coji/natural-japanese

What it solves

AI-generated Japanese text often suffers from a distinct "AI smell"—characterized by generic headings, uniform paragraph lengths, and overly safe, bland conclusions. natural-japanese provides a framework to eliminate these patterns and produce professional, readable, and natural-sounding Japanese business documents (such as meeting minutes, reports, internal guides, and planning documents).

How it works

The project employs a two-pronged approach to ensure quality:

1. Preventative Writing: Instead of trying to fix text after it is written, it establishes a structural skeleton and applies a "Writing Constitution" consisting of 12 rules (e.g., "start with the conclusion") to prevent AI patterns from appearing in the first place.
2. Mechanical Detection: It uses a Python script (lint.py) powered by sudachipy (morphological analysis) to mechanically identify cliché phrases, monotonous rhythms, and translation-style Japanese.

Rather than automatic replacement, the system allows humans or agents to decide which flags to fix based on context, and uses a "6-axis revision rubric" to iteratively refine the text until it meets quality standards.

Who it’s for

Professionals writing business documents in Japanese who want to avoid the generic, robotic tone of LLMs and ensure their writing is clear, concise, and natural.

Highlights

• Agent Skill Integration: Designed as an Agent Skill that can be installed into Claude Code, Cursor, or ChatGPT.
• Rule-Based Linting: Uses morphological analysis to detect specific "AI-smelling" patterns.
• Diagnostic Tooling: Includes a /natural-japanese score command to measure the naturalness of a document on a scale of 0-100.
• Comprehensive Analysis: Provides scripts to check reading load (sentence length, punctuation bias), extract outlines, and verify technical terminology explanations.