For the last eight months I've been running a local Claude Code skill called ORACLE PRIME — a weekly competitive intelligence scanner that reads 40+ sources, identifies market shifts, and produces a structured briefing. It worked well enough. But it had no way to know when it produced a mediocre briefing versus an exceptional one. The output quality varied by 40% depending on which sources were available and what Claude happened to focus on. There was no feedback loop, no quality floor, no way t...