What broke when I used Codex to maintain a multilingual recipe site

I've been using Codex heavily while building PashutFree, a multilingual recipe site with dietary and allergy filters.

The biggest thing I've learned is that AI is much better at making a local change than understanding all the places that change can affect.

A small change to one part of the site can touch recipe pages, different languages, filters, metadata and routing. The code can compile and everything can look fine, while a small number of pages are still wrong.

What has worked better for me is using Codex as the implementation layer, but keeping the important rules deterministic. I try to give it narrow tasks, make it inspect the existing implementation first, then run checks against the things that should never change.

I've also become much more careful about things like allergen data. I don't want an LLM deciding whether something is safe just because the ingredient name looks familiar. That information needs to come from structured data and explicit rules.

So for me the interesting part of AI coding has moved from "how much code can it write?" to "how do I build enough checks around it that I can trust changes across thousands of pages?"

I'm the builder of the project. This is the live site if anyone wants context: https://pashutfree.com/en-us⁠�

For people using coding agents on content-heavy projects, have you found the bigger problem is generating the code or keeping everything consistent as the project grows?