Booksmith: Forging Skills from Books

The problem it addresses

• It turns a methodology book or long document into a reusable, callable skill instead of producing a one-off summary.
• It handles the messy parts of real-world Chinese books: txt/md, docx, epub/mobi, text PDFs, scanned PDFs, inconsistent encodings, missing Word heading styles, double-column layout, and OCR noise.
• It controls context pressure by reading strategically rather than dumping the whole book into the prompt, and it reduces contamination from cases, quotes, and invented concepts.

How it works

• Read and clean: Prefer text editions; detect and normalize encodings for txt/md; convert epub/mobi to Markdown; fall back through python-docx, docx2txt, or pandoc for DOCX; detect scanned PDFs and OCR them with page-wise write-back when needed.
• Layered reading: Scan the table of contents first, then read chapter openings, endings, bold text, and subheadings; deep-read only the chapters that carry methods, frameworks, terms, or decisive examples.
• Model extraction: Find author-named frameworks, verify their definitions against the source text, and map hierarchy, process, and support relationships; mark implicit models as “extended interpretation”; then fill in core themes, terms, methods, thinking patterns, steps, and use cases.
• Skill generation: Produce an English slug, a trigger-rich description, an appropriate template, a term-normalization table, source references, and a self-check list before saving to the user-level skill directory.

Boundaries and notes

• Use it when the goal is “turn this book into a skill”; if the user only wants to read, summarize, or ask questions, confirm intent first.
• WeChat Reading can only serve as an optional accelerator for chapter metadata or user highlights; it cannot provide the full body text.
• Scanned PDFs are more vulnerable to OCR quality issues and cross-column misreads; a text edition is usually safer when available.
• Output must be traceable: key terms are checked against the source, uncertain items are marked “needs verification,” and no references, cases, or data are invented.