How to use the Cangjie Book Distiller agent skill (cangjie-skill)

What an agent skill is and what cangjie-skill makes the agent do

An agent skill is a folder containing a SKILL.md file that tells an AI agent when to use a capability and how to carry it out. The agent reads the skill's description and loads it when a request matches.

cangjie-skill is a meta-skill: its job is to build other skills. Given source text, it reads the material, extracts methods that can be applied in real situations, checks them against the source, and compiles them into a set of focused skills plus a reader-friendly digest.

• Extracts decision frameworks, principles, operating procedures, calculation rules, troubleshooting steps and checklists
• Verifies candidates against the source text before keeping them
• Groups the kept methods into atomic skills that an agent can invoke in real situations
• Produces a digest (DIGEST.md) written for a human reader
• Does not produce plain summaries, book reviews or role-played authors (the listing points to nuwa-skill for persona work)

Install cangjie-skill from the ZIP

Download the ZIP from the skill's page on this directory and unzip it. Place the resulting skill folder in your personal skills directory, ~/.claude/skills/, or in a project's .claude/skills/ directory if you want it available only inside that project.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Install the Python dependency for the compile scripts

The deterministic scripts need Python 3.10 or newer and the PyYAML package. If you run into missing-module errors, run the self-check first; the source says it prints installation guidance when PyYAML is missing.

• Requirement: Python 3.10+
• Requirement: PyYAML

Trigger the skill with natural requests

You do not need a special command. Ask in plain language that describes what you want to extract from a book or long recording. The skill's description lists trigger phrases such as 'turn a book or video into skills' and the Chinese phrases 拆书 and 蒸馏一本书. The source also gives these examples of requests that should trigger it.

Name the source and say what you want back. If you do not have the text yet, the skill asks you for it rather than working from memory, so have the file path or transcript ready.

• 'Distill this book into skills: <path to file>'
• 'I want to turn the methods in this book into a skill I can use.'
• 'Turn this video, podcast or course into skills.'
• 'Help me take apart the book and make its methods usable.'

Give the skill what it needs before it starts

The skill asks for four things. Source text is the most important: a PDF, EPUB, TXT file, subtitles, or accessible plain text. For videos and podcasts, the source suggests first getting a transcript with a video downloader-type tool.

Metadata is used for naming and auditing: for a book, title, author and publication year; for a video, podcast or course, title, creator and publication time. The skill also asks what you plan to do with the output, which decides whether it recommends a single-skill or multi-skill output, and whether this is your first run.

• Source text file or transcript (required; it will not distill from memory)
• Title, author and date metadata
• Intended use: study a book (usually single) or add to daily workflows (usually pack)
• Whether this is a first trial run (the skill recommends starting with one book)

Run the compile step and review the output

After extraction and checking, the skill compiles its capability bundle into skills using the compile script. It writes the output to the folder you choose and shows you a decision report so you can confirm the mode. Compiled files are copied or symlinked into your skills location only after you approve the install location.

The skill keeps a working folder for each book. Useful files include PIPELINE_STATE.md, which lets a stopped run resume, and candidates/ and rejected/, which record what was considered and why items were dropped.

• Working folder: books/<book-slug>/
• Resume point: PIPELINE_STATE.md
• Audit trail: candidates/ and rejected/
• Final digest: DIGEST.md

Safety notes: scripts and network

The skill runs scripts. Its compile and validation steps are Python scripts in the scripts/ folder, so read them before running if you have concerns. The listing says it does not use the network. Review any script output before you rely on it.

The skill's own rules say the compiled output folders are read-only in practice: if it detects local hand edits, it will not overwrite them silently. Keep copies of your source files and of any edits you make.

• Runs local Python scripts (scripts/)
• Listed as not using the network
• Will not silently overwrite hand-edited output folders
• Source copyrighted books: the project itself asks contributors not to submit copyrighted originals, so use your own copy only for your own work

Tips for getting useful results

Start with one book so you can judge the process before running a batch. Review the candidate list and the rejected list, because the skill's own review step is a human confirmation point and the skill notes that confirmation does not replace checking against the source.

Keep the output mode steady across updates. The skill keeps your chosen mode by default so new material does not quietly change the output.

• Start with a single book on a first run
• Check candidates/ and rejected/ for anything important that went missing
• Verify key procedures against the original text
• Use the same output mode for later updates