natural-japanese turns Japanese style rules into an Agent Skill

Large language models often produce Japanese business documents with vague headings, uniformly sized paragraphs, stock transitions, and heavily hedged conclusions. The open-source natural-japanese project addresses those patterns during drafting and revision. Its GitHub repository has passed 1,100 stars.

An Agent Skill packages instructions and tools that an agent can invoke for a specific task. Natural-japanese ships in that format for Claude Code and compatible agents, covering meeting minutes, research reports, internal guides, proposals, and blog posts. Its workflow applies editorial constraints before generating a draft, then checks the result with language-aware scripts.

Lock the outline before drafting

Post-draft requests to “sound more human” often replace one stock phrase with another because the model evaluates prose shaped by its own habits. Natural-japanese divides the work into planning, drafting, linting, and review so those habits can be addressed at several stages.

The first phase fixes the heading structure and applies a style constitution containing 12 drafting rules. Requirements such as leading with the conclusion and varying sentence patterns shape the initial text before formulaic structures spread through the document.