According to the introduction, Book to Skill's on-demand loading approach can save 24 to 51 times the tokens compared to stuffing an entire book into the context.