book-to-skill compiles PDFs into 4,000-token skill files

book-to-skill compiles a PDF into a roughly 4,000-token skill file that an AI agent loads instead of the source document. The project's benchmarks claim 24 to 51 times fewer tokens than pasting the whole book into context, which can run 119,000 to 256,000 tokens. For an agent that needs deep knowledge of one dense document, that's the difference between burning most of a context window on background reading. Instead, it runs a handful of on-demand lookups. The repo has climbed to 14,194 GitHub stars, up 595 in a day.

The split works by keeping a compact index resident and loading roughly 1,000-token per-chapter files only when the agent needs that chapter. Conversion runs through Docling, which parses technical PDFs at about 1.5 seconds a page, for a total cost near a dollar per book.

Three more repos, same bet

cangjie-skill applies the same idea to long videos and podcasts, not just books, and sits at 5,797 stars after gaining 363 in a day. reverse-skill has 10,379 stars, up 612. last30days-skill is the biggest of the four at 56,055 stars, up 660. All four compile knowledge once, offline, and spend tokens only on the piece a query actually needs.

None of the four repos say what happens when the source document changes after the skill gets compiled. For a book, that's rarely an issue. For anything that updates, that gap matters.