ClaudeMap

nihongo

A Japanese-writing skill that brings AI-generated Japanese closer to how a person would actually write it — judging each sentence from reader, situation and writer instead of listing banned AI phrases

Skills & Commandsskilljapanesewritingproofreadingi18n

nihongo is an agent skill for Japanese output, shipped as two layers. Lite is the always-loaded file (ALWAYS.md, about 2,500 tokens) that applies during everyday replies; Full (SKILL.md plus references, about 10,000 tokens) loads only when you are writing something a human will read — email, announcements, job-application essays, documents, README copy, UI strings — and there is a Full proofreading mode that revises existing text without changing its meaning. Instead of a banned-phrase list it decides what a person in this situation would write: keeping set phrases of respect in announcements and messages to superiors while loosening only the casual exchanges, dropping unneeded commas, leaving already-natural sentences alone, and not adding opinions, reasons or promises the request never asked for while keeping figures, dates, conditions and who-does-what unchanged. A Personal feature builds a writing-style memo from three to five messages you actually sent on LINE or email, stored outside the skill so it survives updates, and scene tags keep the memo per context. It works with Claude Code, Codex, the Claude and ChatGPT apps, and Muse; releases also upload directly as app skills. Version 2.3.1 at review time.

What it does

  • Lite/Full split: a 2,500-token ALWAYS.md governs every reply, while the 10,000-token Full skill loads itself only for human-facing writing, with a separate Full proofreading mode for revising existing text
  • Judgment over bans: the skill reasons from reader, situation and writer — keeping honorific templates where the situation demands them, dropping superfluous commas, and leaving natural sentences untouched
  • No fabrication discipline: it must not add opinions, reasons or promises absent from the request, and figures, dates, conditions and who-does-what must stay unchanged
  • Personal style memo: paste three to five of your real LINE or email messages, get a per-scene style note stored outside the skill so it survives updates, and messages under your own name read like you wrote them
  • Multi-host coverage: Claude Code, Codex, the Claude and ChatGPT apps (release zips upload as skills) and Muse; updates via the skills CLI

When to use it

Use when your Japanese output must read as if a person wrote it — team Slack posts, customer mail, announcements, application essays. Skip it if you only ever write Japanese for yourself to parse: the value is in register and restraint, not correctness.

Getting started

Claude Code (install the skill, then load Lite every session):

npx skills add AnIT012/nihongo -g -a claude-code
echo '@~/.claude/skills/nihongo/ALWAYS.md' >> ~/.claude/CLAUDE.md

Omit -a claude-code to be asked where to install; project-only installs reference @.claude/skills/nihongo/ALWAYS.md from the project CLAUDE.md.

Codex: npx skills add AnIT012/nihongo -g, then paste ALWAYS.md into ~/.codex/AGENTS.md (or the project AGENTS.md).

Claude / ChatGPT apps: download nihongo-v<version>.zip from the GitHub releases page and upload it in the skills screen (ChatGPT: Skills > Create > Upload from your computer); a zip from the Code > Download ZIP button is nested one level too deep to upload.

Update: npx skills update nihongo -g

Things to know

The README states a prohibited-use rule up front: no fraud, phishing, impersonation or deception — and the skill itself must not write reasons, promises or figures the request never contained. Its published evaluation (10 prompts, blind AI judging) was run by the author's side, so treat the numbers as indicative only. Output is AI-judged and can vary between runs: check figures, dates and proper nouns yourself before sending. Personal-style memos hold personal message samples — the README suggests redacting names and contacts before pasting. MIT. 270 stars as of writing (2026-11-21).

Repository

AnIT012/nihongo

Charted

Related on the map

  • An instruction-only skill that makes agents research before they ask, plan before they build, and verify before they call anything done.

    • workflow
    • verification
    • quality
    • skill
  • Guizang's agent skill turns your repo's recent updates into a finished promo video — real components, code-rendered animation, an original code-written soundtrack — no video editor involved.

    • agent-skill
    • video
    • promo
    • marketing
  • An agent skill that geolocates a photo with nothing readable in it — OSM geometry, elevation skylines and satellite tiles turn “where was this taken?” into a measured answer with an error radius.

    • agent-skill
    • geolocation
    • osint
    • openstreetmap