A Three-Month Claude Code Pipeline Read 1,189 One Piece Chapters by Vision and Filed 108 Falsifiable Predictions Before the Reveals
A builder published grandlinearchives.com after running a dozen-agent Claude Code pipeline — capture agent, three-pass vision readers per chapter, a synthesizer serializing edits into an Obsidian vault, plus etymology, visual-mirror, and prediction-scoring specialists, with SQLite for state and self-resuming batch passes. Output: 3,073 foreshadowing entries with chapter citations on both the setup and payoff ends, 1,488 cross-linked character pages, 41 tracked open mysteries, and a Theory Gauntlet running 16 community theories through steelman → evidence sweep → adversarial refutation. The part worth stealing is the prediction ledger: falsifiable claims filed and timestamped with the chapter the system knew about at the time, scored as new chapters ship — currently 108 filed, 28 confirmed, 23 partial, 4 refuted, and 85% in the highest-confidence bucket. That's a working self-evaluation harness for a long-running agent, not a demo.
↳ Follow the thread