep learning: the best vibe coding output isn't the most impressive app. It's the one you'd actually use yourself.
Curious what your weekend project would cost with other tools? Try the build cost calculator →
The cheat sheet every vibe coder should have.
Which tool for which job, real pricing for all 17 platforms, prompt templates that work, and a decision tree. One page.
What you get
17 tools mapped to 12 use cases — zero guesswork
5 copy-paste prompt templates that actually work
Real pricing for every tool from free to $25/mo
Weekly digest: new tools, pricing changes, tactics
No spam. Unsubscribe any time.
Not sure which tool to use?
Compare Cursor, Bolt, Lovable, Claude Code, and more side by side.