Office Word Diff

office-word-diff

When to Use

• You need to apply AI-generated edits as tracked changes in Word documents
• You're building a legal tool that produces DOCX output with redlines
• You need deterministic, word-level text comparison (not just line-level diff)
• You want to preserve formatting through document transformations
• You're integrating with Claude Code, Cursor, or other agents that produce document changes

How It Works

Core Problem

Most diff tools operate at the line or paragraph level. Word documents have structural complexity — nested lists, tables, headers, formatting runs — that line-based diffs destroy. office-word-diff solves the kernel-level problem: applying word-level tracked changes to DOCX while preserving formatting.

Architecture

1. Parse — Read DOCX via Office.js API, extract text runs with formatting metadata
2. Diff — Compute word-level differences between old and new text
3. Transform — Insert Word-compatible tracked changes (insertions/deletions) preserving formatting
4. Output — Return valid DOCX with native Word tracked changes

Usage

import { diffWords } from 'office-word-diff';

const result = await diffWords(oldDoc, newDoc, {
  trackChanges: true,
  author: 'LQClaw'
});

Key Features

• Granular tracked changes — Word-level, not paragraph-level
• Format preservation — Bold, italic, underlines survive transformation
• Nested structures — Handles tables, lists, headers correctly
• Agent-compatible — Works with Claude Code, Gemini CLI, Codex as backend
• Deterministic — Same input produces same output (unlike fragile text-matching approaches)

Audience and Work Shape

Audience: lawyers and developers building Word add-ins or document-transformation tools that emit native tracked changes. Output is consumed by Microsoft Word and ultimately signed off by a lawyer.

Work shape: Bounded Transactional. The skill applies a defined set of word-level edits to a DOCX and emits revised DOCX. It does not decide which edits to apply.

Scope and Legal Use

This skill provides legal support, not legal advice. The output is a DOCX with tracked changes — not a redlined version that is approved, signed off, or fit to send to counterparties or file with a court.

Privilege and confidentiality. Runs client-side via Office.js. No network calls. The skill does not transmit document content anywhere. Privilege exposure depends entirely on what the calling application does with the input and output — the lawyer must control the data flow.

Accountability. A qualified lawyer must review every tracked change in the output before accepting, sending, filing, or storing the revised document. The author field (author: 'LQClaw' in the usage example) makes the AI origin visible in Word's revision metadata — keep that visible rather than masking it as a human author.

Confidence Bands

The diff itself is deterministic, but the upstream generation of newDoc (from an AI agent) is not. Tag tracked changes with confidence from the upstream agent where possible:

• High — small, scoped edit (typo, defined-term substitution, cross-reference fix).
• Medium — clause rewrites, paragraph reordering, or any substantive change. Always lawyer-reviewed.
• Low — multi-clause or whole-section rewrites. Treat as a draft proposal, not a redline.

Out of Scope

• Not for executed contracts without separate amendment-protocol controls.
• Not for documents under litigation hold or e-discovery where original metadata must be preserved.
• Not a final-send step — never auto-accept changes and transmit.
• Not for documents with DRM, IRM, or rights-managed protection (will fail or strip protection).
• Not for documents with pre-existing unaccepted tracked changes unless explicitly handled by the calling tool (state must be reconciled first).

Escalation

Stop and route to the responsible lawyer when:

• a nested-table edit cannot be applied without structural loss;
• the input DOCX has pre-existing unaccepted tracked changes from another author;
• the upstream agent produced a Low-confidence edit set or contradictory edits;
• the document is rights-managed, encrypted, or under a litigation hold.

Limitations

• Requires Office.js (needs to run in Word context or headless Word)
• Windows/macOS only (no Linux for Word)
• Very complex nested tables may have edge cases

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: LegalQuants
• Source: LegalQuants/lq-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.