Testing Claude's ability to edit Word documents

Claude-built agent: 50/50 Docx skill: 43/50 Generated Mar 27, 2026
The Experiment

What we tested

50 document editing tasks on real Word files — contracts, memos, invoices, policy handbooks — ranging from simple find-and-replace to multi-field updates across 150-paragraph documents.

AgentA Claude-built agentic approach with structured tools: read, search, replace, diff, verify. Iterates across multiple turns and self-checks its work.
Docx skillClaude Code's built-in docx skill. Claude unpacks the .docx to raw XML, edits it directly, then repacks — no document preprocessing.

The result

50/50 Claude-built agentic approach
43/50 Claude Code docx skill
8 smoke tests — Basic operations: find-and-replace, table edits, multi-field updates across memos, NDAs, invoices, and leases
13 feature tests — OOXML features: comments, footnotes, headers/footers, tracked changes, form fields, and scale tests up to 150 paragraphs
29 capability tests — Benchmark matrix across 6 categories on formatting, structure, reasoning, and real-world legal document tasks
Where the docx skill fell short

7 of 50 tests failed. The failures cluster around a few distinct limitations of the raw XML editing approach.

Form field placeholders
Word content controls — the XML elements behind form field placeholders — sit above the regular paragraph layer that the docx skill edits. It never located or filled them.
Asked toFill in a client intake form by replacing placeholder text in gray fields with real values
What happenedThe placeholder text remained unchanged — the skill edited runs in the document body but couldn't see the content control fields
Long documents
Two tests involved documents over 100 paragraphs. The skill missed occurrences of the target text and, on the longest document, hit its turn limit before completing all edits.
Asked toRename the licensee party throughout a ~100-paragraph enterprise agreement, and separately rename a company throughout a 150+ paragraph policy handbook
What happenedSome occurrences of the old name were left unchanged. The handbook test also hit the turn limit before all edits were complete
Cross-document reference
One test asked Claude to update a contract to match values in a separate term sheet. It made most of the updates but missed a field — a task that requires holding both documents in context simultaneously.
Asked toConform an asset purchase agreement to a term sheet: update purchase price, closing date, escrow amount, representations, and warranty period
What happenedMost fields were updated correctly, but the 18-month warranty period from the term sheet was not reflected in the output
Precise length reduction
Asked to shorten a memo by a specific percentage, the skill reduced the length but not enough. It has no feedback loop to measure word count and iterate.
Asked toShorten a memorandum by 30–40%, removing redundant analysis and tightening the prose while preserving all key conclusions
What happenedThe document was shortened by only 19% (727 → 586 words) — below the minimum 20% threshold
Nuanced legal posture changes
Switching a contract clause from one legal posture to another requires understanding which provisions are affected and how they interact. The edit was applied but didn't fully reflect the intended shift.
Asked toSwitch an asset purchase agreement from buyer-friendly to seller-friendly by adding "to Seller's knowledge" qualifiers to five specific representations
What happenedThe clause edits were made but the knowledge qualifier language was not present in the output as tracked changes
Proofreading
The skill caught most errors in a proofreading task but missed at least one typo — suggesting it's unreliable as an exhaustive error checker.
Asked toProofread a memo for spelling errors, grammatical errors, and inconsistent capitalization of defined terms
What happenedMost errors were corrected, but the misspelling "Reccomendation" was left in the document
Smoke Tests (8)
Simple find-and-replace in memo
Office memo — remote work policy +
Change party names in NDA
Non-disclosure agreement +
Edit table cells in invoice
Consulting invoice +
Modify terms in employment agreement
Employment agreement +
Lease negotiation redline
Commercial lease +
Board resolution - update deal terms
Board resolution +
SOW - change project director and scope
Statement of work +
Shareholder letter - update year and CEO
Annual shareholder letter +
Feature Tests (13)
F1: Add review comments to problematic contract clauses
Consulting agreement +
F2: Delete resolved comments, keep actionable ones
Draft with review comments +
F3: Update legal citations in footnotes and endnotes
Legal brief with footnotes +
F4: Rebrand document — headers, footers, and body
Corporate report +
F5: Fill form fields (content control placeholders)
Client intake form +
F6: Accept all tracked changes (clean up redlines)
Contract with tracked changes +
F7: Edit financial table with merged cells
Financial summary +
F8: Multi-feature edit (header + footnote + comment + table)
Master services agreement +
F9: Scattered edits in long enterprise agreement (~100 paragraphs)
Enterprise software license (~100 paragraphs) +
F10: Update API spec — version, endpoints, and data model
API technical specification +
F11: Update quarterly report data across multiple tables
Quarterly financial report (multiple tables) +
[STRESS] F12: Targeted edits in 150+ paragraph policy handbook
Employee policy handbook (150+ paragraphs) +
F13: Add new article to enterprise agreement
Enterprise software license (~100 paragraphs) +
Capability Tests (29)
A: Core Editing
A1: Simple single find-and-replace
Office memo — remote work policy +
A2: Multi-occurrence replace (same term everywhere)
Service contract (15 parallel clauses) +
A3: Targeted edit in a specific section (not global)
Asset purchase agreement (short form) +
A4: Edit minimal single-paragraph document
Single-paragraph document +
B: Formatting Preservation
B1: Edit text within a heavily formatted paragraph (preserve bold/italic)
Formatted consulting agreement +
B2: Change company name but preserve bold formatting on defined terms
Formatted consulting agreement +
B3: Edit paragraph with bold+italic combined formatting
Formatted consulting agreement +
C: Structural Edits
C1: Edit specific cells in a wide 7-column table
Regional sales report (7-column table) +
C2: Add a new row to a wide table
Regional sales report (7-column table) +
C3: Edit items in a nested numbered list
Technical requirements document +
C4: Add new items to a bullet list
Technical requirements document +
C5: Delete a section and its paragraphs
Asset purchase agreement (short form) +
D: Reasoning
D1: Disambiguate similar clauses (edit only 'training' clauses)
Service contract (15 parallel clauses) +
D2: Conditional edit — update escrow based on new purchase price
Asset purchase agreement (short form) +
D3: Rewrite for audience (legal → plain English)
Asset purchase agreement (short form) +
D4: Proofread document with special characters
Contract with accented characters +
E: Edge Cases
E1: Edit around empty paragraphs (whitespace doc)
Document with blank paragraphs +
E2: Many sequential edits causing index shifts
Asset purchase agreement (short form) +
E3: Preserve summary paragraph while updating table data
Regional sales report (7-column table) +
E4: Convert paragraphs to a table
Single-paragraph document +
E5: Bulk edit 15 similar clauses with a single instruction
Service contract (15 parallel clauses) +
H: Advanced Scenarios
S1:Conform APA to term sheet
Asset purchase agreement (full) +
S2:Define and propagate new term
Asset purchase agreement (full) +
S3:Convert agreement to template
Professional services agreement +
S4:Revise memo with meeting notes
Legal strategy memorandum +
S5:Shorten memo preserving substance
Legal strategy memorandum +
S6:Repurpose APA for California
Asset purchase agreement (full) +
S7:Switch APA to seller-friendly posture
Asset purchase agreement (full) +
S8:Proofread memo for errors
Legal strategy memorandum +