book-to-skill-distillation

agent
Security Audit
Pass
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 10 GitHub stars
Code Pass
  • Code scan — Scanned 10 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Agent skill for distilling books and long PDFs into agent-native skills

README.md

book-to-skill-distillation

⚡ Quick Start

git clone https://github.com/9s5bz2jvd2-lang/book-to-skill-distillation
cd book-to-skill-distillation
pip install pymupdf
python scripts/setup_distill.py your-book.pdf
# → Workspace created. Now follow SKILL.md for distillation.

What this does: Takes a book/PDF and converts it into an agent-native skill — routers, checklists, decision trees, not a summary.


What is this?

This is a loop engineering skill for LingTai agents. It converts books, long PDFs, EPUBs, manuals, guideline collections, and databases into agent-native skill structures — the branching, callable artifacts an agent uses at runtime.

"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." — Peter Steinberger

This skill IS a loop. You don't prompt the agent to "read chapter 3 and take notes." You design a system that:

  1. Finds relevant material via sparse routing (ROUTING.yaml)
  2. Hands out work to parallel workers (daemon fan-out)
  3. Checks quality via missed-case sweep
  4. Records what's done in structured stores (source_catalog.jsonl, node_store.jsonl)
  5. Decides the next step based on graph edges and budget tiers
  6. Returns compressed artifacts to cache for future calls

Why not just summarize?

Most "book → notes" tools produce shorter prose. That's the wrong shape for an LLM agent. A linear chapter summary doesn't help an agent:

  • Route a task to the right procedure
  • Check whether it missed an edge case
  • Decide whether it needs to load a heavy reference
  • Verify whether a claim depends on current law or pricing

This skill encodes a repeatable workflow for the rewriting step: taking linear human text and turning it into the branching, callable artifacts an agent uses at runtime.

How it works — the Loop

                    ┌─────────────────────────────────────────────┐
                    │                                             │
                    ▼                                             │
           ┌──────────────┐                                      │
           │  Shared Core  │  ← Always loaded first (200-800 tok) │
           │  (copyright,  │                                      │
           │   safety,      │                                     │
           │   red lines)   │                                     │
           └──────┬───────┘                                      │
                  │                                               │
                  ▼                                               │
        ┌─────────────────┐        ┌──────────────────┐          │
        │  ROUTING.yaml   │ ─────▶ │  Top-k Experts   │          │
        │  (sparse        │        │  (800-3000 tok   │          │
        │   activation)   │        │   each)          │          │
        └────────┬────────┘        └────────┬─────────┘          │
                 │                          │                     │
                 ▼                          ▼                     │
        ┌─────────────────┐        ┌──────────────────┐          │
        │  Missed-case    │        │  Route Log       │          │
        │  Sweep          │◀──────│  (what was       │          │
        │  (systematic    │        │   loaded/skipped) │          │
        │   edge-case     │        └──────────────────┘          │
        │   checklist)    │                                      │
        └────────┬────────┘                                      │
                 │                                                │
                 ▼                                               │
        ┌─────────────────────────────────┐                     │
        │  Cyclic Return-to-Cache         │─────────────────────┘
        │  (compress results back into    │   (better cache hit
        │   ROUTING, GRAPH, CACHE, evals)  │    next time)
        └─────────────────────────────────┘

Key concepts

Concept What it means Where it lives
Sparse activation Load only what the task needs — not the whole skill ROUTING.yaml
Shared core Always-on safety/copyright/red-line rules (≤800 tokens) SKILL.md
Top-k routed experts The smallest sufficient expert set for the task reference/
Missed-case sweep Systematic edge-case checklist after sparse activation ROUTING.yaml
Budget tiers Token budget per component type ROUTING.yaml
Cyclic return-to-cache Compress results back into routing/cache for next time CACHE.md
Source packs Multi-book domain routing with conflict gates ROUTING.yaml

What's in this repo

SKILL.md                     # Thin sparse router — first file an agent loads
RUNBOOK.md                   # Demand-sensitive sparse/full reading execution manual
ROUTING.yaml                 # Machine-readable activation hints, budget tiers, triggers
CACHE.md                     # Cache-friendly layout design (stable prefix + variable suffix)
RULES.md                     # Ownership, reuse boundaries, content safety

scripts/
  setup_distill.py           # Bootstrap: PDF → workspace (TOC + fulltext + catalog)
  route_plan.py              # Generate demand-sensitive sparse reading plan
  fanout_daemons.py          # Generate parallel daemon tasks from a doc list
  quality_check.py           # Lightweight release checks
  extract_bookmarks.py       # Dump PDF TOC as JSON
  scout.sh                   # Quick PDF reconnaissance (pdfinfo + sample text)
  render_pages.sh            # Render PDF pages to PNG for OCR

reference/
  sparse-distillation-methodology.md   # Full methodology
  toc-recovery.md                      # TOC extraction patterns
  validation-checklist.md              # Output validation
  worked-example-legal-dd.md           # Worked example

assets/
  output-template.md                   # Distilled skill module template
  route-log-template.md                # Route log template
  eval-cases.md                        # Evaluation cases
  route-log-examples.md                # Route log examples
  reference-doc-template.md            # Reference doc template

examples/
  quickstart-example.md                # Concrete before/after worked example

Operational discipline

  • Copyright-safe: Extract operational know-how (procedures, checklists, taxonomies); do not reproduce prose or examples beyond tiny necessary snippets.
  • Privacy-safe: Source-derived OCR substrate stays under a private work/book-distill/<slug>/ workspace, not inside the published skill.
  • Verification-aware: Any claim depending on current law, pricing, or API behavior is marked for refresh against an authoritative source.
  • Token-aware: Budget tiers in ROUTING.yaml control token consumption per component type. The loop is designed to be token-efficient by default.

Multi-book domain support

For same-domain multi-book distillation (e.g., a nutrition guideline library, a legal handbook collection):

  1. Each book becomes a source pack with its own source_map
  2. A domain parent skill routes over source packs
  3. Source anchors preserve which book said what
  4. Conflict gates log contradictions between sources
  5. Unified graph index connects nodes across all books

Rule: Do not blend sources into anonymous consensus; preserve source identity and log conflicts.

Relation to 生万物论 (Sheng Wanwu)

This skill is the 成库 (library) layer for 生万物论 — a hypothesis discovery engine that builds on distilled knowledge bases:

道 (papers) → 一 (distill to library) → 二 (graph) → 三 (hypothesize) → 万物 (discover)

book-to-skill-distillation handles the 道→一→二 layers. shengwanwu handles the 二→三→万物 layers.

License

© 2026 王润圆 (Wang Runyuan) · Kunming Medical University · Master of Nutrition and Food Hygiene

  • ✅ Allowed: Share, adapt, learn, research
  • ❌ Prohibited: Any form of commercial use
  • Attribution required when referencing

Author

王润圆 (Wang Runyuan)
Kunming Medical University · Master of Nutrition and Food Hygiene
China Registered Dietitian


Powered by LingTai — agent-native loop engineering.



📜 许可 · License

本项目为公益开源,采用 MIT 许可证:

  • ✅ 随意使用:学习、研究、转载、二次创作、商用也可以 —— 保留原作者署名(王润圆 Runyuan Wang)就好啦 💛
  • 🌱 开源是为了帮助更多的人。

This project is public-welfare open source under the MIT License — free for anything, including commercial use; just keep the original credit ("Runyuan Wang") 💛

© 2026 王润圆 Runyuan Wang · MIT License

Reviews (0)

No results found