coji / natural-japanese
A Claude Code skill that mechanically detects and prevents 'AI-smell' in Japanese business writing through a writing constitution and sudachipy-based linting.
星标趋势
AI 分析
项目摘要
A Claude Code Agent Skill that helps write and revise natural-sounding Japanese business documents by preventing 'AI-smell' patterns. It combines a 12-rule writing constitution applied before writing with a sudachipy-based linting script that mechanically detects clichés, monotonous rhythms, and unnatural expressions.
为什么值得关注
Addresses a real and widespread pain point — AI-generated Japanese text has distinctive, off-putting patterns that are hard to fix after the fact. The two-phase design (prevent via constraints, detect via mechanical linting, judge via agent) is a thoughtful architectural response, and the project has gained significant traction with 1,903 stars and 12 releases in six months.
优势
- Well-designed two-phase architecture: writing constitution for prevention + mechanical linting for detection
- Clear before/after examples demonstrating tangible quality improvement
- Active maintenance with 12 releases in 6 months and CI pipeline
- Multiple installation methods (npx skills add, openskills, Claude Code plugin, manual)
- Thoughtful philosophy of mechanical detection with human/agent judgment for nuanced decisions
局限性
- No automated tests despite having CI
- Japanese-only, limiting audience to Japanese speakers
- Single contributor, creating bus-factor risk
- No Docker support or standalone examples directory
使用场景
- Writing meeting minutes, internal reports, and company guides in natural Japanese
- Revising AI-generated Japanese drafts to remove robotic patterns
- Diagnosing and scoring 'AI-smell' in existing Japanese documents
- Blog post and essay drafting with more human-like tone