# Cangjie Skill

> Distills methodologies from books, long-form videos, and podcasts into installable, callable AI Agent Skills.

Cangjie Skill is an open-source meta-skill by kangarooking that turns books, video transcripts, podcast transcripts, interviews, and other long-form content into executable AI Agent Skills. The current release is v2.5.0, and it works with OpenClaw, Claude Code, and DeepSeek Harness. Its output is a structured skill repository or a compact router-style skill rather than a summary document.

## What It Is

Cangjie Skill is a skill generator. A user supplies source text, such as a book file or the subtitles and transcript of a video or podcast, and the tool extracts methodologies that are verifiable and transferable, then compiles them into skills an agent can invoke. The README says it is designed for structured reuse by agents rather than for human-oriented summaries or notes.

## How the RIA-TV++ pipeline works

The README describes a seven-stage process:

- Adler-style whole-content analysis, producing a book overview.
- Parallel extraction by five extractors: frameworks, principles, cases, counter-examples, and glossary.
- Triple verification and a promotion gate.
- RIA++ capability cards covering Reading, Interpretation, Appropriation, Execution, and Boundary.
- Zettelkasten-style linking of dependencies, contrasts, and compositions.
- Pressure testing with bait questions and cross-skill confusion tests.
- Deterministic compilation into delivery artifacts.

## Outputs and delivery modes

Extraction first produces a Capability Bundle, which compiles into either a single router-style skill or a compact pack with a router plus promoted standalone skills. Generated artifacts include BOOK_OVERVIEW.md, INDEX.md, a reader-facing DIGEST.md, GLOSSARY.md, multiple SKILL.md files, and test-prompts.json for trigger testing. A local script, scripts/cangjie.py, covers diagnostics, compilation, replanning, incremental updates, repair, rollback, evaluation, and benchmarking. It requires Python 3.10+ and PyYAML.

## Setup path

The generic skill is distributed as a release ZIP that is extracted and installed as the complete cangjie-skill directory. A separate tarball installs into DeepSeek Harness through the dsh plugin command. For video sources, the author recommends first using the companion video-downloader skill to obtain subtitles or transcripts.

## Generated skill packs

The README lists example packs distilled from books and video courses, including Buffett's shareholder letters, Influence, Contagious, and Andrew Ng's AI for Everyone course.

## Features
- RIA-TV++ seven-stage distillation pipeline
- Works on books, video and podcast transcripts, interviews, talks, and courses
- Five parallel extractors for frameworks, principles, cases, counter-examples, and glossary
- Triple verification and promotion gate
- Capability Bundle as single source of truth
- Single or compact pack compilation modes
- Pressure testing with bait and cross-skill confusion prompts
- Unified local toolchain for diagnostics, compile, update, repair, rollback, evaluation, and benchmarking
- Generates INDEX, DIGEST, GLOSSARY, SKILL.md files, and test prompts
- DeepSeek Harness plugin package

## Integrations
OpenClaw, Claude Code, DeepSeek Harness

## Platforms
API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
2.5.0

## Links
- Website: https://github.com/kangarooking/cangjie-skill
- Repository: https://github.com/kangarooking/cangjie-skill
- EveryDev.ai: https://www.everydev.ai/tools/cangjie-skill