January 16, 2026
Bibliobots vs Bookworms
Reading across books with Claude Code
AI stitches book-to-book 'idea trails' — commenters split: genius guide or reading cheat
TLDR: An AI agent built “idea trails” across 100 books, impressing the author but igniting a brawl. Commenters split between excitement and eye-rolling, with critics warning AI can’t grasp context and purists saying “just read the book,” making this a flashpoint for how we’ll study big ideas next.
Claude Code, an AI assistant, skimmed 100 nonfiction books and stitched "trails" of related excerpts — including a spicy leap from Steve Jobs to Theranos — by chunking text and mapping a giant topic tree. The author says the bot outperformed a hand-built workflow, even re-editing trails on request. The taxonomy is wild, with quirky top topics like “Names beginning with ‘Da’,” yet it’s “good enough” for AI to map ideas. Cue the bookworms vs bots brawl.
On Hacker News, the repost police quickly popped in with an earlier thread link (here), while curious types wanted more on the topic tree — one even confessed, “Whoops, found the details at the end.” But the loudest energy came from eye-reading purists calling the project a "huge waste" and "LLM scam tech," plus a vibe that the cool kids of the future will be the ones still flipping paper pages. The semantic sticklers delivered the toughest critique: meaning shifts by context, so AI might miss truly novel connections. Jokes about “reading with your eyes” turned into the day’s meme, as the crowd split between AI librarians and analog loyalists. Verdict: thrilling idea-miner or glittering shortcut — depends which camp you’re in.
Key Points
- •A 100-book non-fiction library was processed into ~500-word chunks, indexed by topic for search and exploration.
- •Over 100,000 topics were extracted and organized into a hierarchical topic tree of about 1,000 top-level topics.
- •Claude navigates the corpus via CLI tools to find similar topics, co-occurrences, and sibling relationships across books.
- •Trail generation occurs in stages: propose ideas, build trails from seed topics by ordering excerpts, then add highlights and edges.
- •Using Claude Code as an agent with minimal prompting outperformed a hand-tuned pipeline and enabled nuanced corpus-wide edits.