cangjie-skill: Distill a Book or Long Video Into a Pack of Executable Agent Skills

A 600-page book, a 90-minute lecture, a five-hour podcast: useful methodology is buried in prose. Most people extract a summary and lose the part that would actually be reusable. kangarooking/cangjie-skill runs the RIA-TV++ pipeline — Zhao Zhou's RIA note method extended with Triple Verification and agent-execution hooks — and turns the source into a pack of callable, testable Agent Skills, complete with SKILL.md modules, a glossary, an index, and a test-prompts.json file.

Why This Skill Matters

Five extractor prompts run in parallel: framework, principle, case, counterexample, and terminology. Each produces candidates. Triple Verification then filters ruthlessly — only 25–50% of candidates pass. The gate requires at least two independent cross-domain citations, predictive power on unstated questions, and non-trivial uniqueness. The output is not a digest; it is a small library of skills with trigger conditions and scope boundaries, so the reader can later invoke "the framework skill" or "the counterexample skill" without re-reading the source.

For anyone who consumes long-form content and re-uses the methods inside it — researchers, operators, course designers, technical writers — this converts passive reading into active, callable assets.

Installation

There is no Claude Code plugin entry in the README. The cleanest install path is the DeepSeek Harness CLI plugin release:

mkdir -p ~/.dsh/packages
curl -fL "https://github.com/kangarooking/cangjie-skill/releases/download/v2.0.0/dsh-cangjie-skill-2.0.0.tgz" \
  -o ~/.dsh/packages/dsh-cangjie-skill-2.0.0.tgz
dsh plugin --profile web add ~/.dsh/packages/dsh-cangjie-skill-2.0.0.tgz
dsh web

For Claude Code use, clone the repo into a working directory and reference the meta skill from a project-local instruction:

git clone https://github.com/kangarooking/cangjie-skill ~/work/cangjie-skill

Then point your agent at ~/work/cangjie-skill/SKILL.md (the meta skill definition) so it knows how to invoke the five extractors and the templates.

Real Workflow: Turn a Book Into a Skill Pack

You just finished a 350-page product strategy book. You want the methodology available as skills you can call later — without re-reading the book.

Step 1. Drop the book's text (or a clean PDF extract) into a working directory, for example ~/work/books/product-strategy/. Make sure the text is searchable, not a flat scan.

Step 2. In a Claude Code session, ask:

Use cangjie-skill at ~/work/cangjie-skill to distill the text under
~/work/books/product-strategy/ into a skill pack. Run the five extractors
in parallel, apply Triple Verification, and write the output to
~/work/books/product-strategy/distilled/.

Step 3. The skill runs framework / principle / case / counterexample / terminology extraction, then filters. Expect 25–50% of candidates to survive. The output directory looks like:

~/work/books/product-strategy/distilled/
├── BOOK_OVERVIEW.md       ← global understanding
├── INDEX.md               ← skill map with trigger conditions
├── DIGEST.md              ← reader-facing summary
├── GLOSSARY.md            ← term dictionary
├── skills/
│   ├── pricing-framework/
│   │   └── SKILL.md
│   ├── positioning-counterexamples/
│   │   └── SKILL.md
│   └── qualitative-research-loops/
│       └── SKILL.md
└── test-prompts.json      ← verification scenarios

Step 4. Open INDEX.md and skim the trigger conditions for each skill. When one matches a future prompt, invoke it directly instead of re-deriving the method.

Real Workflow: Distill a Long YouTube Lecture Into Callable Skills

You watched a 90-minute lecture with a strong methodology (for example: a thinking framework the speaker teaches through examples). You want the framework as a callable skill, not a transcript summary.

Step 1. Get the transcript. The README pairs cangjie-skill with the companion video-downloader skill; use that to fetch the video and a clean subtitle file, then pass the subtitle .srt or .txt into the cangjie pipeline:

Distill the lecture transcript at ~/work/lectures/framework-90m.txt into a
skill pack using cangjie-skill. Bias toward counterexamples — the
speaker's methodology is clearest in the cases they reject.

Step 2. The five extractors run on the transcript. Triple Verification then filters: a framework survives only if the speaker demonstrates it on at least two distinct domains in the lecture.

Step 3. Read DIGEST.md first — it is the short orientation. Then read INDEX.md and pick the skill you most want to internalize next. Open its SKILL.md, paste it into a fresh Claude session, and run one of the prompts from test-prompts.json to verify the skill behaves as described.

Step 4. Repeat for the next lecture. Over a quarter, you build a personal library of ~/work/lectures/<topic>/distilled/skills/... packs that you can grep, invoke, or chain together.

• Bias the extractors when you invoke them — for a methodology book, weight framework + counterexample; for a case-study book, weight case + principle.
• Triple Verification is the gate that keeps quality high; if you see only a handful of skills survive, that is the intended outcome, not a bug.
• Keep the source text alongside the distilled output so you can re-run cangjie-skill later with a new prompt bias and compare outputs — useful for capturing methodology you missed the first pass.
• Treat test-prompts.json as a regression suite: re-run it on the skill after any edit to make sure the trigger condition still holds.

When Not to Use This

If the source is short, opinion-only, or lacks verifiable methodology (a motivational talk, a one-hour vlog, a thread of personal takes), the Triple Verification filter will strip almost everything and you will end up with an empty pack. Skip cangjie-skill for sources under ~30 pages of dense content, and skip it entirely for opinion pieces.

See the leaderboard for more content-distillation skills.