DOCX (Windows + Cyrillic)

Two CLI scripts on top of Python (read) and docx-js (write). Works on Windows, doesn't crash on cp1251, doesn't require pandoc / soffice.

When to use

• The user attached a .docx file and wants its content extracted
• Need to generate a letter / commercial offer / spec / report as .docx
• Need to save your work as a real Word document, not markdown

For tracked changes, comments, or find-and-replace inside existing documents - use the official anthropic-skills:docx instead.

Reading

python "<SKILL_DIR>/scripts/read_docx.py" "/absolute/path/to/file.docx"

Where <SKILL_DIR> is typically:

• Linux / macOS: ~/.claude/skills/docx-ru
• Windows: %USERPROFILE%\.claude\skills\docx-ru

Output is plain text split by paragraphs (UTF-8 stdout). Add --tables to render tables as pipe-separated rows: cell | cell | cell.

Creating

1. Write a JSON config with blocks (see schema below) using the Write tool. A typical scratch path is ./_tmp_docx.json.
2. Run the builder:

node "<SKILL_DIR>/scripts/build_docx.js" --config "./_tmp_docx.json" --out "./result.docx"

The script automatically replaces em-dashes and en-dashes with regular hyphens so generated text doesn't look AI-written. Pass --keep-dashes to disable.

JSON config schema

{
  "font": "Arial",
  "fontSize": 22,
  "pageSize": "A4",
  "margins": 1134,
  "blocks": [
    { "type": "h1", "text": "Heading 1" },
    { "type": "h2", "text": "Heading 2" },
    { "type": "p",  "text": "Regular paragraph", "bold": false, "italic": false, "center": false },
    { "type": "pRich", "runs": [
        { "text": "Bold fragment: ", "bold": true },
        { "text": "and the rest in plain text." }
      ]
    },
    { "type": "bullet", "text": "Bulleted list item" },
    { "type": "number", "text": "Numbered list item" },
    { "type": "divider" },
    { "type": "image", "path": "./logo.png", "width": 300, "center": true },
    { "type": "table",
      "widths": [3600, 2700, 2700],
      "header": ["Column 1", "Column 2", "Column 3"],
      "rows": [
        ["cell", "cell", "cell"]
      ],
      "firstColBold": true,
      "shadeHeader": "E8E8E8"
    }
  ]
}

All fields except blocks are optional. Defaults: Arial 11pt, A4, 2 cm margins, tables with grey header and thin grey border. Column widths are in DXA (1440 = 1 inch, 1134 = 2 cm). The table width is the sum of widths.

For image blocks: path is required, width/height are in pixels (if only one is given, the other scales proportionally from the file's natural size; if neither, natural size is used). Supported formats: PNG, JPG, GIF, BMP — type is detected by extension and magic bytes. center aligns the image; before/after set paragraph spacing in twips.

A complete example is in examples/basic.json next to this file.

Dependencies

Must be installed globally on the user's machine (skill assumes they exist):

• Node.js + npm i -g docx (tested with docx@9.6.1)
• Python 3 with built-in zipfile

Working rules

• For a new .docx, always use this skill instead of writing an inline JS-script with docx-js boilerplate from scratch
• For reading an existing .docx, try read_docx.py first; only escalate to anthropic-skills:docx if you actually need tracked changes / complex formatting
• In Windows paths, use double backslashes inside JSON ("C:\\Users\\...") or forward slashes
• After creating a .docx, verify file size with ls -l; a normal document is > 8 KB

По-русски

Облегчённый skill для чтения и создания .docx на Windows с нормальной поддержкой русского. Альтернатива anthropic-skills:docx, когда не нужны tracked changes / find-replace.

Чтение: python <SKILL_DIR>/scripts/read_docx.py file.docx [--tables] - выдаёт plain text с разбивкой по параграфам, опционально рендерит таблицы. UTF-8 stdout всегда (на cp1251 не падает).

Создание: написать JSON-конфиг (схема выше) → node <SKILL_DIR>/scripts/build_docx.js --config spec.json --out result.docx. Автоматически заменяет em-dash и en-dash на обычный дефис, чтобы текст не выглядел сгенерированным ИИ. --keep-dashes отключает.