Menu
Advanced โ structural OOXML edits via run_code + lxml + zipfile. Use ONLY when the docx-editing tools (edit_document, add_comment, accept/reject, revert_edit) can't express the change. For ordinary text edits, use the docx-editing skill instead.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Compare two document versions and produce analysis or a Word redline. Use when the user has two versions of a document and wants to understand what changed.
Edit existing DOCX files โ tracked-change redlines, placeholder fills, range deletes, comments, and accept/reject finalization. Plain-text-only tools; no raw OOXML. For structural edits beyond text replacement, invoke `Skill('docx-xml')`.
Draft a NEW document from scratch via docx-js (run_code language="node"). Use when the user wants to create new legal content from a blank page โ agreements, memos, letters, resolutions, term sheets. For filling an existing template with placeholder values, use the `docx-editing` skill instead.
Quality assurance review โ 6-dimension checklist covering template compliance, factual accuracy, completeness, track changes, research audit, and instruction cross-reference.
Research a legal topic with web search and inline citations. Use when the user needs legal information, market standards, regulatory guidance, or case law references.
Review a document against playbook positions. Analyze clauses for risks, flag issues, suggest specific changes, and apply edits. Subsumes redlining โ review + fix is one workflow.
| name | docx-xml |
| emoji | ๐ง |
| description | Advanced โ structural OOXML edits via run_code + lxml + zipfile. Use ONLY when the docx-editing tools (edit_document, add_comment, accept/reject, revert_edit) can't express the change. For ordinary text edits, use the docx-editing skill instead. |
| requires | {"tools":["read_document","run_code","list_documents"]} |
You are looking at this skill because a structural edit needs raw OOXML manipulation. Always check first whether edit_document can do it โ see the docx-editing skill. Only drop to run_code for:
<w:p><w:pPr><w:rPr><w:del/></w:rPr></w:pPr> pattern)word/document.xml (styles.xml, headers/footers, numbering.xml)A .docx file is a ZIP archive containing XML files. The main content is in word/document.xml. The OOXML namespace for Word is http://schemas.openxmlformats.org/wordprocessingml/2006/main (prefix w:).
Use language="python" + lxml + zipfile on the XML parts directly.
Why NOT docx-js: docx-js is write-only โ there is no Document.load(...) API. It's for creating new DOCX files from scratch (the draft skill uses it). You cannot use it to load Contract.docx, mutate, and save back.
Why NOT python-docx for run-level mutation: python-docx silently corrupts complex legal templates. It loses non-target glyphs (other placeholders, special characters) when run.text is reassigned, has no reliable API for footnotes / bookmarks / internal hyperlinks / TableOfContents, and collapses <w:rPr> run properties when rewriting text โ losing bold / font / size on surrounding runs. Never use python-docx for run.text.replace(...) or any run-level mutation. python-docx for reading and additive operations (adding images, adding paragraphs) is acceptable.
# run_code(language="python", input_files=["Contract.docx"], code=...)
import zipfile, io
from lxml import etree
NS = "http://schemas.openxmlformats.org/wordprocessingml/2006/main"
W = f"{{{NS}}}"
# 1. Read document.xml from the input DOCX.
with zipfile.ZipFile('/sandbox/input/Contract.docx') as zin:
doc_xml = zin.read('word/document.xml')
parts = {name: zin.read(name) for name in zin.namelist()}
# 2. Mutate with lxml (example โ delete a specific paragraph including its mark).
root = etree.fromstring(doc_xml)
for p in root.iter(f"{W}p"):
text = "".join(t.text or "" for t in p.iter(f"{W}t"))
if "text-to-delete" in text:
p.getparent().remove(p)
# 3. Repack, preserving every other part.
parts['word/document.xml'] = etree.tostring(root, xml_declaration=True,
encoding='UTF-8', standalone=True)
buf = io.BytesIO()
with zipfile.ZipFile(buf, 'w', zipfile.ZIP_DEFLATED) as zout:
for name, data in parts.items():
zout.writestr(name, data)
with open('/sandbox/output/Contract_Edited.docx', 'wb') as f:
f.write(buf.getvalue())
One document = one run_code call. A single call should open the file, make all changes, and save one output. Never split a document edit across multiple scripts.
