| name | book-to-skills |
| description | Use when the user wants to turn a book into reusable skills, methodology documents, or a book-derived thinking system for Claude Code, OpenClaw, or similar agent environments. Trigger on phrases like "从书生成skill" or "book to skill" or when user wants to extract methodologies from a book. |
Book to Skills
Overview
Transform a book into reusable, evidence-based skills that an agent can actually call. This skill enforces systematic extraction—not summary generation—to produce a structured bundle including author thinking models, methodology cards, routing rules, and execution-ready Markdown outputs.
Core principle: Every conclusion must have evidence from the book. No guesswork, no relying on training data about the book.
When to Use
Use this skill when:
- User says "帮我从《书名》生成skill"
- User wants to extract methodologies, frameworks, or thinking patterns from a book
- User needs a book-derived thinking system for agent environments
- User wants to convert book content into reusable MethodCards
Do NOT use when:
- User just wants a book summary or review
- User wants subjective interpretation without evidence
- The book has no clear methodology or framework
Decision Flowchart
digraph usage_decision {
"User mentions book + skill/methodology?" [shape=diamond];
"User wants summary only?" [shape=diamond];
"Book has extractable methods?" [shape=diamond];
"Use book-to-skills" [shape=box,style=filled,fillcolor=lightgreen];
"Refuse or redirect" [shape=box,style=filled,fillcolor=lightcoral];
"Proceed with summary" [shape=box,style=filled,fillcolor=lightyellow];
"User mentions book + skill/methodology?" -> "User wants summary only?" [label="yes"];
"User wants summary only?" -> "Proceed with summary" [label="yes"];
"User wants summary only?" -> "Book has extractable methods?" [label="no"];
"Book has extractable methods?" -> "Use book-to-skills" [label="yes"];
"Book has extractable methods?" -> "Refuse or redirect" [label="no"];
}
Outcomes
This skill produces a structured bundle that helps an agent answer:
- Should this book's thinking style be applied to the user's problem?
- Which methodology modules from the book are relevant?
- Which subagents should be assigned?
- Which Markdown documents should be produced?
Red Flags - STOP and Reassess
- Thinking "I know this book" → STOP. Use NotebookLM to query the actual content, not your training data.
- Creating content without evidence → STOP. Every MethodCard needs evidence references.
- Skipping the five-stage query process → STOP. This ensures systematic extraction.
- Producing just a summary → STOP. This skill creates executable skills, not summaries.
Required Inputs
Collect or infer a minimal BookBrief:
title: string
author: string
domain: string
audience: string
goal: string
language: zh | en
If some fields are missing, continue with reasonable assumptions and record them in the output.
Default Workflow
Phase 1: Setup (Sequential)
- Create NotebookLM notebook for the selected book
- Verify source processing before proceeding
Phase 2: Parallel Extraction (Parallel Execution)
Execute these packs in parallel where dependencies allow:
┌─ Pack 1: 全书结构图 ─┐
│ (Book Profiler) │
│ ↓ │
│ 00-overview.md │
└──────────────────────┘
↘
┌─ Pack 3: 方法论抽取 ─┐
│ (Methodology Miner) │
│ ↓ │
│ MethodCards[] │
└──────────────────────┘
↗ ↘
┌─ Pack 2: 作者思维 ───┐ ┌─ Pack 4: 场景路由 ─┐
│(Author Thinking │ │ (Scenario Router) │
│ Analyst) │ │ ↓ │
│ ↓ │ │ 03-scenario-router │
│ 01-author-thinking │ └────────────────────┘
└──────────────────────┘ ↘
↘
┌─ Pack 5: 冲突校验与综合 ─────────────────────┐
│ (Skill Packager / QA) │
│ ↓ │
│ Final Skill Bundle │
└─────────────────────────────────────────────┘
Parallel Rules:
- Pack 1 & Pack 2: Execute in parallel (no dependencies)
- Pack 3: Starts after Pack 1 AND Pack 2 complete
- Pack 4: Starts after Pack 3 completes
- Pack 5: Starts after Pack 4 completes, validates all outputs
Phase 3: Packaging (Sequential)
- Assign subagents using references/subagent-playbooks.md
- Package into Markdown output bundle per references/output-schemas.md
- Prepare infographic structure (optional) using references/visual-infographic-spec.md
Parallel Agent Orchestration
Agent Dependency Graph
| Agent | Dependencies | Parallel Groups |
|---|
| Book Profiler | None | Group A |
| Author Thinking Analyst | None | Group A |
| Methodology Miner | Pack 1, Pack 2 outputs | Group B |
