Why You Shouldn't Let AI Do Your Fortune Telling — And How to Do It Right

The first lesson I learned building a saju (Korean four-pillar fortune telling) app: it's not about what you ask AI to do — it's about what you don't . Revenue is still $0. This isn't a success story — it's a debugging diary. "Tell me the fortune for someone born March 15, 1990." I threw this straight at an LLM. The response looked great. Smooth sentences, Five Elements this, Wood-Fire-Earth-Metal-Water that. But there was a problem. The base calculations were wrong. Me: "Analyze the Four Pillars for March 15, 1990, 6 AM" Claude: "The year pillar is Geng-Wu, month is Ji-Mao..." Me: "...Ji-Mao is wrong." LLMs can't do manseryeok (traditional Korean astronomical calendar) calculations. More precisely, they appear to get it right probabilistically, but they don't actually compute anything. The Heavenly Stems and Earthly Branches follow a 60-cycle system. Month pillars shift at solar term boundaries. Hour pillars depend on the day's Heavenly Stem. This isn't reasoning — it's arithmetic. Wh