Editorial correction — October 11, 2026: Astro is a web framework. The original account attributed the Tea breach to “basic Firebase vulnerabilities”; a supporting primary technical investigation is unavailable here, and the incident does not establish that AI-generated code caused the breach. The account of my own product and security concerns is preserved.
Published August 21st, 2025 • Based on Founder Reality Episode 6
Also available on: Apple Podcasts • Spotify • YouTube
Today I'm going to tell you about one of my most expensive mistakes this year - and why it might save your business from a similar disaster.
I fell hard for "vibe coding." If you haven't heard the term, it basically means using AI to code based on vibes rather than understanding. You know what you want, AI knows how to build it, but neither of you knows how it works.
Sounds amazing, right? That's what I thought too.
The $200/Month AI Coding Addiction
When Cursor first came out, I had that "holy shit, this changes everything" moment. I immediately jumped on the $200/month Claude plan - 20x more usage than normal users. I couldn't get my head out of it.
Within hours, I rebuilt SimpleDirect's entire marketing website using Claude Code. Learned Astro (a web framework I'd never used). The results looked professional, worked perfectly, and I felt like a coding genius.
That first hit was pure dopamine. I was hooked.
Project 1: The Mobile App Success Story
Our SimpleDirect mobile app had been dormant since 2021. We had new designs sitting in Figma for a year and a half - beautiful mockups that never saw daylight because we kept prioritizing other features.
With vibe coding, I fed Claude the Figma designs (screenshots, dev mode exports, the whole thing). Two days later? Complete app redesign. New look, modern interface, everything polished.
This project actually worked. It solved a real problem - bringing a dead feature back to life. I thought I'd found the holy grail of development.
I was wrong.
Project 2: The SimpleDirect Chat Nightmare
Riding high from the mobile app success, I decided to tackle SimpleDirect Chat - our internal AI team tool that had been collecting dust since 2022.
I ignored my dev team's advice and just started vibe coding. Two terminals open: Claude on the left for frontend, Claude on the right for backend.
Friends would give me feature requests, and I'd go home and build them the same night.
The velocity felt incredible. One evening I'd knock out 2-3 features. Within weeks, I had everything on the wishlist: file sharing, team channels, AI integration, custom workflows.
But here's what I didn't realize: I was creating a monster.
Hundreds of commits. Environment variables scattered everywhere. Markdown files in random directories. Architecture that made no sense. I was using GitHub Copilot to check Claude's work - AI checking AI's work.
When I asked my dev team to review the codebase, they went silent. Now I know why.
The repo was incomprehensible. Even to me. I had built a functioning product that I couldn't explain, maintain, or confidently release to customers.
Project 3: The Ghost Theme Disaster
The third project broke me completely.
The Handlebars theme I asked the model to change failed in my workflow. That is not a test of every model’s ability to work with Handlebars.
Claude confidently told me it had perfectly implemented my Figma designs. The reality? 72-point fonts everywhere. Broken margins. Massive white spaces. Text with no spacing. The site looked like a teenager's first HTML project.
I spent an entire evening trying to fix it. Hours on the subway, tweaking code I didn't understand, chasing problems I couldn't identify.
Finally, I gave up. The site was unusable, and I had no idea how to fix it.
That's when I realized: I was building on quicksand.
The Tea App Wake-Up Call
Right around this time, the Tea app got completely hacked. If you missed it, T was a popular dating app where women could anonymously report toxic men. It hit #1 on the App Store.
The hack was devastating. Driver's licenses stolen. Personal addresses leaked. Lawsuits incoming.
The breach does not, by itself, establish that AI-generated code caused the problem. My own concern was simpler: I had put a tool into production without understanding its security boundaries.
I could see my own projects in that story. SimpleDirect Chat, running in production, built with code I didn't understand. What if we got hacked? What if customer data was compromised?
The liability scared me. The embarrassment would be worse.
Why I Quit Vibe Coding
AI made prototyping available to me without understanding every generated change. That was useful, but it did not establish that I could safely maintain the resulting products.
But here's the thing: when everyone can do something, the industry changes completely.
The camera analogy made me wonder how easier creation changes competition. It did not prove that only the best photographers or developers could earn a living.
I expected commercial rewards to concentrate among a small number of builders. I did not have evidence for the precise 1–5%/90–95% forecast.
I thought skilled developers using AI well would have an advantage. I could not measure that advantage from these experiments.
The Market Concentration Reality
I expected crowded software markets to put pressure on similar products. That was a forecast, not proof that every category would choose a single winner.
How concentrated those markets become remains uncertain.
For the products I wanted to maintain, I needed people who could understand and review the system, not just generate more changes.
The changes I had generated became hard for us to explain and maintain. That does not prove every AI-assisted prototype has the same outcome.
What I'm Doing Now
I've basically stopped using Cursor and Claude for coding. Instead, I'm focusing on:
Editorial update, October 11, 2026: That was my August 2025 tool-use decision. It is not my current rule never to use these tools. The later coding-agent account and delegation reflection document changed usage and continuing limits.
Empowering my actual developers. At that point, I wanted developers to use AI while keeping review and quality standards. The 5–10× speed claim was not backed by a matched task measurement here.
Design and strategy. My time is better spent on problems that require human judgment, not code generation.
Planning and architecture. Making sure we build the right things, not just building things quickly.
The mobile app project taught me that vibe coding works great for:
- Prototypes and demos
- Reviving dormant features
- Quick MVPs for testing ideas
But it's dangerous for:
- Production applications
- Commercial products
- Anything involving user data
- Long-term codebases
For Developers Worried About AI
If you are a developer reading this, my failures do not establish what will happen to your job.
My team still needed engineering judgment and review. That need is different from a guarantee that every developer remains employed or works 5–10× faster.
I still value system design and security fundamentals. They are useful skills, not a guarantee of a developer’s future income or employment.
The Bigger Picture
This experience taught me something crucial about building in an AI-first world: when everyone can build complex features cheaply, simplicity becomes your competitive advantage.
For my products, I want to solve real problems with code we can understand and maintain.
I learned that more complexity did not automatically improve those products.
That's a lesson worth hundreds of thousands of dollars - even if I had to learn it the hard way.
What's your experience with AI coding tools? Have you run into similar issues, or found better approaches? Email me at [email protected] - I read every message.
New episodes of Founder Reality drop Monday, Wednesday, and Friday at 9am EST. Subscribe wherever you get podcasts, or visit founderreality.com
Editorial update, October 11, 2026: This was the schedule announced in the original episode, not a present guarantee of three weekly releases. Current reading is available through Founder Reality resources.
Daily thoughts: @TheGeorgePu on Twitter/X