下载地址
https://www.zhengdianxue.com/wp-content/uploads/skills/user_5fc3c48c--book-deep-reading-skill-compiler.zip将提示词发送给你的 AI 安装该 skills
请根据 https://skillhub.cn/install/skillhub.md，安装 @user_5fc3c48c/book-deep-reading-skill-compiler。把书籍和长篇资料编译为可追溯、可验证、可复用的 AI Skill 与知识包。Use when users ask for 深度阅读、拆书、提炼方法、跨书综合、把书变成 Skill/book2skill, or convert PDF/EPUB/DOCX/HTML/TXT/Markdown sources into operational frameworks with source anchors, evidence gates, pressure tests, boundaries, and lineage; not for casual plot summaries.
阅读全文 （展开查看完整说明）
概述
Book to Skill（拆书Skill）
Turn a book from passive reading material into a reusable method system. Preserve source traceability, reject weak ideas, merge repeated knowledge, and create a Skill Pack only when the method is stable enough.
中文介绍
这不是“把一本书压缩成几页摘要”，而是把长篇内容编译成一套能被 AI Agent 稳定调用、能追到原文证据、也经得起反例检验的方法系统。
- 适合谁：需要深度拆书、建立主题知识库、融合多本书观点，或把成熟方法发布成 Skill 的研究者和团队。
- 解决什么：避免摘抄堆积、概念重复、来源丢失，以及未经验证的观点直接进入知识库。
- 标准产出：研究问题、来源地图、候选方法单元、证据锚点、压力测试、知识合并记录和可安装 Skill Pack。
- 核心价值：不仅告诉你“书里说了什么”，还回答“哪些方法值得复用、适用边界在哪里、遇到什么证据应当推翻”。
Operating contract
- Classify the request as
note-pack,knowledge-pack,skill-pack,cluster-merge, orcross-book-synthesis. - Confirm source rights and confidentiality before extraction. Keep raw copyrighted text local by default.
- Move through the seven modules in order. Do not jump from raw text directly to a skill.
- Use research questions to decide reading depth unless the user explicitly requests a full read.
- Keep a source anchor for every reusable claim. Quote only the minimum text required for traceability.
- Extract only supported units across five lanes: framework, principle, case, counter-example, and glossary.
- Verify every candidate. Mark it
verified,rejected, orneeds_review; never leave it unclassified. - Search existing concepts and aliases before creating a new concept. Preserve genuine disagreements.
- Create an L4 Skill Pack only when triggers, steps, boundaries, and self-tests are clear.
- End with files created or updated, gates passed or failed, rejected-unit count, merge decisions, and open questions.
Choose the output level
| User intent | Stop after | Required result |
|---|---|---|
| Understand one book deeply | Module 4 | Note Pack |
| Build reusable knowledge | Module 6 | Knowledge Pack |
| Turn stable methods into an AI skill | Module 7 | Skill Pack |
| Add a book to an existing topic | Module 6, then reassess | Updated concepts and synthesis |
| Compare several books | Module 6 | Cross-book synthesis with divergences |
If the user has not supplied the source yet, produce an execution plan and input checklist; do not fabricate book content.
Seven-module workflow
| Module | Decision | Core outputs | Gate |
|---|---|---|---|
| 1. Problem Lens | What real problem should the reading solve? | intake.yaml, research-questions.md | questions and output level are explicit |
| 2. Source Map | Is the source complete, readable, and safe to process? | local full-text, headings.md, metadata.json | extraction and rights are recorded |
| 3. Depth Router | Which parts deserve full, targeted, skim, or no reading? | relevance.json, reading route | depth has evidence |
| 4. Method Foundry | What reusable methods does the book actually support? | single-book-note.md, candidate-units.jsonl | anchors and boundaries exist |
| 5. Evidence Gates | Which ideas survive transfer, prediction, and uniqueness tests? | verified-units.jsonl, pressure-tests.json | every candidate is classified |
| 6. Knowledge Merge | What should support, refine, contradict, or extend existing knowledge? | concepts, synthesis.md, merge log | no silent duplication or flattened conflict |
| 7. Action Pack | Can the stable method be called reliably by an AI agent? | installable skill, lineage, self-tests | triggers, steps, boundaries, tests pass |
Read references/workflow.md before running Modules 1–7. Read references/data-contracts.md before creating structured artifacts and references/quality-gates.md before declaring a module complete.
For strategy, competitive advantage, planning, management operating systems, or resource-allocation books, also read references/strategy-profile.md.
Execution rules
- Create a dedicated run directory under the user-approved workspace. Use relative paths in durable artifacts.
- Keep raw extraction in a local scratch subdirectory excluded from shareable outputs.
- Prefer structural anchors such as chapter, section, page, heading, or paragraph ID.
- Mark OCR uncertainty, missing chapters, broken tables, or parser degradation explicitly.
- Treat author anecdotes as examples, not proof of universality.
- Reject generic advice that lacks operational form, predictive value, or a clear boundary.
- Preserve the author's distinctive concept only when it adds explanatory or decision value.
- Use exactly one merge tag per contribution:
SUPPORTS,REFINES,CONTRADICTS,NEW_SUB_ASPECT, orNEW_CONCEPT. - Never overwrite an existing synthesis when a contribution is tagged
CONTRADICTS; create a review item. - Keep fiction, memoir, and narrative-heavy books at Note Pack unless the user explicitly requests a method extraction and the text supports it.
Output assets
Copy and adapt only the templates needed for the selected output level:
assets/templates/intake.yaml.templateassets/templates/single-book-note.md.templateassets/templates/candidate-units.jsonl.templateassets/templates/verified-units.jsonl.templateassets/templates/pressure-tests.json.templateassets/templates/synthesis.md.templateassets/templates/skill-lineage.yaml.template
Do not copy placeholder values into final artifacts.
Validation
Before handoff:
- Apply the relevant checklist in
references/quality-gates.md. - Run the behavioral prompts in
references/self-test.mdfor a Skill Pack. - Run the available skill-folder validator on the generated skill.
- Scan shareable outputs for long copyrighted passages, credentials, private IDs, absolute local paths, and unresolved placeholders.
- Report any unrun test or unresolved review item instead of presenting it as complete.
