Word documents

You already know Word's features. This skill adds what you can't see from the code: a library that makes tracked changes and comments correctly, validation that catches untracked edits, a way to see pages as a reader will, and the docx-js and OOXML behaviours that silently produce broken files.

Running scripts

<skill> is the directory containing this file. Run bundled scripts by absolute path from the user's working directory, so relative paths to their files keep working:

uv run <skill>/ooxml/scripts/unpack.py contract.docx unpacked

Run Python you write with its dependencies declared, either uv run --with defusedxml script.py or a PEP 723 header. Run Node scripts as NODE_PATH=<skill>/node_modules node build.js; if <skill>/node_modules is missing, run npm install in <skill> once. Rendering pages needs LibreOffice (soffice) and poppler (pdftoppm); reading needs pandoc.

Choosing a workflow

Redlining workflow

1. Read. Convert the document with pandoc (above) and list every change you will make. Locate each one by its surrounding text, never by markdown line numbers, which have no relation to the XML.
2. Unpack. uv run <skill>/ooxml/scripts/unpack.py contract.docx unpacked prints a suggested RSID. Pass it as Document(..., rsid=...) so the session's edits share one revision ID.
3. Edit. Read ooxml.md, then write a script that uses the Document library. tracked_replace(old, new) handles most edits: it finds text across runs and marks only the words that differ. Work in batches of related changes, a few to a dozen per script, so a failure is easy to locate. Grep unpacked/word/document.xml again before each batch, because line numbers move after every save.
4. Save. doc.save() refuses to write a document with untracked edits, or with XML that breaks the schema, and prints what differs. Fix the script rather than skipping validation.
5. Pack. uv run <skill>/ooxml/scripts/pack.py unpacked reviewed.docx repacks the file and checks that LibreOffice can open it.
6. Verify. Run pandoc --track-changes=all reviewed.docx -t markdown and confirm every intended change appears and nothing else changed. Then check the output (below).

To accept or reject existing changes, call doc.accept_all_changes() or doc.reject_all_changes() before making edits of your own.

Checking the output

1. Validate. uv run <skill>/ooxml/scripts/validate.py reviewed.docx --original contract.docx checks the schema, and that every change from the original is tracked; doc.save() runs the same checks. Pass --author NAME to let that author's earlier insertions be edited directly. For a document generated from scratch, leave out --original to check the schema alone.
2. Look. uv run <skill>/ooxml/scripts/render.py reviewed.docx pages prints one PNG path per page (--pages 2-4 for a range). Read them before delivering a document you created or restructured. LibreOffice renders a docx-js table of contents empty; Word fills it in when it opens the file.

Pitfalls that fail silently

• Partial edits inside other authors' changes. To remove text inside someone else's insertion, nest your <w:del> inside it. Never edit their text directly.
• pandoc as an editor. It rewrites the whole document and drops formatting, comments and tracked changes. Use it to read, never to produce the deliverable.