word/document.xml
<w:body>
<w:p> โ paragraph
<w:pPr> โ paragraph properties (style, numbering, spacing)
<w:pStyle w:val="Heading1"/>
<w:numPr>...</w:numPr>
<w:spacing>...</w:spacing>
<w:ind>...</w:ind>
<w:jc>...</w:jc>
<w:rPr>...</w:rPr> โ always last child of pPr
</w:pPr>
<w:r> โ run (contiguous text with same formatting)
<w:rPr> โ run properties (bold, italic, font, size)
<w:b/> โ bold
<w:i/> โ italic
<w:sz w:val="24"/> โ font size in half-points (24 = 12pt)
<w:rFonts w:ascii="Arial"/>
</w:rPr>
<w:t xml:space="preserve">text content</w:t>
</w:r>
</w:p>
| Rule | What goes wrong if ignored |
|---|---|
<w:t> with leading or trailing whitespace must carry xml:space="preserve" | Word strips the spaces during round-trip |
Inside <w:del>, deleted text content lives in <w:delText>, not <w:t> | Tracked changes corrupt; Accept All leaves stale text |
<w:rPr> is always the last child of <w:pPr> | Schema violation; some viewers reject the document |
When wrapping a run in tracked-change markup, copy the original <w:rPr> into your <w:ins> / <w:del> runs | The change preserves bold / font / size of surrounding text |
When creating a tracked change, replace the entire <w:r> element rather than injecting markup inside it | Avoids invalid run nesting and preserves Word's display logic |
Every w:id in tracked changes is unique | Word silently merges or drops changes with duplicate IDs |
xml:space="preserve" is the most-missed of these rules โ most run-level edits need it | Spaces disappearing is invisible in source diff but obvious to the user |
DOCX documents use smart quotes stored as Unicode. When adding text via lxml, use these:
| Code | Char | Description |
|---|---|---|
โ | ' | Left single quote |
โ | ' | Right single quote / apostrophe |
โ | " | Left double quote |
โ | " | Right double quote |
In lxml: elem.text = 'Hereโs a quote: โHelloโ'
Insertion:
<w:ins w:id="1" w:author="Anylegal.ai" w:date="2026-05-04T00:00:00Z">
<w:r><w:rPr>...</w:rPr><w:t>inserted text</w:t></w:r>
</w:ins>
Deletion:
<w:del w:id="2" w:author="Anylegal.ai" w:date="2026-05-04T00:00:00Z">
<w:r><w:rPr>...</w:rPr><w:delText>deleted text</w:delText></w:r>
</w:del>
Minimal edit โ only mark what changes (governing law change from Delaware to England and Wales):
<w:r><w:t>This Agreement is governed by the laws of </w:t></w:r>
<w:del w:id="1" w:author="Anylegal.ai" w:date="...">
<w:r><w:delText>the State of Delaware</w:delText></w:r>
</w:del>
<w:ins w:id="2" w:author="Anylegal.ai" w:date="...">
<w:r><w:t>England and Wales</w:t></w:r>
</w:ins>
<w:r><w:t>.</w:t></w:r>
Deleting an entire paragraph โ mark the paragraph mark as deleted too, otherwise "Accept All" leaves an empty line:
<w:p>
<w:pPr>
<w:rPr>
<w:del w:id="1" w:author="Anylegal.ai" w:date="..."/>
</w:rPr>
</w:pPr>
<w:del w:id="2" w:author="Anylegal.ai" w:date="...">
<w:r><w:delText>The deleted clause text...</w:delText></w:r>
</w:del>
</w:p>
Rejecting a counterparty's insertion โ nest your deletion inside their insertion (preserves their authorship in the audit trail):
<w:ins w:author="Counterparty" w:id="5">
<w:del w:author="Anylegal.ai" w:id="10">
<w:r><w:delText>their inserted text</w:delText></w:r>
</w:del>
</w:ins>
Restoring a counterparty's deletion โ add an insertion after their deletion (do not modify their deletion):
<w:del w:author="Counterparty" w:id="5">
<w:r><w:delText>deleted text</w:delText></w:r>
</w:del>
<w:ins w:author="Anylegal.ai" w:id="10">
<w:r><w:t>deleted text</w:t></w:r>
</w:ins>
Inserting images is an additive operation; python-docx handles relationships and content types correctly:
from docx import Document
from docx.shared import Inches, Cm
doc = Document('/sandbox/input/contract.docx')
doc.add_picture('/sandbox/input/logo.png', width=Inches(2.0))
# Or in a specific paragraph:
para = doc.paragraphs[0]
run = para.add_run()
run.add_picture('/sandbox/input/signature.png', width=Cm(5), height=Cm(2))
doc.save('/sandbox/output/contract_with_images.docx')
This is the only place python-docx is the right tool. For text mutation, always use lxml on the raw XML.