| Scenario Router | MethodCards[] | Group C |
| Skill Packager / QA | All outputs | Group D (final) |
Execution Protocol
-
Launch Group A agents simultaneously
- Both agents query NotebookLM independently
- Outputs:
00-overview.md, 01-author-thinking.md
-
Merge Group A outputs before Group B
- Combine book structure + author thinking model
- Pass to Methodology Miner as unified context
-
Group B: Methodology extraction
- Can split into 2 parallel sub-agents:
- Sub-agent B1: Extract explicit methods (named frameworks)
- Sub-agent B2: Extract implicit methods (recurring patterns)
- Merge results into single
MethodCard[]
-
Group C: Routing logic
- Single agent, processes all MethodCards
- Output:
03-scenario-router.md
-
Group D: Validation and packaging
- Parallel validation tasks:
- Validate evidence completeness per MethodCard
- Check cross-references between documents
- Verify routing logic coverage
- Final packaging into skill bundle
MethodCard Schema
Every extracted method must be documented as:
name: string
thesis: string
triggers:
- string
steps:
- string
scenarios:
- string
limits:
- string
evidence:
- "Chapter X, Section Y: specific quote or reference"
assigned_agents:
- string
Subagent System
Use the following fixed roles:
| Role | Primary Duty | Key Output |
|---|
Book Profiler | Build book map, structure, core issues | 00-overview.md |
Author Thinking Analyst | Extract author's judgment patterns | 01-author-thinking.md |
Methodology Miner | Convert scattered techniques to MethodCards | 02-method-catalog.md |
Scenario Router | Map methods to problem types | 03-scenario-router.md |
Skill Packager / QA | Compile skill spec and validate evidence | 04-subagent-playbooks.md, 05-master-skill-spec.md |
Read references/subagent-playbooks.md before assigning work.
Output Bundle
Produce these Markdown artifacts:
00-overview.md - Book goals, structure, concept map, keywords
01-author-thinking.md - Author's observation patterns, judgment framework
02-method-catalog.md - All MethodCards, grouped by theme
03-scenario-router.md - Problem-to-method mapping and combination rules
04-subagent-playbooks.md - Subagent duties, inputs, outputs, handoffs
05-master-skill-spec.md - Top-level skill logic, delegation, output rules
Schemas defined in references/output-schemas.md.
Generated Skill Directory
When the extraction is materialized on disk, the final target is a standard skill directory rather than a loose output/ folder:
skills/<book-slug>/SKILL.md
skills/<book-slug>/workflow.yaml
skills/<book-slug>/references/00-05*.md
skills/<book-slug>/queries/pack-*.yaml
skills/<book-slug>/logs/run-*.jsonl
Use the lightweight runtime scripts to create this layout:
scripts/init_generated_skill.py
scripts/wait_notebooklm.py
scripts/finalize_generated_skill.py
Operating Rules
| Rule | Violation Consequence |
|---|
| Use NotebookLM MCP as primary interface | Risk using training data instead of actual book content |
| Use staged, purpose-built prompts | Open-ended chat leads to inconsistent extraction |
| Every conclusion must have evidence | Skills drift from book content, become unreliable |
| Weak conclusions → mark as pending | Publishing unverified methods damages skill quality |
| First version: single book only | Scope creep, unfinished multi-book mess |
| Logical parallel, actual serial execution | Race conditions, inconsistent state |
Common Mistakes & Fixes
| Mistake | Why It Happens | Fix |
|---|
| Relying on training data about the book | Faster than reading | Force NotebookLM query first, every time |
| Creating generic summaries | Easier than structured extraction | Follow the five-stage query packs strictly |
| Missing evidence references | Forget to record during extraction | Add evidence field to every MethodCard template |
| Weak conclusions promoted | Want to appear complete | Mark as pending_confirmation instead |
| Skipping subagent assignments | Simpler to do it all yourself | Use fixed role system, delegate systematically |
Reference Guide
Quick Reference: Default Prompt Shape
When invoked without detailed spec, assume the user wants:
- Structured extraction of the book's thinking system
- Methodology cards with evidence
- Routing logic for real user problems
- Master skill specification for Claude Code / OpenClaw execution
Start with: "I'll help you convert this book into reusable skills. First, let me set up the extraction workflow..."