An honest, no-hype look at what happens when you let AI build your whole app — and then try to refactor it.
Photo by Steve A Johnson on Unsplash
TL;DR
- I spent 30 days vibe coding a web app from logo to launch using only AI — no manual code.
- Total spend: ~$1.45 in Claude API tokens, plus two full Claude Code sessions in one night.
- The app worked. The codebase didn't. I sunset the project after the AI couldn't understand its own code.
A few months back, I started vibe coding to build a web application that helps users practice hard conversations — things like asking for a raise, prepping for an interview, or having difficult talks at work.
To really test the vibe coding craze, I used AI for everything from the logo to the final publish. Here's what I learned, honestly and from my own perspective.
If this story caught your interest but you don't have a Medium membership, don't worry — I've got you covered. [Non-member link].
The Experiment: Building a Real App With Zero Manual Code
I decided to build a web application with no login and no paywall. Users could just jump in and test how it works. No database to save user information — all conversations and decisions are handled by Anthropic's API.
And yes, before you scream it at me, I'll admit it: just another AI wrapper. :)
I used Next.js paired with Vite — an obvious choice from Claude. React.js and Angular were on the table, but I went with Next since Claude recommended it.
Here's the mindmap of how it works from an end-user POV:
Mindmup.com
The questions are AI-generated. It's essentially a conversation between you and the AI.
The Budget: How Cheap Vibe Coding Actually Is
My budget was very low — like, super low. I spent $5 on Claude API tokens, and I shared those same tokens across a few other small projects too.
For the AI voice, I just used the browser's native text-to-speech. It's a bit robotic, but it does the job.
How I Actually Used Claude Code
I spent one full night on this experiment across two complete Claude Code sessions (5 hours + 4 hours), staring at the limit screen waiting for the 5-hour reset and compacting context here and there.
More than 80% of my prompts were written by me and polished with Claude Chat. I never wrote a vague prompt — not once.
Sample Prompt:
Act as a Senior Full-Stack Developer and UI/UX expert.
We are going to build a [describe the app/feature, e.g., real-time cryptocurrency dashboard].
Tech Stack: [e.g., React, Tailwind CSS, and Node.js].
UX/UI Style: [e.g., A minimalist, neon-dark mode aesthetic with glassmorphism elements, inspired by cybercore design].
Key Features:
1. [Feature 1, e.g., Live ticker fetching from an API]
2. [Feature 2, e.g., Interactive drag-and-drop dashboard widgets]
3. [Feature 3, e.g., Responsive layout for mobile and desktop]
Execution Guidelines:
1. First, analyze the requirements and output a step-by-step implementation plan.
2. Wait for my approval before proceeding with the code.
3. For every file you create or modify, ensure the code follows SOLID principles and includes structured, typed error handling.
4. Before finishing your response, verify your code against my initial goals.
Let's start by outlining the implementation plan.
One user session with Claude Sonnet 4.6 cost me roughly $0.05. So in the testing phase alone, I burned about $0.95. Real users — mostly friends — burned another $0.50.
Day 1: Reading the Code I Didn't Write
Everything seemed fine until I sat down to actually read the code.
On day one, I started noticing Claude's architectural choices. It was a Vite build alongside Next.js. I checked the README and package.json — there was a Next.js version mismatch.
It used Tailwind for styling, but there were a huge number of classes, many of them completely unnecessary. On top of that, it wrote custom CSS in the stylesheet for the same things it had already styled with Tailwind.
I opened one component and found it stuffed with redundant code — almost jargon for a simple task. Something that should've taken 10 lines was written in 45.
Redundant classes were everywhere, and it leaned on absolute positioning instead of using proper media queries.
Where the Refactor Fell Apart
Since I couldn't understand what the AI had written, I started refactoring it with the same AI. Even with context compacting and a CLAUDE.md file in place, it lost the thread after a while. It started overthinking, made the code more complicated, and turned it even less readable.
If I didn't know what it wrote, and it couldn't keep the code consistent — and didn't even know what it was writing — there was no path forward.
Thanks to hallucinations, I now understood nothing in the codebase. I decided to stop further production and sunset the project to move on to something else.
The Verdict: Is Vibe Coding Worth It?
I'm not completely against vibe coding — it's just not my cup of tea. I still use AI as a helping tool, not to build complete web applications.
Here's where AI actually earns its keep in my workflow:
- Dataset generation for testing projects — it's great at cooking things up.
- Error finding — then I cross-verify with Stack Overflow and Reddit communities before implementing fixes.
- Boilerplate and breath-taking low-effort work — the stuff that drains hours without adding value.
These tools aren't copilots. They help you here and there — but they won't build you a cool web application that lasts for years.
If you've tried vibe coding yourself, I'd love to hear how it went. Drop a comment, and if this saved you from burning $5 the way I did, a like wouldn't hurt either.
