nc://ja-writing-guard

JA Writing Guard: scores Japanese text for AI-style phrasing, with line, column and a fix hint for each. A Nodecode rewrite of TakeshiTGAL/hermes-plugin-ja-writing-guard by TakeshiTGAL.

Get it

Tell the agent "add the ja-writing-guard cell", or run nodecode add nc://ja-writing-guard.

Set it up

Nothing is required: the cell scores every Japanese answer on its own, offline, and adds one italic line under it. All of it is in the ja-writing-guard section of /settings or config.jsonc; config.example.jsonc is that section written out.

• mode — report (the line under the answer) or off.
• surface — chat (default), business_email, external, sns or tech_doc: where the answers go. It scales each rule's weight, sets the threshold, and business_email and external also report 常体 where です・ます is needed.
• threshold — the 0-100 score at which a text counts as AI-like; 0 uses the surface's own (chat 40, business_email 35, external 40, sns 35, tech_doc 45).
• min_chars, footer, max_findings — how short an answer is skipped, whether the line is shown, and how many findings the tool returns.

Use it

Ask for a Japanese mail, notice, post or page and read the answer. Where it carries the phrasing that reads as unedited AI output, one line says so:

ja-writing-guard: AI的な言い回し 72/100（chat の基準 40）。指摘 3 件（定型句・常套句・誇張・空虚な言葉）

What it finds: stock closing phrases (〜と言えるでしょう, いかがでしたか), inflated words (シームレス, 重要な役割を果たします), English-calque grammar (〜することができます, 〜を行う), template shapes (3つのポイント, お役に立てれば幸いです), em dashes and Markdown where prose belongs, monotonous sentence endings, and 常体 where the surface needs です・ます. It checks wording, not authorship.

• /ja-writing-guard TEXT checks a text now; /ja-writing-guard business_email TEXT judges it as an outside email; bare, it shows the settings and the last report; off, on, reset control it.
• ja_writing_check — the tool: a whole draft in, the score, the threshold and every finding with line, column, excerpt and a fix hint out. Use it before sending a Japanese text and again after rewriting, to confirm the score went down.
• (help :ja-writing-guard) — the manual, for the model.

Code fences, inline code, URLs, quoted lines, HTML and text inside 「」 are masked out first, so a quotation or a code sample is never counted against the writer.

What is in it