Never Trust LLM Output — 6 Defenses from Building a PDF-to-Anki CLI

Introduction In the previous article , I wrote about my first 10 days of real development in an Everything Claude Code (ECC) environment. It was the story of a beginner who didn't even know git, learning development by running the PDCA cycle over and over. This time, I'll talk about pdf2anki — a tool I built during the second half of those 10 days. The Problem Was Content, Not the App To study for the G-Kentei exam, I built a quiz app based on spaced repetition. A web version in Python/Streamlit, an iOS version in Swift/SwiftUI. I implemented them with TDD in the ECC environment, recording design decisions in ADRs, and had working software in two weeks. After it was done, I realized: I had reinvented Anki. There's no point spending two weeks rebuilding something that Anki has refined over 20+ years. But during development, the real bottleneck became clear. It wasn't the app — it was the content . In the iOS version, I implemented an algorithm that extracted 410 Q&A pairs from PDF text