book-to-skill: Compressing a technical book into a skill pack, and only reading the correct chapter when asking questions, saves 24-51 times the token than cramming the entire book into the contextbook-to-skill：把一本技术书压成技能包，提问时只读对的那一章，比把整本书塞进上下文省 24–51 倍 token

1. Point to it:/book-to-skill./my-book.pdf -- You can also give a folder, a glob, or a string of files.

II. It distills: Draws out the frame, decision-making rules, and anti-patterns, and writes out a file for each chapter. Note the structure, not the summary - the summary is compressed and the structure is retrievable.

Third, the agent loads on demand: after you knock/my-book replication, it reads the corresponding chapter and answers as it is, without you posting the book again.

SKILL.md: Core mental model + chapter index, about 4,000 tokens, which is the resident cost.

chapters/ch01- * .md…: one chapter, one file, about 1,000 tokens each, only loaded when asked, does not account for the skill budget.

glossary.md (glossary of terms, arranged alphabetically and labeled with chapter references), patterns.md (all tricks, algorithms, design patterns), cheatsheet.md (decision tables and cheatsheet rules).

By default, it falls in the cross-agent skills directory ~/.agents/skills<slug>//, Copilot CLI, Amp, Codex, OpenCode, OpenClaw can be found directly; under Claude Code, an additional attempt will be made to create a verified symbolic link to ~/.claude/skills/.

Formats: PDF, EPUB, DOCX, HTML, RTF, MOBI/AZW3 (walk Calibre), TXT/Markdown/reStructuredText/AsciiDoc (built-in, no dependencies).

PDF is divided into two ways: technical books (with code, tables, formulas) use docling, which can retain Markdown tables and code blocks, about 1.5 seconds per page; plain text books use pdftotext, which is almost instantaneous. Before the conversion begins, it will ask you what category the book belongs to, and then automatically select the tool.

The scanned version (full page is an image, no text layer) must first OCR: ocrmypdf input.pdf output.pdf, and then take the results to convert. It detects and stops on the first few pages, and it doesn't give you an empty skill by running through it.

A command checks the environment: python3 scripts/extract.py --check, which lists which tools are installed in each format, what is missing, and which installation command to run.

Internal documentation: architectural decision records, runbooks, onboarding guides - fold the entire docs/into one skill and ask as you write the code.

Branding and design systems: Tone guides, component principles, turning a 60-page branding manual into a team searchable skill.

Research cluster: a bunch of papers plus their own notes, merged into a skill, new material until it folds in (update/fold-in mode).

Specifications and Standards: RFCs, API Contracts, Compliance Documentation - those that you will revisit but will never remember.

One line: npx skills add virgiliojr94/book-to-skill; or manually git clone to the skills directory and register the/book-to-skill command.

The project itself does not contain any book content, it is just a converter; the extraction and analysis are all running on your machine, and the files are not uploaded. The generated skill is "your own reading notes", which is a derivative arrangement, and do not redistribute the copyrighted works of others.

After conversion, you can also publish to GitHub (default private) with one click, and other machines can be installed with npx skills add.

Only work with documents that you own or have permission to read; skill suggestions generated by third-party books remain private. This article compiles self-public information and accompanies supporting resources, and the function is subject to the official warehouse.