If you need raw XML control (custom positioning, wrapping):
<w:drawing>
<wp:inline distT="0" distB="0" distL="0" distR="0">
<wp:extent cx="914400" cy="914400"/> <!-- EMUs: 914400 = 1 inch -->
<wp:docPr id="1" name="Picture 1"/>
<a:graphic xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main">
<a:graphicData uri="http://schemas.openxmlformats.org/drawingml/2006/picture">
<pic:pic xmlns:pic="http://schemas.openxmlformats.org/drawingml/2006/picture">
<pic:blipFill>
<a:blip r:embed="rId5"/>
</pic:blipFill>
<pic:spPr>
<a:xfrm><a:ext cx="914400" cy="914400"/></a:xfrm>
</pic:spPr>
</pic:pic>
</a:graphicData>
</a:graphic>
</wp:inline>
</w:drawing>
To add via raw XML you must also (1) copy the image to word/media/, (2) add a relationship in word/_rels/document.xml.rels, and (3) register the content type in [Content_Types].xml.
findall(), iterate over a copy: for elem in list(body.findall(...)).elem.get('{http://schemas.openxmlformats.org/wordprocessingml/2006/main}id'), not elem.get('w:id').w:id in tracked changes must be unique. Use max(existing_ids) + 1.When run_code produces a DOCX file, validate it before reporting success. The most common validation failures and their fixes:
| Failure | Fix |
|---|---|
<w:t> inside <w:del> | Use <w:delText> for deleted text, not <w:t> |
Missing xml:space="preserve" | Add the attribute to <w:t> elements that have leading/trailing whitespace |
| Corrupt ZIP | Ensure zipfile.ZipFile(..., 'w') is properly closed before reading the bytes back |
Schema violation in <w:pPr> | Verify element order: <w:pStyle>, <w:numPr>, <w:spacing>, <w:ind>, <w:jc>, <w:rPr> (rPr last) |
If the script produces a corrupt DOCX, fix the Python code and retry rather than retrying the same code.
import openpyxl
from openpyxl.styles import Font, PatternFill
wb = openpyxl.Workbook()
ws = wb.active
ws.title = "Fee Schedule"
headers = ['Service', 'Rate (USD/hr)', 'Estimated Hours', 'Total']
for col, header in enumerate(headers, 1):
cell = ws.cell(row=1, column=col, value=header)
cell.font = Font(bold=True, color="FFFFFF")
cell.fill = PatternFill(start_color="4472C4", fill_type="solid")
services = [
('Contract Review', 350, 8),
('Negotiation Support', 400, 12),
('Due Diligence', 375, 20),
]
for row, (service, rate, hours) in enumerate(services, 2):
ws.cell(row=row, column=1, value=service)
ws.cell(row=row, column=2, value=rate)
ws.cell(row=row, column=3, value=hours)
ws.cell(row=row, column=4, value=rate * hours)
ws.column_dimensions['A'].width = 25
wb.save('/sandbox/output/fee_schedule.xlsx')
from datetime import datetime, timedelta
from dateutil.relativedelta import relativedelta
def add_business_days(start, days):
current = start
added = 0
while added < days:
current += timedelta(days=1)
if current.weekday() < 5:
added += 1
return current
signing_date = datetime(2026, 5, 4)
deadlines = {
"Signing Date": signing_date.strftime("%Y-%m-%d"),
"Effective Date (T+5 business days)": add_business_days(signing_date, 5).strftime("%Y-%m-%d"),
"First Payment Due (30 days)": (signing_date + timedelta(days=30)).strftime("%Y-%m-%d"),
"Warranty Period Ends (12 months)": (signing_date + relativedelta(months=12)).strftime("%Y-%m-%d"),
}
for name, date in deadlines.items():
print(f"{name}: {date}")