cangjie-skill How to use: Turn books, videos and podcasts into executable Agent Skills

Many people have a collection of books, long videos, and podcasts, and end up with only summaries, and still can’t remember which method to use when it comes to actually doing things. The goal of cangjie-skill is to convert high-value content into Skills that Agent can call in specific scenarios, rather than just compressing it into a reading note.

Project address: kangarooking/cangjie-skill

Quick Answer

Materials suitable for input include:

• Books or complete chapters;
• Long videos with subtitles or transcribed text;
• Podcasts, interviews and course transcripts;
• Methodologically dense long articles and collections of resources.

If the material only has opinions, news or emotional expressions, and no verifiable and transferable operation methods, it is not suitable to be forcibly made into a skill.

What is the difference between it and a normal summary?

A normal summary answers “What does this content say?” Skill also answers:

1. What scenes should be triggered;
2. What steps should be followed;
3. Under which conditions it does not apply;
4. Easily confused with any method;
5. How to use testing to prove that it really works.

The project adopts the RIA-TV++ process, which includes overall understanding, parallel extraction, triple verification, structured construction, inter-Skill linking, stress testing and final delivery.

What will be generated by a complete conversion?

Typical output is not a single SKILL.md, but a set of files:

This structure is suitable for subsequent maintenance: when updating raw materials, only the affected Skills can be redone instead of rewriting the entire summary.

Recommended workflow

1. First obtain the legally available original text, subtitles or personal transcription;
2. Clean up advertisements, repeated fragments and transcription errors;
3. Identify the real problem you want to solve;
4. Let cangjie-skill extract candidate methods;
5. Delete common sense, single cases and unverifiable opinions;
6. Add trigger conditions and boundaries for retained Skills;
7. Verify with confusing test questions.

When processing videos, subtitle extraction or speech transcription should be completed first before plain text is entered. Don’t let the Agent infer the entire paragraph from the title and introduction alone.

What content is worth distilling?

A piece of material is suitable for conversion into a Skill and usually has the following characteristics:

• Contains steps that can be repeated;
• Applicable conditions and counterexamples are given;
• Methods can be migrated outside the original case;
• The original text is supported by multiple pieces of evidence;
• Able to answer “when to use” and “when not to use”;
• Can design tests to verify results.

Pure news, personal reflections, collections of stories, and opinion pieces lacking methodology are generally more suitable for summaries or notes. Forcing skills will generate empty shell rules that look formal but cannot actually be triggered.

How to organize input materials

Books

Try to provide legally obtained, well-structured text and retain chapter titles. When scanning a PDF, you should first do an OCR spot check and correct headers, footnotes and line breaks, otherwise the extractor will treat noise as text.

Long video

Priority is given to using official subtitles; if there are no subtitles, phonetic transcription will be performed. Keeping timestamps helps to check the source and delete pre-roll advertisements, repeated oral broadcasts and interactions unrelated to the topic.

Podcasts & Interviews

Indicate the speaker. Methodologies may only represent an individual guest and should not be misattributed to the host or the show as a whole.

Dataset

Record the title, author, date, and permission of each source. Multi-source content especially requires distinguishing between shared conclusions and conflicting viewpoints.

RIA-TV++ seven stages detailed explanation

1. Overall content understanding

First do the four-step disassembly of structure, explanation, criticism and application to generate BOOK_OVERVIEW.md. This step is not a summary competition, but rather establishing the global context required for subsequent extraction.

2. Parallel extraction

Look for frameworks, principles, examples, counterexamples, and terminology respectively. Multiple perspectives can reduce the problem of extracting only eye-catching sentences and ignoring boundary conditions.

3. Triple verification

At a minimum, candidate methods need to be supported by multiple independent sources, have predictive power, and not be common knowledge. It is normal for a large number of candidates to be eliminated here.

4. RIA++ construction

Keep the original text basis, explain it in your own words, and supplement the book’s cases, future trigger scenarios, implementation steps, and applicable boundaries. A valid Skill must have both Practices and Restrictions.

5. Zettelkasten link

Establish dependencies, comparisons and combination relationships between Skills to generate INDEX.md. Otherwise dozens of orphaned Skills will just become another favorite.

6. Stress testing

Design normal questions, boundary questions and bait questions. Testing must not only prove that the skill will trigger, but also that it will not accidentally trigger when it is not applicable.

