The article describes a system where an agentic Claude Code setup explores a corpus of 100 non‑fiction books to build “trails” of thematically linked excerpts instead of simple summaries.

Books are chunked, embedded, clustered into a hierarchical topic tree, and explored via CLI tools that let Claude propose trail ideas, assemble ordered excerpt sequences, and annotate connections. The author reports that treating Claude as a coworker with tools, asking it what capabilities it needs, and biasing search toward novel topics enables effective syntopic reading, with a relatively simple stack (EPUB parsing, SQLite + vector search, Gemini for topic extraction, Claude for trail generation).