docx-redline-ja

A skill that writes AI proofreading results for Japanese text into a Word document as tracked changes. It diffs character by character, so each tracked change covers only the characters that changed.

View on GitHub Install

Three rows: the original sentence, docx-cli, and this skill. docx-cli strikes out whole phrases and reinserts them; this skill marks only the deleted の, で and 画像上 and the lowercased C and R. — The same two findings applied with docx-cli, which diffs by word, and with this skill. Both rows are drawn from real output.

The problem with Japanese

Ask an AI such as Claude to proofread a manuscript and the suggestions come back as chat text. Applying them by hand in Word takes a while, and afterward nobody can tell what was changed.

Tools that write suggestions as Word tracked changes exist, but most assume a language with spaces between words. Japanese has none, so everything up to the next punctuation mark is treated as one word, and fixing a single character deletes and reinserts all of it. The reviewer cannot see what was actually fixed.

docx-redline-ja diffs character by character. Each tracked change covers only the characters that changed, so the author can check and accept or reject each correction in Word's review pane.

What it does

Reads a JSONL file of old_text/new_text pairs and writes them into a copy of the .docx as native Word tracked changes. The original is never modified.
Before writing, checks that rejecting every change restores the original text and that other people's tracked changes are intact. If any check fails, nothing is written.
Skips a finding that matches more than one place, and says why.
Edits inside table cells and text boxes, and handles superscripts and subscripts such as reference numbers.
Needs only Python 3.11+ and lxml. It runs locally and sends no document content anywhere.

Benchmark: cost and tokens

Claude Code (Sonnet 5.5) was given a 40-paragraph document and 20 edit instructions and asked to produce the redlined .docx. Means of 3 runs.

Language	Tool	Cost (USD)	Input tokens	Marked-up chars
Japanese	This skill	0.060	51,933	1.0x minimum
Japanese	None (model alone)	0.068	70,370	1.0x
Japanese	Anthropic's `docx` skill	0.079	74,055	1.0x
Japanese	docx-cli	0.080	92,508	28.1x
English	This skill	0.060	56,726	1.0x
English	None (model alone)	0.096	86,113	1.1x

In both languages this skill used the least money and tokens of the four. Against the model alone it cost 12% less in Japanese and 37% less in English. The full table and method are in the README.

Install

Claude Code

/plugin marketplace add akchan/docx-redline-ja
/plugin install docx-redline-ja@docx-redline-ja

Codex CLI

codex plugin marketplace add akchan/docx-redline-ja
codex plugin add docx-redline-ja@docx-redline-ja

Claude Desktop (Cowork)

Open Customize in the sidebar, then Plugins. Select Add marketplace, enter akchan/docx-redline-ja, and install docx-redline-ja.

Usage

Once installed, ask the agent something like:

Proofread manuscript.docx, collect the findings in findings.jsonl, and apply
them with docx-redline-ja as tracked changes.

Findings are JSONL, one per line. The script also runs on its own:

{"old_text":"所見がみられか？","new_text":"所見がみられるか？"}

python3 plugins/docx-redline-ja/skills/docx-redline-ja/scripts/apply_redline.py \
  --docx manuscript.docx --findings findings.jsonl --out reviewed.docx

What it does not do

It does not proofread. An agent or a person decides what to change; this skill writes those decisions into the document.
It edits body paragraphs only, not comments, headers, footers or footnotes.
Each edit must stay within one paragraph.
It does not add comments.