7. Delivery and installation

Generate DIGEST.md, glossary, Skill module and test files, and then install them into the Agent’s skill directory. Copyright, trigger range and script permissions still need to be manually reviewed before installation.

What should a good SKILL.md contain?

At least it should state:

1. Skill name and goal;
2. Clear trigger conditions;
3. Scenarios that should not be triggered;
4. Input requirements;
5. Executable steps;
6. Judgment and stopping conditions;
7. Output format;
8. Source and boundaries;
9. Test examples.

If a Skill only has a prompt word “Please analyze carefully and give suggestions”, it is still a general prompt word and is not a reliable execution module.

How to design test questions

Forward trigger

Give typical applicable scenarios and check whether the Agent recognizes them correctly and executes them according to the steps.

Negative trigger

Provide tasks that have similar keywords on the surface but are not actually applicable, and confirm that the Skill will not preempt other methods.

Boundary conditions

Remove critical inputs, create insufficient evidence, or conflict with goals, and check if the Agent will stop and request supplementary information.

Cross-Skill Confusion

Make both relevant methods appear to be available and check that the Agent can interpret the selection and combination order.

test-prompts.json Expected triggers, expected rejections, and acceptance points should be logged instead of just saving a few demo questions.

Review before installation into Agent

Skills generated from books and videos may also contain scripts or external links. Check before installation:-Whether a large section of copyrighted content is quoted;

• Whether to include personal privacy in the example;
• Whether it is required to read a directory that is too wide;
• Whether original materials will be uploaded;
• Whether the trigger description will cover unrelated tasks;
• Whether there is unverifiable author attribution;
• Test for inclusion of failed paths.

Version management and updates

When source material is updated, subtitles corrected, or methodology is overturned by new evidence, you need to know which Skills are affected. It is recommended to save source locations and versions in each module and maintain dependencies in INDEX.md.

Don’t silently rewrite skills that are already used for production decisions. Important updates should document the changes, reasons, and retest results.

Common failure modes

FAQ

Can I publish the entire book directly as a Skill?

There are copyrights and licensing to consider. Private study and public distribution are two different things. Public repositories should retain methodological extraction and necessary short citations to avoid copying large sections of copyrighted original text.

Why are there many Skills generated, but none of them are useful?

Usually the screening is not strict enough. Not every chapter is worth turning into a Skill. Prioritize the few methods that can be transferred across scenarios and have clear steps and boundaries.

How long of material should be processed at one time?

Depends on model context, text quality and structure. Extra-long materials should be processed according to chapters while retaining the overall overview; they cannot be simply cut into pieces and left in the dark.

Can methods from multiple authors be combined into one Skill?

Yes, but commonalities, conflicts and respective evidence must be noted. Don’t forcefully average conflicting methods into an empty principle.

Can the generated Skill be directly used for important decisions?

It should not be used without verification. High-risk content such as investment, medical, legal and security requires professional review, up-to-date information and clear boundaries of responsibility.

Practical example of making Skill from a video

Suppose the input is a two-hour interview on “How to review project failure.” Instead of directly asking to generate ten Skills, you can handle it like this:

1. Obtain subtitles and retain timestamps and speakers;
2. Delete advertisements and repeated greetings;
3. First generate the interview structure and main arguments;
4. Extract candidates such as “failure classification”, “evidence collection” and “action review”;
5. Check whether each candidate has multiple references in the interview;
6. Reducing a single character’s experience to a case and not treating it as a universal rule;
7. Supplement applicable project scale and stopping conditions for each method;
8. Design confusing questions such as “General Weekly Report”, “Accident Review” and “Employee Performance”;
9. Only keep modules that can correctly distinguish scenes;
10. Generate index, digest and source location.

You might end up with only three solid Skills, which is more useful than fifteen hint words that lack boundaries.

Quality score before delivery

Each candidate can be evaluated on a scale of 0–2:

Low-scoring candidates should be withdrawn or deleted; standards should not be lowered to maintain numbers.

How to distance yourself from the original author’s point of view

Skills should clearly distinguish between “original author assertion”, “extractor interpretation” and “verifiable operation”. For methods that are controversial or have obvious historical background, the source time and counter-evidence should be stated in the boundary. Agent should not package a book’s opinions into universal facts.

Summary

cangjie-skill is suitable for transforming “read but not used” content into Agent workflow. The key to quality is not the quantity generated, but the complete source, strict verification, clear triggers, and the courage to eliminate candidate skills that do not have independent value.