Title: I built a 13-app 'Zoo' using Gemini Pro 3. The constraint: I wasn't allowed to inspect the code.

Published: (December 14, 2025 at 03:38 PM EST)
2 min read
Source: Dev.to

Source: Dev.to

Introduction

I recently decided to test the absolute limits of “Vibe Coding.” I wanted to know: can I build a robust platform of micro‑applications without ever acting like a traditional engineer? To find out, I built the AppZoo.

Rules of the Experiment

  • Zero manual code inspection – I was not allowed to read any generated code.
  • If Gemini wrote a React component, I couldn’t audit the syntax.
  • If Copilot suggested a Terraform configuration, I had to deploy it as‑is.
  • My role was limited to “Product Manager” and “QA.” I could only verify behavior, not implementation.

Blind Trust Methodology

Logic Check (Math)

For the Dual Fuel Calculator, I first built a model in a spreadsheet to know the expected answers. I then fed the requirements to Gemini. The AI not only matched my spreadsheet but refined the logic, handling temperature cutoff points with better precision than my manual model. I verified the output and trusted the code to get there.

Vibe Check (Visuals)

For the Fractal Explorer and the Video‑to‑GIF Converter, the test plan was experiential. If the fractal zoomed smoothly or the GIF rendered quickly, the app passed.

Architecture: Sandbox Security

Running uninspected AI code is risky, so I mitigated it by having the AI build a Domain Isolation Model via Terraform:

  • Every micro‑app lives on its own subdomain (e.g., zork.appzoo..., dualfuel.appzoo...).
  • This leverages the browser’s Same‑Origin Policy, preventing a compromised app from accessing cookies or localStorage of the main site.
  • The entire “Zoo” is 100 % client‑side; there is no backend to hack.

Friction Points

The AI frequently got stuck in “circular reasoning” loops, proposing the same broken fix three times in a row.

Fix

I discovered that model switching is the best debugger. When Gemini Pro got stuck, I switched Copilot to a different underlying model, which snapped it out of the loop.

Result

The outcome is a collection of 13 working apps, ranging from practical tools (Heat Pump cost analysis) to pure toys (Infinite Zork, 3D CSS Studios). It’s messy and still has bugs I haven’t found yet, but it exists.

You can explore the Zoo here:

Back to Blog

Related posts

Read more »