Under the hood, .docx is a ZIP container holding XML parts. Creation, reading, and modification all operate on this XML structure.
Visual and typographic standards: Consult skills/design-foundations/SKILL.md for color palette, typeface selection, and layout principles (single accent color with neutral tones, no decorative graphics, WCAG-compliant contrast). Use widely available sans-serif typefaces like Arial or Calibri as your baseline.
Choosing an approach
Objective
Technique
Reference
Create a document from scratch
docx npm module (JavaScript)
See CREATION.md
Edit an existing file
Unpack to XML, modify, repack
See EDITING.md
Pull out text
anydoc document.docx -o output.md
For redline content use pandoc document.docx -o output.md --track-changes=all
System tools such as anydoc, pandoc, soffice, and pdftoppm are installed in the sandbox. Python packages such as lxml, python-docx, and pdf2docx are pre-installed — run scripts with python3. Use uv run --with <package> only for a package that is not pre-installed.
Editing professional documents
Finance, legal, and other corporate documents carry conventions that are easy to break without noticing. Honor these whenever you create or edit one — violating them is what makes output read as obviously machine-generated:
Format with Word styles, never literal markdown. Map # to a Heading style, **bold** to a bold run, *italic* to an italic run, and > to a block-quote style. Never leave raw #, **, >, or - characters in the document body — literal markdown is the clearest tell that a file was generated. Apply named paragraph styles (Heading 1, Normal) rather than hand-setting fonts, so the document's theme and any brand template stay intact.
Treat tracked changes as a record, not scratch space. If a document already has tracked changes, keep your own edits tracked, and never accept or reject someone else's redlines — each is attributed to a named reviewer. For contract edits, default to a tracked-changes redline and offer a clean accepted copy only if asked.
Keep fields and links live. Cross-references, tables of contents, page numbers, dates, and linked Excel/OLE objects (REF, PAGEREF, TOC, DATE, DOCPROPERTY, NUMPAGES, LINK) must stay as fields. Re-typing "Section 2.1" or a linked figure as flat text severs every downstream reference. Don't renumber legal clauses (1.1, (a), (i)) without instruction — referenced numbering will desync.
Match the document's number dialect. Sample existing figures before adding new ones: currency style (\$1.2bn vs \$1.2B vs \$1.2MM), comma separators, decimal precision per column, and negatives in parentheses for finance ((\$1,234), not -\$1,234). In legal text, follow the spelled-out-then-numeral convention already in use (Thirty (30) days).
Defined terms are load-bearing. In contracts, "the Agreement" and "the agreement" mean different things. Scan the definitions section and inline ("Defined Term") / "Defined Term" means … patterns, then preserve their exact capitalization, pluralization, quote style (curly vs straight), and emphasis (bold/underline) in any new text you generate.
Preserve markings and placeholders verbatim. Keep confidentiality and privilege banners (PRIVILEGED & CONFIDENTIAL, ATTORNEY WORK PRODUCT) in the page header, and never add new ones unless asked. Leave deliberate blanks and notes ([•], [TBD], [NTD: …], [XX]) untouched — on a "finalize" request, flag what's still missing rather than guessing a value.
Don't disturb template scaffolding. Content controls and custom XML parts databind firm-managed templates. Insert text into a control's range only; never delete the control, rename its tag, or strip XML parts, and don't "tidy" branded tables or theme fonts.
Check metadata before sharing. Before any export or hand-off, scan for leftover tracked changes, comments, hidden text, and author metadata, and surface them — never claim metadata was scrubbed without actually removing it.
PDF to Word
Start by running pdf2docx to get a baseline .docx, then correct any artifacts. Never skip the automated conversion and attempt to rebuild manually.
pdf2docx is pre-installed — run the conversion code with python3.
python
from pdf2docx import Converterparser = Converter("source.pdf")parser.convert("converted.docx")parser.close()
Once you have the converted file, address any problems (misaligned tables, broken hyperlinks, shifted images) by unpacking and editing the XML directly (see EDITING.md).
Image rendering
bash
soffice --headless --convert-to pdf document.docxpdftoppm -jpeg -r 150 document.pdf pagels page-*.jpg # always ls to discover actual filenames — zero-padding varies by page count
After generating the document, run a verification pass yourself: inspect extracted text, render preview images when useful, and fix issues before delivering the file.