book-to-skill: Hallucination-free answers with 24–51x fewer tokens than full-book context

When given technical documents in various formats such as PDF, EPUB, and Markdown, it generates skills composed of frameworks, decision rules, anti-patterns, and per-chapter files. It creates a structured knowledge system rather than a simple summary, designed so that agents load specific chapters only when needed. This enables accurate answers without placing the entire book in the context.

Previous methods, such as PDF search or asking agents directly about book content, often returned only page lists or caused hallucinations. This tool pays the structuring cost only once at conversion time, providing accurate information based on the actual book content without hallucinations. Measurements show that token usage is 24–51 times lower than dumping the entire book into the context.

It works identically on all hosts supporting the Agent Skills standard, including GitHub Copilot CLI, Amp, and Claude Code. Based on the SKILL.md format, it is not tied to any specific platform. Any structured text referenced repeatedly, such as internal documentation, brand guidelines, or research paper collections, can be converted.

It is useful for managing technical books that are easily forgotten after reading, as well as architecture decision records, runbooks, and API specs that need frequent reference during coding. Extraction and analysis are performed locally, so files are not uploaded externally, and it is released under the MIT license. It consists of a Python-based extractor and a spec-based generator, offering high extensibility.