kangarooking/cangjie-skill

Cangjie Skill is a Chinese-language project that turns the reasoning found inside books, long videos, and podcasts into structured AI skills that an agent like Claude Code can actually call on. The idea builds on a related project called nuwa skill, which distills a specific person's way of thinking into a usable skill. Cangjie Skill focuses instead on distilling the systematic reasoning that a person has already written down or spoken at length, such as in a book, a long interview, a podcast episode, or a long form video. The problem it addresses is a common one. People read many books, watch many videos, and listen to many podcasts, but struggle to actually apply what they learned at the moment a real decision needs to be made. Summaries and notes tend to compress information without making it something an agent can reliably call up and use. The tool works through a seven stage pipeline it calls RIA-TV++. It first analyzes a whole piece of content using a structured reading method, then runs five parallel extractors that pull out candidate frameworks, principles, cases, counterexamples, and terminology. Each candidate has to survive three checks, including independent verification within the source and the ability to answer new questions the source never directly addressed. Only 25 to 50 percent of candidates typically pass. Surviving material is then built into a structured skill file with clear sections for the original quote, a rewritten explanation, book examples, future trigger scenarios, executable steps, and known boundaries. The project also generates a stress test set of trick questions for each skill and links related skills together. The repository already includes many generated skill packs distilled from sources such as Warren Buffett's shareholder letters, Robert Cialdini's Influence, and various Chinese business and philosophy texts, each represented as its own linked GitHub repository. The project is licensed under GNU AGPL v3 and is built to work with platforms like Claude Code and OpenClaw.