AnIT012/nihongo
ProductivityAn agent skill that rewrites AI-generated Japanese into natural, human-sounding prose for email, Slack, docs, and other writing.
Overview
nihongo is an agent skill for Claude, ChatGPT, Claude Code, Codex, and compatible skill systems that makes AI-written Japanese read more like human writing. Instead of relying on a banned-word list, it gives the model a decision procedure based on the reader, purpose, medium, and writer, so the text fits the situation rather than merely avoiding stock phrases. It covers emails, announcements, Slack posts, READMEs, entry sheets, opinion pieces, and similar output.
The skill is split into two parts. Lite (nihongo/ALWAYS.md, roughly 1,500 characters) holds principles for everyday replies and is loaded every turn. Full (nihongo/SKILL.md plus references/) contains the detailed, situation-specific guidance and is read when writing text for other people, either automatically via Lite or by invoking it by name. The estimated token cost is about 1,100 for Lite and about 9,000 for Full.
A Python script in nihongo/scripts/ can compare text before and after revision and revert sentences where numbers, negation, certainty, strength, assignees, or times changed. The skill still works without Python.
Getting started
With Node.js available, install the skill globally:
npx skills add AnIT012/nihongo -g -a claude-code
Drop -a claude-code to choose the target AI interactively; if you are using Claude Code, select Claude Code in the list, otherwise the skill lands in ~/.agents/skills/ and Claude Code will not see it.
For Claude Code, add this line to ~/.claude/CLAUDE.md so Lite loads every session:
@~/.claude/skills/nihongo/ALWAYS.md
If installed without -g, reference @.claude/skills/nihongo/ALWAYS.md in the project's CLAUDE.md instead. For Codex, paste the contents of ALWAYS.md into ~/.codex/AGENTS.md or a project AGENTS.md. For the Claude or ChatGPT apps, upload the nihongo/ folder as a zip; in ChatGPT use Skills → Create → "Upload from your computer", then place the Lite content in the app's instruction field.
Full is used automatically when a task calls for writing to people; otherwise ask with something like "nihongoで直して", or send /nihongo in Claude Code. Optional meaning-preservation checks run through the scripts under nihongo/scripts/.
Constraints
- The core is Markdown guidance interpreted by an AI, so output is not deterministic even for the same request.
- The Python scripts only check for meaning changes after revision; they are optional and depend on the standard library.
- Content correctness is not guaranteed: verify numbers, dates, and proper nouns before sending.
- Evaluation is reported at version 2.1.3 with Claude Sonnet 5.5 as writer and Claude Opus 5.5 as judge; ChatGPT and Codex models have not been evaluated the same way, and judging was done by the author or an AI from the same company, so the figures are indicative rather than independent benchmarks.
- MIT licensed.
HotRepo's take
On the 2026-11-18 new-project list nihongo recorded 109 stars with 9 gained and 2 forks, plus one fork gained in the measured window; it has no recorded all-language daily rank. That is a small, early sample, so it supports no trend claim. What the repository does provide is a concrete design choice worth noting: the Lite/Full split keeps everyday token overhead near 1,100 while keeping the detailed procedure for high-stakes writing, and the author has published the evaluation prompts and win/loss records under evaluations/ for inspection.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- conorbronsdon/avoid-ai-writingAn agent skill that audits and rewrites prose to remove AI writing patterns, with a deterministic detector and CI gate.
- blader/humanizerAn agent skill that rewrites AI-sounding prose to read more naturally without inventing facts.
- Dodaios/ste80-humanizerAn agent skill that strips AI writing tells and rewrites text toward Karpathy's 80% ASD-STE100 style.
- AminBlg/SimpleEnglishAn agent skill that makes LLMs write documentation and replies in ASD-STE100 Simplified Technical English.
- duyunzhoudyz/dyz-med-paper-skillsA 32-skill Agent Skills pack for Chinese clinicians and medical graduate students to plan, draft, translate, and revise medical research papers.