The cangjie-skill repository is a Python-based system designed to distill high-value content from books, long-form videos, podcasts, and other sources into executable AI skills that can be called and composed by agents. Rather than creating summaries or notes that remain unused, the project transforms methodologies embedded in lengthy content into structured, actionable skill packages that AI systems can invoke in real-world scenarios.
The repository addresses a specific problem: people consume substantial amounts of content through books, videos, and podcasts but struggle to apply that knowledge in practice. Information typically remains at the consumption level—watched, read, or bookmarked—without being converted into tools that agents can actually use. The project's core insight is that systematic methodologies expressed in long-form content deserve extraction and transformation into callable tools, similar to how colleagues' expertise can be distilled into skills but focused on the structured knowledge they have deliberately developed and published.
The distillation process follows a seven-stage pipeline called RIA-TV++. The first stage applies Mortimer Adler's analytical reading method to understand the entire content's structure, interpretation, criticism, and application, producing an overview document. Five specialized extractors then work in parallel to identify candidate methodologies across frameworks, principles, case studies, counterexamples, and terminology. These candidates undergo triple verification: they must have at least two independent supporting references in the source material, demonstrate predictive power by answering questions not explicitly stated in the content, and represent genuinely unique insights rather than common knowledge. Typically only 25 to 50 percent of candidates pass this filtering stage.
Verified content is then restructured according to the RIA++ framework, which organizes information across six dimensions: original source references, rewritten interpretations, case studies from the source material, future trigger scenarios, executable steps, and boundaries or blind spots. The system identifies relationships between skills using Zettelkasten linking principles, generating connection maps and indexes. Each skill undergoes stress testing with specially designed test cases that include decoy questions and cross-skill confusion tests. Failed skills are reworked until they pass. The final output includes a reader-focused digest document, a comprehensive index, a glossary, individual skill modules, and test prompts for validation.
The repository has generated numerous skill packs from diverse sources. Book-based packs include distillations from Warren Buffett's shareholder letters, works on cognitive dividends, investment logic, copywriting, influence psychology, and classical texts like the Yellow Emperor's Inner Classic and Mao's selected works. Video-based packs have been created from Andrew Ng's AI for Everyone course and Loop Engineering video collections. The project maintains a curated list of 21 completed skill repositories with skill counts ranging from 6 to 25 per source.
The cangjie-skill project positions itself within a larger ecosystem that includes nuwa-skill for distilling human expertise and darwin-skill for evolving any skill over time. The system is designed to work with any content containing extractable, verifiable, and transferable methodologies, whether from traditional books, transcribed videos with subtitles, podcasts with transcripts, interviews, lectures, courses, long-form articles, or resource collections. The repository recommends pairing video distillation with a video-downloader skill to first extract subtitles, audio transcriptions, and key materials before applying the methodology extraction and skill generation process.
