书到技能 · 把技术书/PDF 蒸馏成 AI 技能

【知识技能化】把技术书/PDF/白皮书/设计规范/运营手册蒸馏成可导入的 AI 技能（知识包）——不是 RAG 检索原文，而是把书里方法论内化为决策规则、模板、心智模型，让 AI 边干活边套用专业方法。适合：把《Clean Code》变代码审查 Skill、把品牌规范变审图 Skill、把投资问答录变投研 Skill。当用户说"把这本书变成 skill""把 PDF 转成技能""知识即服务"时使用。

Book-to-Skill

Turn a technical book (or any structured document) into an importable Skill: a distilled, procedural knowledge pack an agent loads on demand so it applies the book's methods while working — not merely retrieves its text.

Requirements

• pdfplumber for Step 1 (PDF parsing): pip install pdfplumber. Step 2–4 (distilling + assembling) need no extra deps.
• Works best on text PDFs. If metadata.json reports scanned_pages, OCR the PDF first — image-only files have no extractable text.

Why this is not RAG

• RAG = retrieve text chunks to stuff into context. Skill = load distilled procedure + curated references so the agent acts per the book's methodology.
• The output is a portable skill folder another agent can import and trigger by its description. That portability is the product.

Workflow (run in order)

1. Ingest & scaffold

Parse the PDF into a navigable structure:

bash

python scripts/extract.py --pdf "<path-to.pdf>" --out "<output-dir>" --name "<skill-slug>"

This produces in <output-dir>:

• structure.md — table of contents (section #, title, page range, size)
• sections/sec_NNN.md — per-chapter raw extracted text
• metadata.json — page count, font stats, likely_scanned flag

If metadata.json shows "likely_scanned": true, STOP and tell the user to OCR the PDF first (image-only file → no extractable text).

2. Distill (the value step — never skip)

Read structure.md, then load the relevant sections/*.md and rewrite each into procedural knowledge following references/distill-guide.md:

• Convert "what the book says" → "how to do X" checklists, decision rules, anti-patterns.
• Extract reusable templates, code patterns, naming/schema rules.
• Capture the book's mental model, not its prose.

3. Assemble the generated Skill

text

<skill-slug>/
├── SKILL.md            # name + description (trigger text) + how to apply the book
└── references/
    └── <book>.md       # distilled, sectioned knowledge (the curated pack)

• SKILL.md description must list concrete triggers ("when implementing X from …").
• Keep SKILL.md lean (<5k words); put detail in references/.
• For deterministic deep lookups, ship scripts/search_knowledge.py pointing at references/.

4. Verify & package

• Spot-check that references/<book>.md covers the high-value sections from structure.md.
• Tell the user to import via the Skills UI → 上传技能 (or run the WorkBuddy skill packager). Do not claim auto-publish.

Notes

• Deterministic retrieval while an agent uses the generated skill: python scripts/search_knowledge.py --base references --query "topic" --top 5
• Very large books: split references/ by part; keep it one level deep from SKILL.md.
• This skill creates the scaffold + distilling method; the actual distillation of a specific book is done by the agent at use time (or by a human reviewer for quality-gated packs).