Document DOCX Skill - Quick Reference

This skill covers the specialist .docx lanes: docxtpl template pipelines, python-docx structural edits and extraction, LLM/RAG ingestion, and accessibility/EAA compliance.

Boundary: when the official Anthropic docx skill is available, prefer it for plain create/edit (docx-js), tracked-change (redline) authoring via <w:ins>/<w:del> with validate.py, and comment authoring. When the user asks for "a document" without naming a format and a document connector or artifact type is available, prefer that over generating a .docx.

Core Decision Rules

• Non-developers own the layout in Word, and code supplies data: docxtpl. Code decides the structure: python-docx (Python) or docx (Node).
• Render templates fail-closed. docxtpl's defaults render a missing variable as blank and insert &/< unescaped, so the file opens fine and passes a tag scan. Use scripts/docx_render_template.py (StrictUndefined + autoescape); see references/template-workflows.md.
• Tracked changes: never promise high-level library support, and never fake redlines with colour or strikethrough. Route authoring to the official docx skill's OOXML workflow. See references/tracked-changes.md.
• Semantic HTML from a trusted file: mammoth, then sanitise (it does no sanitisation). Markdown/JSON for RAG: a Markdown converter or scripts/docx_extract.py; see references/llm-extraction-workflows.md.
• DOCX feeding a KB or index: tables inside content controls are silently missing from Document.tables, tables pasted as images have no text, tracked changes need an explicit accept/original policy, and filenames can lie. Run the table-count detection check in references/docx-extraction-failure-modes.md before indexing any file (load it whenever DOCX is ingestion input).
• PDF output: Word automation for the highest fidelity; LibreOffice headless for CI and batch.
• .doc: convert to .docx first. .docm/.dotm: treat the macros as untrusted and never execute them.
• .docx is the editable source; PDF is the release artifact.

Quick Reference

Task	Tool	When
Template fill, mail merge	docxtpl + scripts/docx_render_template.py	Word-authored templates, batch documents
Structural edits in Python	python-docx (references/docx-patterns.md)	Styles, tables, sections, fields, hyperlinks
Comments	python-docx add_comment (version-gated) or the official docx skill	Review notes without revisions
DOCX to HTML	mammoth via scripts/docx_to_html.mjs	Trusted documents, semantic HTML
DOCX to JSON/Markdown	scripts/docx_extract.py, Markdown converters	Automation, RAG ingestion
Inspect revisions/comments	scripts/docx_inspect_ooxml.py	Before editing any reviewed document
Release gate	scripts/docx_quality_gate.py	Every generated or edited file
DOCX to PDF	Word automation / LibreOffice headless	Release artifacts, CI smoke tests

Format And Safety Caveats

• python-docx comments: main body only, no anchors in headers or footers, no threaded replies or resolved state, and anchors are whole runs.
• python-docx has no tracked-change authoring, no alt-text property, no field API and no public hyperlink authoring API. The workarounds in the references use private attributes; re-verify them after upgrades.
• A TOC and other Word fields stay empty or stale until updated in Word or rendered by LibreOffice.
• New python-docx documents can inherit author, date and comments from the default template. Inspect core properties before distribution.

Version-Gate Before Promising A Feature

Do not assume the environment has a current library:

• Document.add_comment() is a recent python-docx addition and raises AttributeError on older pins. Check hasattr(docx.document.Document, "add_comment") before promising comment support.
• If a user asks for something flagged unsupported here (tracked-change authoring in python-docx, threaded replies, resolved comments), say so plainly and route to the official docx skill. Do not approximate it with formatting.

Default Workflow

1. Identify the file type and trust level: .docx/.dotx, .docm/.dotm, or legacy .doc.
2. If the document may carry review metadata, run scripts/docx_inspect_ooxml.py --json first.
3. Generate or modify: docxtpl through docx_render_template.py, or python-docx/docx for structure.
4. For a newly generated output, set doc.core_properties to the intended author, clear template comments, and set created/modified to datetime.now(timezone.utc).replace(tzinfo=None) (python-docx expects naive UTC). For edits to an existing file, preserve provenance unless the user requested a metadata reset.
5. Run scripts/docx_quality_gate.py. It exits 2 on any issue, including a requested --check-libreoffice that could not run.
6. Render to images and look at every page before shipping: soffice --headless --convert-to pdf out.docx && pdftoppm -png -r 80 out.pdf page, then open the page-*.png files. For external distribution, also check in Word and apply the accessibility checklist.

python3 scripts/docx_render_template.py template.docx context.json out.docx   # exit 2 on missing vars
python3 scripts/docx_quality_gate.py out.docx --json                          # exit 2 on any issue
python3 scripts/docx_extract.py in.docx --include headers footers comments --out extracted.json
node scripts/docx_to_html.mjs in.docx out.html --style-map map.txt --extract-images-dir assets/

Round-Trip Release Proof

• Capture the expected counts of sections, tables, images, hyperlinks, comments and tracked changes before the final save.
• Reopen the saved DOCX and compare the counts.
• Render a copy and inspect page breaks, numbering, table overflow, missing fonts and unresolved tags.

A successful OOXML parse or text extraction does not prove layout fidelity. A visually correct PDF does not prove that comments, hyperlinks, fields or accessibility metadata survived in the editable DOCX. Record which viewer did the primary review and which secondary viewer was checked.

Navigation

Resources

• references/docx-patterns.md - python-docx gotchas: styles, tables, sections, fields, hyperlinks, corruption pitfalls
• references/template-workflows.md - docxtpl tag placement, fail-closed rendering, batch rules
• references/review-comments-workflows.md - Comments, review notes, programmatic redline routing, comment limits
• references/tracked-changes.md - What is and is not feasible for tracked revisions
• references/llm-extraction-workflows.md - Mammoth, MarkItDown, Docling, HTML/Markdown/JSON extraction; includes Extraction Stack tool candidates by job (native parse, structured converter, OCR)
• references/docx-extraction-failure-modes.md - Silent-loss modes for KB/RAG ingestion (content-control tables, image tables, revisions, lying filenames) and the table-count detection check
• references/accessibility-compliance.md - Word accessibility, EN 301 549 context, manual checks
• references/cross-platform-compatibility.md - Word, Google Docs, LibreOffice, PDF conversion
• references/document-automation-pipelines.md - CI/CD, batch generation, quality gates
• data/sources.json - External documentation links

Scripts

• scripts/docx_inspect_ooxml.py - Dependency-free OOXML inspection for tracked changes and comments
• scripts/docx_extract.py - Extract text, tables, metadata, and optional headers/footers/hyperlinks/comments/images to JSON
• scripts/docx_render_template.py - Render a docxtpl template from JSON, failing on missing variables (--allow-missing for drafts)
• scripts/docx_to_html.mjs - Convert trusted .docx to HTML with style maps and optional image extraction
• scripts/docx_quality_gate.py - Validate parseability, unresolved template tags, tracked-change/comment signals, and optional LibreOffice conversion; exits 2 on any issue

Templates

• assets/docx-template-authoring-checklist.md - Template authoring and handoff checklist

Related Skills

• ../document-pdf/SKILL.md - PDF generation and release workflows
• ../document-xlsx/SKILL.md - Spreadsheet generation and exports
• ../document-pptx/SKILL.md - Presentation generation
• ../docs-codebase/SKILL.md - Technical writing patterns

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.