Making AI-Sounding Japanese Natural — the natural-japanese Skill
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
Original article: AIっぽい日本語を自然に直す ― natural-japaneseスキル (cccskills)
Skill page: natural-japanese
What you'll learn from this article
score returns only a naturalness score without rewritingYou handed in meeting minutes or a report written by AI, and someone said, "You used AI to write this, didn't you?"
Sound familiar?
natural-japanese is a skill for writing and revising Japanese business documents so they read naturally.
Suspicious words and literal-translation phrasing are detected mechanically, while the agent judges each case in context to decide whether to leave it or fix it.
Word order and the placement of commas are also subject to review.
This article covers what the skill can do, how to install it, example prompts, and things to watch out for.
natural-japanese is a skill for writing and revising Japanese business documents so they read naturally.
SKILL.md lists the following as targets: meeting minutes, research and analysis reports, internal guides and manuals, research memos, project proposals, proposals, reports, emails, and slide outlines.
It can also be used to produce meeting minutes from transcripts, write note or blog articles and essays from scratch, or rewrite existing text.
Before starting to write, the skill reads a "template" file prepared for each document type.
For meeting minutes it's references/doctypes/minutes.md, for research reports report.md, for slide outlines slide.md, and so on.
For text that doesn't fit a template, such as note or blog posts, this step is skipped.
Work proceeds in the order: design → draft → check → finalize.
The first things to determine are the intended reader and the main message summarized in a single sentence.
Next, a skeleton of headings is built and checked to ensure the story holds with headings alone before the body is written.
If the main message can't be expressed in one sentence, the skill treats this as a sign that the material is insufficient rather than a writing problem.
The core principle of this skill is "detection by machine, judgment by AI."
Because AI tends not to notice its own habits, suspicious passages are flagged mechanically by the bundled Python script scripts/lint.py.
Detection targets include banned words, translation-style phrasing, repeated negative-positive contrasts, uniform sentence length, and similar patterns.
Sentence-ending noun phrases, a high ratio of conjunctions at paragraph openings, and literal translations that follow English sentence structure are also caught.
However, the detection results are not an instruction to "fix everything."
The agent reviews each item in context and records whether it was "fixed" or "kept (with reason)."
A second core principle is that it's better to prevent AI-like patterns from appearing in the first place through constraints during writing, rather than removing them after the fact.
The 12 rules in references/writing-constitution.md serve as those constraints.
These rules include things like: start with the conclusion; make headings messages that contain the conclusion; use bullet points only for items that are truly parallel; and don't repeat the same sentence structure three times in a row.
Starting from the idea that giving every section equal weight is itself unnatural, the skill also calls for "density" design — writing important sections in depth and lighter sections more briefly.
For readability, the skill addresses principles such as word order, comma placement, one idea per sentence, and the distance between subject and predicate.
Parts that are hard to judge mechanically are checked visually using the catalog readability-antipatterns.md.
This catalog classifies 27 types of bad writing patterns in order of how much they burden comprehension.
You can also request only a diagnosis without any rewriting.
Specifying score returns a naturalness score (0–100, higher meaning less AI-like) and the reasons for it, without touching the document.
There are two modes — Quick and Full — with Quick as the default.
According to the rough benchmarks in SKILL.md, short documents around the length of meeting notes finish in about 30 seconds, and even documents around 10,000 characters take roughly 3 minutes.
Full mode is used when you ask for something "thorough" or say "take your time."
It is also used for high-stakes documents like those intended for external or executive audiences, or for very long documents exceeding 10,000 characters.
In Full mode, structural and readability reviews are conducted in parallel by sub-agents.
The skill uses a 6-axis evaluation framework and keeps revising until all axes score 90 or above.
Estimated time is around 7 minutes for short documents and 15–20 minutes for 10,000-character documents.
First and foremost, it suits people who have been told their AI-written text "sounds like AI" or "reads like machine translation."
If you're bothered by phrases like "〜することができる" or "と言えるだろう," or by paragraphs that all have the same rhythm, you can delegate both detection and judgment to the skill.
It also suits people who want to get the basics of business writing right.
The skill explicitly covers responding to instructions like "start with the conclusion," "keep headings concise," and "explain technical terms clearly."
If you frequently write meeting minutes or reports, it's useful regardless of your role.
People who want text written in their own voice are also targeted.
Feed the skill 3–5 samples of your past writing and it will extract tendencies in perspective, vocabulary, and rhythm into style-profile.md, following the template in assets/style-profile-template.md.
This file then serves as the foundation for subsequent writing.
On the other hand, the skill is not suited for structuring chapters in technical documents or formatting Markdown.
