InterfaceX-co-jp/genshijin

What it solves

Many AI coding agents use excessive tokens by generating polite, redundant, and wordy responses in Japanese. This project provides a set of communication skills and plugins to compress AI responses, reducing token usage by approximately 75% while maintaining 100% technical accuracy.

How it works

It optimizes Japanese communication by removing honorifics, filler words, cushion expressions, and redundant particles. It offers three levels of compression intensity:

• Polite Mode: Removes cushion words and vague expressions but keeps honorifics (suitable for business).
• Normal Mode: Drops honorifics and uses noun-ending sentences (the basic "caveman" style).
• Extreme Mode: Uses abbreviations, arrow notation, and one-word answers for maximum compression.

Beyond simple prompting, it integrates with AI agents via hooks (SessionStart and UserPromptSubmit) to prevent "style drift" and provides a middleware called genshijin-shrink to compress MCP server tool descriptions.

Who it’s for

Developers using AI coding agents like Claude Code, Cursor, Windsurf, Cline, or GitHub Copilot who want to lower their API costs and increase the efficiency of their context window.

Highlights

• Multi-Agent Support: Native rules and installers for Cursor, Windsurf, Cline, and GitHub Copilot.
• Token Savings: Benchmarked at an average of 83% reduction in Japanese token usage.
• Specialized Sub-skills: Includes dedicated modes for concise commit messages (/genshijin-commit) and one-line PR reviews (/genshijin-review).
• Memory Compression: A tool to compress CLAUDE.md or other memory files to permanently reduce input tokens for every session.
• MCP Middleware: Wraps MCP servers to shrink tool descriptions without losing technical identifiers.