kangarooking/cangjie-skill

把书、长视频、播客等高价值内容蒸馏成可执行的 Agent Skills（Distill high-value content from books, long-form videos, podcasts, and more into executable Agent Skills）

What is Cangjie Skill?

Cangjie Skill is a toolchain that turns the methodologies, frameworks, principles and concrete procedures found in books, long‑form videos, podcasts or other dense content into callable “AI Skills” – small, self‑contained modules an AI agent can invoke at runtime. In other words, instead of just summarising a book, it extracts the actionable knowledge and packages it so an LLM‑powered agent can execute the ideas directly.

How does it work?

Input – Any text source (book, transcript, notes).
RIA‑TV++ pipeline – a seven‑stage deterministic process:
- Whole‑content comprehension (Adler analysis) creates a high‑level overview.
- Parallel extraction runs five specialised extractors (frameworks, principles, cases, counter‑examples, glossary).
- Triple verification checks each candidate for source sufficiency, executability and task utility.
- Capability construction builds Capability Cards (R/I/A1/A2/E/B) stored under .cangjie/capabilities/.
- Zettelkasten linking weaves a graph of dependencies and cross‑references.
- Pressure testing generates test prompts to ensure the skill behaves as intended.
- Deterministic compilation produces either a single router‑style skill (single) or a compact pack (pack).
Output – A directory containing:
- BOOK_OVERVIEW.md, DIGEST.md, GLOSSARY.md
- Individual */SKILL.md files for each extracted methodology
- test‑prompts.json for validation
- Installable artifacts (ZIP/tarball) ready for platforms such as OpenClaw, Claude Code or DeepSeek Harness.

Key Features (as of v2.5.0)

Capability Bundle – a single source of truth for all extracted cards and metadata.
Two deterministic delivery modes – single (router‑only) or pack (router + stand‑alone skills).
Unified local toolchain – scripts/cangjie.py handles diagnostics, incremental updates, rollback, evaluation and benchmarking.
Safety mechanisms – content‑addressed preprocessing, diff‑based impact analysis, transactional patches, snapshots and rollback.
Registry integration – supports both Registry v1 and v2, with website‑visible capability counts.
Platform adapters – ready‑to‑install plugins for OpenClaw, Claude Code and DeepSeek Harness.

Who might use it?

AI product teams that want to embed domain‑specific knowledge (e.g., investment reasoning from Buffett’s letters) into their agents.
Knowledge workers who have accumulated many books or video courses and want the insights to be actionable rather than just archived.
Researchers building “skill ecosystems” where skills distilled from people (nuwa‑skill) and from texts (cangjie‑skill) can evolve together via darwin‑skill.

Example Use‑Case

A user has a PDF of The Lean Startup. Running the pipeline produces a skill pack containing separate modules such as “Validated Learning Loop”, “Minimum Viable Product”, and “Pivot Decision Framework”. An LLM‑agent can then be prompted:

Use cangjie‑skill to run the "Validated Learning Loop" on my product idea.

The agent executes the concrete steps defined in that skill, rather than merely reciting a summary.

Ecosystem & Related Projects

nuwa‑skill – distills people (thinking styles, expression DNA) into AI skills.
cangjie‑skill – this repo, focused on books and other structured content.
darwin‑skill – evolves any skill over time.

Getting Started

Install Python 3.10+ and pyyaml.
Run the self‑check: python3 scripts/cangjie.py doctor.
Follow the website’s “Skill Pack browsing” guide or use the DeepSeek Harness plugin as shown in the README.

License

MIT – free to use, modify and redistribute.

In short: Cangjie Skill is a reproducible, safety‑first pipeline that converts dense learning material into modular, testable AI skills that agents can call on demand, turning passive knowledge into active capability.