SKILL.md states that formatting tasks such as one-sentence-per-line conversion, blockquotes, and footnote syntax are out of scope and belong to a different skill.
The distribution source is coji/natural-japanese on GitHub.
Run the following command in your terminal to install it.
npx skills add https://github.com/coji/natural-japanese --skill natural-japanese
If you'd like Claude Code to install it for you, send the following message as-is.
次の Agent Skill をインストールしてください。
https://github.com/coji/natural-japanese/tree/main/skills/natural-japanese
まず SKILL.md と付属ファイルの中身を確認し、不審な処理（外部への送信、認証情報の読み取りなど）がないことを確かめてから、~/.claude/skills/natural-japanese/ に配置してください。
There are two places to put the skill.
To use it across all projects, place it in ~/.claude/skills/; to use it only in a specific project, place it in that project's .claude/skills/.
There are also two ways to invoke it.
You can call it as a command like /natural-japanese, or simply ask in natural language.
The skill's description is written so that it also responds to requests like "make this more natural Japanese" or "give me an AI-ness score."
The first example is a request to create meeting minutes from a transcript.
この会議の文字起こしから議事録を作ってください。
読むのは会議を欠席した部長です。
決まったことと宿題を先に書いてください。
The skill will draft the minutes using the appropriate template, run a lint check, and return the finished minutes.
The second example is rewriting a published article.
このブログ記事を「AIで書いたでしょう」と言われました。
もっと自然な日本語に直してください。
What comes back is a revised version where detection results have been judged and addressed one by one.
Applying the same type of correction uniformly across the whole text actually makes it smell more like AI.
SKILL.md takes this view, and adopts a policy of fixing only the places where it can add real value.
The third example is requesting only a diagnosis without rewriting.
/natural-japanese score quick draft.md
In natural language, "この文章、どれくらいAIっぽいか判定して。書き換えはしないで" works the same way.
You'll receive a naturalness score from 0 to 100 and the reasons behind it.
Rewrites may be suggested, but the text won't be changed until you ask.
Scripts such as scripts/lint.py used for checking assume they are run with uv run.
In environments where uv is not available, the skill switches to a method of manually going through the same checkpoints following references/manual-checklist.md.
SKILL.md lists Claude.ai as an example of such an environment.
scripts/semantic.py, used with score exp, is an experimental detector.
It depends on torch and sentence-transformers, and downloads a model of approximately 1 GB on first run.
It is not included in the standard check and is treated as an optional addition for full workflows or when the environment allows.
Since the skill includes Python scripts, it's a good idea to review the contents before installing.
The license is MIT.
Intermediate files created during the task are designed to be deleted upon completion.
These include judgment logs, lint JSON output, and draft backups.
Only the finished document and, if requested, style-profile.md will remain.
yomiyasu is a skill for proofreading AI-generated Japanese.
It targets technical articles, specifications, PR descriptions, internal reports, and essays.
It is described as correcting unnatural metaphors and ambiguous subject-predicate relationships while preserving meaning and conditions.
copy-editing is a skill for editing and reviewing marketing copy.
It's a candidate to reach for when you want to revisit ad copy or landing page text.
slide-story is a skill covering how to structure a story in slides for presentations or lectures.
It's worth reading alongside this skill when building a slide outline.
Yes.
Emails are included as a target, and short documents are processed in the default Quick mode.
According to SKILL.md's estimates, text around the length of meeting notes takes about 30 seconds.
Feed it 3–5 samples of your past writing and it will summarize your stylistic tendencies in style-profile.md.
This file then serves as the foundation for subsequent writing.
If you tell it there's a particular expression you "don't use," it will be recorded as a phrase to avoid.
Specifying score or asking it to "grade the AI-ness" will perform only a diagnosis.
You'll receive a naturalness score and reasons, and no rewriting will occur unless you ask.
Markdown formatting tasks such as one-sentence-per-line conversion, blockquotes, and footnote syntax, as well as chapter structuring for technical documents, are out of scope.
This skill focuses exclusively on the naturalness, readability, and clarity of text.
natural-japanese is a skill specialized in writing and revising Japanese business documents for readability.
Banned words and literal-translation phrasing are detected mechanically, and the agent decides in context whether to fix each one.
This makes it less likely to stop at superficial word substitutions.
From drafting meeting minutes and reports, to rewriting blog articles, scoring AI-ness, and learning your personal style — you can delegate the full range of tasks related to natural writing.
Start with Quick mode and try it on a document you have on hand.
For more details, see the cccskills page.
This article is a repost of the cccskills piece "AIっぽい日本語を自然に直す ― natural-japaneseスキル."
Full details about the skill and installation instructions can be found on the cccskills natural-japanese page.