简体中文 · English · 日本語

Cangjie Skill

Distill methodologies from books, long-form videos, and podcasts into callable AI Skills

Finish reading, watching, or listening—and leave with a methodology you can invoke.

Official Website

🌐 Visit the Cangjie Skill official website

The website provides visual Skill Pack browsing, a beginner-friendly usage guide, Skill detail pages, and a contribution submission entry. This GitHub repository remains the sole source for cangjie-skill code, methodology, and templates; the website provides presentation, navigation, and usage guidance.

What's New in v2.5.0

• Capability Bundle as the single source of truth: extraction produces stable capability cards and metadata before any installable output is compiled.
• Two deterministic delivery modes: compile one router-style Skill (single) or a compact pack with a router plus promoted standalone Skills (pack).
• A unified local toolchain: scripts/cangjie.py now covers diagnostics, compilation, output replanning, incremental updates, repair, rollback, evaluation, and benchmarking.
• Safer evolution: content-addressed preprocessing, source diffs, impact analysis, transactional patches, edit detection, snapshots, and rollback are included.
• Registry v2 and website support: output mode and capability counts are visible without breaking existing Registry v1 entries.

See the v2.5.0 release notes and changelog for the complete scope and migration notes.

2026-09-13 refresh (still v2.5.0): task-first validation now retains complete procedures and formulas explained in a single source location. Output scoring counts missing runs and checks numeric values/units; compiled Skills can carry declared scripts and text templates. Download the refreshed generic Skill ZIP · SHA256. Extract it and install the complete cangjie-skill/ directory. Existing users must download the refreshed package; check BUILD_INFO.json for the source commit and refresh date. The original tag is unchanged, so GitHub's automatic source archives do not contain this refresh.

DeepSeek Harness Plugin

cangjie-skill also provides a standalone installation package for DeepSeek Harness. The adapter layer is bundled in the Release package, so no platform-specific wrapper files are added to this repository.

After installing DeepSeek Harness, download the v2.5.0 package and checksum, verify it, then install from the local tarball:

mkdir -p ~/.dsh/packages
curl -fL "https://github.com/kangarooking/cangjie-skill/releases/download/v2.5.0/dsh-cangjie-skill-2.5.0.tgz" \
  -o ~/.dsh/packages/dsh-cangjie-skill-2.5.0.tgz
curl -fL "https://github.com/kangarooking/cangjie-skill/releases/download/v2.5.0/dsh-cangjie-skill-2.5.0.tgz.sha256" \
  -o ~/.dsh/packages/dsh-cangjie-skill-2.5.0.tgz.sha256
(cd ~/.dsh/packages && shasum -a 256 -c dsh-cangjie-skill-2.5.0.tgz.sha256)
dsh plugin --profile web add ~/.dsh/packages/dsh-cangjie-skill-2.5.0.tgz
dsh web

Download the DeepSeek Harness plugin (for Cangjie Skill v2.5.0) · SHA256 checksum

After starting a new task, you can say:

Use cangjie-skill to distill this book into a set of executable Agent Skills: <file path>

Why This Exists

There's a recent viral idea: distilling colleagues into AI skills. Even after someone leaves, their experience, tone, and work style can be partially replicated by AI. nuwa-skill does exactly this — creating "human skills" like an Elon Musk skill or a Warren Buffett skill. The companion darwin-skill handles automatic skill evolution.

Distilling people is valuable — nuwa-skill has already proven this. Distilling the content people have expressed systematically is a complementary dimension: a book, a long-form interview, a podcast episode, or a long Bilibili or YouTube video can contain methodologies that took the creator years to refine. Rather than imitating someone's expression style, extracting those methodologies and turning them into tools that solve real problems is equally valuable.

There's also a real pain point: you may read many books, save many videos, and listen to many podcasts, yet still struggle to apply what you learned. Content-rich long videos are published every day, are often time-sensitive, and can be difficult to absorb in one viewing; they may not be represented in an AI model's training data at all. Once this content is distilled into skills, an AI agent can invoke the knowledge in real scenarios instead of letting it gather dust in notes, bookmarks, or watch-later lists.

So cangjie-skill has one clear goal: distill every piece of high-value content worth distilling. It works not only with books, but also with videos that have subtitles or transcripts, podcasts, interviews, talks, courses, long-form articles, and document collections. Whenever content contains extractable, verifiable, and transferable methodologies, cangjie-skill can turn them into independently callable, composable, and pressure-testable AI skill packs.

For video content, we recommend using the video-downloader skill alongside cangjie-skill. Use it first to download the video, extract subtitles or audio transcripts, and collect key materials; then pass the resulting text to cangjie-skill for methodology extraction, skill construction, and pressure testing.

What Problems It Solves

• Reading many books, watching many videos, or listening to many podcasts without applying them — knowledge stays at the "read/watched/listened/saved" level and cannot be invoked in real decisions
• Summaries, notes, and organized transcripts are compression, not structured reuse — after reading or watching, you still do not know "what to use when"
• Only a small fraction of high-value content deserves to become a tool — strict filtering is needed, not wholesale inclusion
• Existing methods for reading, watching, and learning are designed for people, not agents — distillation must be execution-oriented rather than consumption-oriented

How It Works

cangjie-skill uses the RIA-TV++ pipeline to transform source texts—including books, video transcripts, podcast transcripts, and interview notes—into a reusable Capability Bundle, then compiles that source into installable skills. The process has seven stages:

1. Whole-Content Comprehension (Adler Analysis) — Structural, interpretive, critical, and applicability analysis using Mortimer Adler's method, producing BOOK_OVERVIEW.md
2. Parallel Extraction — Five specialized extractors (frameworks, principles, cases, counter-examples, glossary) run simultaneously to pull candidate units from the source text
3. Triple Verification + Promotion Gate — Check source sufficiency, executability, and task utility by candidate type. A complete procedure or formula explained once can qualify; repetition or author originality is not mandatory. References and unresolved candidates remain auditable, and standalone entrypoints are decided separately
4. RIA++ Capability Construction — Verified content is structured into R / I / A1 / A2 / E / B capability cards inside .cangjie/capabilities/
5. Zettelkasten Linking — Dependencies, contrasts, and compositions are encoded in the Bundle's capability graph and shared glossary
6. Pressure Testing — Test prompts including bait questions (and cross-skill confusion tests) are designed for each skill; failures go back for full reconstruction
7. Deterministic Compilation and Delivery — The same Bundle compiles to single or compact pack, alongside a reader