The Day I Told Claude Code 'You ARE the LLM'

Auto-generating structured explanations for 397 questions — and the quality was catastrophic I'm building an iOS app for a certification exam as a solo project. About 400 questions across 8 categories. I can't share the actual exam content, so for this article, imagine the questions are about Baki the Grappler — a martial arts manga. All example content below uses anonymized Baki-themed placeholders. The app started as a simple spaced repetition (FSRS) quiz, but I wanted richer explanation UI. Instead of plain text explanations, I decided to implement a 4-layer "structured explanation" format. public struct EnhancedExplanation : Codable , Equatable , Sendable { public let correctSummary : String // Layer 1: Answer summary (≤200 chars) public let contrastTable : [ ContrastEntry ] // Layer 2: Per-choice judgment + explanation public let keyPhrases : [ String ] // Layer 3: 3-6 technical terms public let relatedConcepts : String ? // Layer 4: Related concepts (optional) } I started by hand