Claude Code connects excerpts across 100 books

Pieter Maes wanted an agent to help him read across books rather than reduce them to summaries. He indexed 100 nonfiction books in SQLite, using topic extraction, embeddings and a hierarchical topic tree. Claude Code received command-line tools for exploring those topics and retrieving relevant passages.

The agent worked in stages. It scanned the library and existing trails, proposed unexplored ideas, then followed a chosen theme through many text chunks. It selected and ordered excerpts before adding highlights and connections between consecutive passages. Maes published the resulting trails, including one connecting Steve Jobs, Theranos, startup cultures and mass movements.

Maes had first built a pipeline of discrete model calls, but found that letting Claude choose its own tool sequence produced more interesting results with less orchestration. When he changed his preference for excerpt length, the agent revised existing trails while preserving their meaning. It also suggested missing tool functions, which helped him improve the browsing interface.

The quality of a connection remained a human judgment. Maes notes that topic groups could be strange and that Claude gravitated toward secrecy and hidden systems. The reported roughly £10 covered preprocessing the book collection, not the total cost of every agent run. Published trails demonstrate the artifact; they do not establish that all suggested connections are useful or faithful interpretations.

Prompt
Original prompt

Source: Blog · Pieter Maes