natural-japanese

Why we included this project

Anyone who drafts work documents in Japanese with an AI assistant has hit the same wall: output that is technically correct yet reads flat or unmistakably machine-translated. This Claude Code skill tackles that directly by loading writing rules into every stage of the pipeline. You settle on the audience, core message, and heading skeleton before drafting, generate under a twelve-article style constitution, then run a sudachipy-based linter that deterministically flags clichés, monotonous rhythm, and English-syntax calques. The human or the AI decides what to fix, and the loop keeps running until the lint findings converge. It ships templates for meeting minutes, reports, guides, essays, and slide outlines, and the linter scripts also run standalone, making them a useful reference for building AI-smell detection into your own tooling.