I've been an iOS developer for 10 years, but I had never built my own app. While on parental leave, I finally built one. But I never wrote a single line of code myself.
Why I decided not to read the code
From the start, I decided I wouldn't read a single line of code. Because I knew I couldn't verify AI-generated code. After 10 years of working mostly with WebView, I couldn't tell if SwiftUI code was right or wrong.
So I decided to verify the results instead of the code.
The Korean particle problem
Korean has grammatical particles that change depending on the word before them. For example, 나는 and 내가 mean the same thing, but use different particles.
나 + 는 → 나는 ("I" + topic particle)
내 + 가 → 내가 ("I" + subject particle)
So when comparing what the user said with the saved affirmation, even a single different particle made it a different sentence.
At first, I told the AI to hardcode Korean particle handling. But the number of edge cases exploded. That's when I changed my approach: anything above an 85–90% similarity score would pass. So I switched to similarity scoring instead of exact matching.
Making this decision didn't require a single line of code.
Letting the AI verify itself
If I can't verify the code myself, why not let the AI do it? I deliberately fed it wrong inputs and tested whether it could catch them.
What I found in the logs
When I opened the oldest log on my MacBook, it wasn't my words — it was messages exchanged between AI agents. The first thing I actually typed was: "You are the orchestrator of this repo."
The plan is to add on-device AI to this app.
The app I built is AffirmCount — App Store · Google Play
Top comments (0)