japanese-ai-writing-proofreader

name	japanese-ai-writing-proofreader
description	Proofread Japanese prose in three passes: textlint (mechanical), AI-smell removal, and deep naturalness (sentence-end variety, rhythm, connectives). Use when asked to 「校正して」「推敲して」「AI臭を消して」「自然な日本語にして」, when polishing article drafts / PR bodies / docs written in Japanese, or as the proofreading phase of the article-writing skill.
argument-hint	<file or text to proofread>

Japanese AI-Writing Proofreader

Three passes, mechanical first. The textlint pass exists because deterministic checks beat LLM judgment where both can do the job — run it before reading the text yourself, so your review starts from machine-verified ground.

Mode

• report (default for text the user wrote): list findings with before/after suggestions; change nothing
• fix (default for Claude-drafted text, or when the user asks for 修正): apply the fixes, then summarize what changed by category. When the target is a file, edit it in place AND show the corrected text (or the changed passages, for long files) in the response; when it is pasted text, return the corrected text in full (no config probe needed — go straight to the fallback)

When unclear which applies, ask.

Scope guards

• Never touch code blocks, command output, quoted text (>), or verbatim excerpts — flag issues inside them at most
• Voice preservation: if the text has an intentional style (e.g. article-writing's style profiles — 括弧ツッコミ, 取り消し線, 「普通に」, casual fragments), those devices are NOT findings. Proofreading removes AI-smell, not personality. The protection covers the device itself, not errors inside it: real 誤字・誤用 inside a device are still findings; orthography preferences inside one are LOW at most. When invoked from article-writing, re-read the active style profile first.

Pass 1: textlint (mechanical)

Prefer the project's own setup when it exists:

# look for config in the target file's directory and its repo root, not your cwd
ls <target-dir>/.textlintrc* 2>/dev/null
npx textlint --format compact <file>            # or the repo's lint script

Otherwise run the bundled fallback config (no project pollution):

npx -y -p textlint \
  -p textlint-rule-preset-ja-technical-writing \
  -p @textlint-ja/textlint-rule-preset-ai-writing \
  -p textlint-rule-prh \
  -p textlint-rule-terminology \
  -p textlint-rule-ja-hiragana-keishikimeishi \
  -p textlint-rule-ja-hiragana-fukushi \
  -p textlint-rule-ja-hiragana-hojodoushi \
  -p @textlint-ja/textlint-rule-no-dropping-i \
  -p textlint-filter-rule-comments \
  textlint --config ~/.claude/skills/japanese-ai-writing-proofreader/assets/textlintrc.json \
  --format compact <file>

The fallback covers, beyond the two base presets: 表記ゆれ via the bundled assets/prh.yml dictionary (grow it when new variants surface in real drafts), English term casing (terminology), 漢字→ひらがな for 形式名詞/副詞/補助動詞, and い抜き言葉 (@textlint-ja/no-dropping-i).

• The fallback handles .md/.txt. For .mdx without a project config, copy to a temp .md first
• For bare text (not a file), write it to a temp .md and lint that
• Treat findings as advisories, not auto-applied truth: judge each against the intended voice. The fallback config already disables no-exclamation-question-mark and ja-no-weak-phrase because 「？」 and 「〜と思います」 are legitimate in this user's prose. Likewise, い抜き inside 括弧ツッコミ can be intentional — judge before fixing
• Findings that contradict an explicit repo/media style rule (e.g. no-mix-dearu-desumasu firing on a だ・である規約の記事) are residual: leave them, and note the count as intentionally残置 in the summary
• An intentional passage that keeps tripping rules can be fenced with <!-- textlint-disable [rule] --> … <!-- textlint-enable --> (the comments filter). Use sparingly and prefer the narrow per-rule form — never weaken the config itself, and do not weaken a project's textlint config to make findings go away

Pass 2: AI-smell (the 5 categories)

Aligned with ~/.claude/rules/japanese-writing.md. Hunt each category explicitly:

1. 機械的リスト: **ラベル**: 説明 templates, emoji markers (✅❌💡), identical structure forced onto every bullet → rewrite as prose or destructure the list
2. 誇張: 「革命的」「劇的に」「圧倒的」「究極の」 → replace with the concrete number or comparison, or delete
3. 過剰な強調: bolded adverbs, emoji-led headings, repeated 「！」 → keep at most the one emphasis that matters
4. 英語式コロン: 述語直後の「:」（「実装した: 〜」）→ receive with a noun or split the sentence
5. 冗長・受動: 「することができる」→「できる」、「〜を行う」→ verb form, passive → active, dropped particles restored（「設定変更」→「設定を変更する」）

This manual read is also where plain 誤字・誤用・文法ミス get caught — textlint misses many（「とゆう」等）. Report them under Pass 2 in the findings table.

Pass 3: Deep naturalness (what surface rules miss)

This is where text that passes Pass 1-2 still reads AI-written:

• 文末の単調: same ending 3+ sentences in a row（です。です。です。）→ rotate でした／ません／た。／体言止め／問いかけ
• 文長の均質: equal-length sentences in a flat march → vary; place a short punch sentence after a long explanation
• 接続詞の機械的連結: paragraph after paragraph opening with 「また」「さらに」「次に」 → cut (juxtaposition often suffices) or vary（ただ／そこで／つまり／一方で）
• 段落の均質: every paragraph the same size → merge or split by content
• 読点過多/過少: 一文に読点4つ以上は分割を検討。読点ゼロの長文は補う
• 情報の出し順: conclusion buried at paragraph end, examples before the point they illustrate → lead with the load-bearing sentence
• 翻訳調: 「〜することによって」「〜という形で」「〜の方(ほう)」の多用 → 直接的な構文に組み替える
• 未消化の専門用語・カタカナ英語: English or loanword jargon a general reader stumbles on → replace with settled Japanese, or gloss on first use（ドロップ→配線・ケーブル、ネゴシエーション→つながる・リンクする、ラジオ→帯域、律速→ボトルネック・頭打ち、Traffic Rule→トラフィックルール）. Keep field-standard terms as-is（API・PoE・VLAN・SSID）. textlint cannot catch this — read for it manually
• 初出の専門用語の導入: don't start using a term in the body before its alias/definition appears — a VLAN labeled Default in a table, then called Trusted in prose with no bridge, reads as 唐突 → introduce on first use as「A（＝B、その役割）」

Output

Findings table (report mode) or applied-fix summary (fix mode):

| Pass | 重要度 | 箇所 | Before → After |

• 重要度: CRITICAL（意味が壊れている・事実が変わる）/ IMPORTANT（誤字・誤用・文法ミス、および明確なAI臭）/ LOW（表記の好み）。 Present CRITICAL and IMPORTANT in the table. LOW: one aggregate line — count plus rule/kind names only（「ほか LOW 2件（訳→わけ等の表記）」), no Before→After. This keeps Pass 1 evidence visible without itemizing
• In fix mode, end with counts grouped by pass（textlint / AI臭 / 自然さ）and anything intentionally left alone (voice devices, quoted text)