Cangjie Skill
Cangjie Skill applies the RIA-TV++ pipeline to turn books, transcripts, podcasts, courses, and interviews into verified, composable, pressure-tested Agent Skills.
Project overview
Cangjie Skill is a meta-skill that creates other Skills. Instead of compressing a book or long video into one summary, it first models the whole source, runs framework, principle, case, counter-example, and glossary extractors, filters candidates with three verification tests, and builds the accepted units into Skills with triggers, execution steps, boundaries, and test prompts. It is designed for users who want long-form knowledge to become reusable agent tools.
Core capabilities
Whole-source comprehension and audit
The Adler stage covers structure, interpretation, criticism, and application, producing BOOK_OVERVIEW.md with resumable state.
Five-way methodology extraction
Framework, principle, case, counter-example, and glossary extractors can run in parallel or serially with the same output format.
Triple verification and RIA++
Candidates need cross-passage support, predictive power, and non-obviousness before being structured for reading, interpretation, application, execution, and boundaries.
Pressure-tested delivery
Each Skill receives positive, negative, and ambiguous trigger tests; failures return for reconstruction before installation.
Installation and usage
Place the repository in a host Skills directory to load its root SKILL.md, or install the documented v2.0.0 package for DeepSeek Harness.
Let an AI Agent install it
Send this prompt to Codex, Claude Code, or another AI agent that can work with your local environment.
Install Cangjie Skill from https://github.com/kangarooking/cangjie-skill. Read README.md, README.zh-CN.md, SKILL.md, and the v2.0.0 release. Install the repository into the user- or project-level Skills directory I specify. For DeepSeek Harness, use only the documented v2.0.0 tgz download and `dsh plugin --profile web add` command. Verify with a short local text and confirm it first asks for source text, metadata, and pilot scope. Ask before downloading a release, running dsh, creating or overwriting directories, copying Skills, reading source files, spawning multiple sub-agents, or requesting additional permissions.- Readable book, subtitle, transcript, or course text
- A Skills-capable agent host or DeepSeek Harness
- Project write access
- Enough context and execution time for sub-agents or serial extractors
Project URL: https://github.com/kangarooking/cangjie-skill
Copy the URL and follow the steps below to complete setup.- 1Connect the Skill or DSH package
Place the repository in the host Skills directory, or use the README's v2.0.0 tgz and dsh plugin command for DeepSeek Harness.
- 2Prepare source text and metadata
Provide a PDF, EPUB, TXT, subtitle, or transcript plus title, author, and date; agree on one pilot source for a first run.
- 3Review comprehension and filtered candidates
Generate BOOK_OVERVIEW.md, run the five extractors and triple verification, then confirm accepted and rejected candidates.
- 4Construct, test, and install outputs
Build complete RIA++ Skills, links, and bait tests; reconstruct failures and install only the tested results.
Run `Use cangjie-skill to distill this book into a set of executable Agent Skills: <file path>`. A correct installation should ask for the content source, metadata, and whether this is a first pilot instead of generating Skills without text.
- Stop when readable source text is missing instead of distilling from model memory.
- Start with one source on a first run before scaling to a batch.
- Confirm the installation destination before copying or linking generated Skills.
Use cases
Book methodologies as Skills
Convert decision frameworks, principles, and cases into atomic capabilities with evidence and boundaries.
Long video or podcast reuse
Filter transferable methods from subtitles or transcripts instead of producing a one-time summary.
Course and document Skill Packs
Organize multi-part sources into linked Skills, a glossary, an index, and repeatable trigger tests.
Assessment
The README, SKILL.md, methodology, extractors, templates, and listed examples show that Cangjie Skill treats knowledge distillation as an auditable filter-build-test pipeline rather than summary packaging. The tradeoff is a long process that depends on complete source text and checkpoints. This is a documentation-based assessment, not a reproduction of its example quality.
Why it may be useful
- Explicitly rejects memory-based generation and simple summarization
- Preserves candidates, rejected units, citations, boundaries, and tests
- Covers books, videos, podcasts, courses, interviews, and document collections
What to know first
- The full seven-stage process requires substantial time
- Parallel extraction and blind testing depend on sub-agents or serial fallback
- Outputs still require user review, source-rights decisions, and quality control
README
Cangjie Skill
Overview
Cangjie Skill applies the RIA-TV++ pipeline to turn books, transcripts, podcasts, courses, and interviews into verified, composable, pressure-tested Agent Skills. Distill high-value content from books, long-form videos, podcasts, and more into executable Agent Skills.
Getting started
- Place the repository in a host Skills directory to load its root SKILL.md, or install the documented v2.0.0 package for DeepSeek Harness.
- RIA-TV++ covers whole-source understanding, five-way extraction, triple verification, RIA++ construction, linking, pressure testing, and delivery.
- Generated Skills need triggers, boundaries, executable steps, and bait tests rather than generic notes.
- Run `Use cangjie-skill to distill this book into a set of executable Agent Skills: <file path>`. A correct installation should ask for the content source, metadata, and whether this is a first pilot instead of generating Skills without text.
Configuration
Project URL: https://github.com/kangarooking/cangjie-skill
Copy the URL and follow the steps below to complete setup.