cangjie-skill

Why we included this project

Anyone who keeps a growing pile of books, lecture videos, and podcast episodes they intend to get back to will recognize the problem cangjie-skill targets: all that material sits in notes and watch-later lists, never turning into something an agent can use. The tool feeds full text through a seven-stage pipeline that pulls candidate methods out in parallel, filters them with three checks (two independent corroborations inside the source, ability to answer questions the content never states, and not being common sense; usually only a quarter to half make it through), then rebuilds the survivors as skill modules with trigger conditions, executable steps, and boundary notes. Those modules install into Claude Code or OpenClaw, so the knowledge stops being a saved file and starts being something an assistant reaches for during real work. Each skill also gets stress-tested with decoy questions and cross-skill confusion cases, and the pipeline emits a digest for readers who want the takeaways without rereading the source. It is a distillation process rather than a summarizer, and the strict filtering is what keeps the output useful.