kangarooking/cangjie-skill

There is a growing gap between static knowledge repositories and the agents meant to act on them. This project tries to close that gap by taking a book or long document and converting its core ideas into discrete, callable skill units that an LLM agent can invoke at runtime rather than simply retrieve as raw text. The approach treats a book not as a chunk to be embedded but as a structured curriculum of reusable behaviors, which is a meaningfully different philosophy from standard RAG pipelines. For builders working on domain-specific agents, think legal research tools, medical reference assistants, or technical training products, this could reduce the gap between "the agent knows something" and "the agent can do something with that knowledge." The codebase and documentation are written in Chinese, which will create a real friction point for teams outside that ecosystem. It is not a plug-and-play library; expect to spend time understanding the extraction logic before adapting it. -> Best for: Agent developers building specialized vertical tools who read Chinese or are willing to navigate a translation workflow.