# Legal/TOS Diff-er    A semantic diff tool for Terms of Service and legal documents. Unlike standard text diffs that spot character changes, this skill understands legal meaning — catching when "may" becomes "will" in a data-sharing clause or when a forced arbitration clause quietly appears. ## The Problem Companies update their Terms of Service frequently, and the changes are often buried in pages of dense legal text. A standard code diff looks for character changes, but legal changes require **semantic understanding**: - Changing "may share data" to "will share data" is a single word, but a massive privacy shift - Adding "mandatory arbitration" to a dispute section strips users of their right to sue - Changing a refund policy from "within 30 days" to "at our discretion" eliminates a financial right ## How It Works ``` ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ Fetch URL │────▶│ Extract │────▶│ Snapshot │────▶│ Compare │ │ (node-fetch)│ │ (cheerio) │ │ (JSON) │ │ (Claude) │ └─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘ ``` 1. **Fetch** — Retrieves the legal page HTML 2. **Extract** — Two-pass engine strips noise (nav, ads, popups) and scores content blocks to isolate legal text 3. **Snapshot** — Stores timestamped versions with SHA-256 hashes 4. **Compare** — Outputs a structured prompt for Claude to semantically analyze changes ## Change Categories | Category | What It Detects | Example | |----------|----------------|---------| | **Privacy Risks** | Data collection, sharing, tracking, cookies | "may share" → "will share" with third parties | | **Financial Changes** | Pricing, fees, billing, refunds, auto-renewal | "30-day refund" → "at our discretion" | | **User Rights** | Termination, ownership, arbitration, governing law | New mandatory arbitration clause | ## Quick Start ### Commands ``` # Track a new legal document add_url --url "https://example.com/terms" --label "Example Corp TOS" # See what you're tracking list_tracked # Capture the current version fetch_current --url "https://example.com/terms" # Compare current version against last snapshot diff --url "https://example.com/terms" # Stop tracking remove_url --url "https://example.com/terms" ``` ## Installation ```bash cd legal-tos-differ npm install ``` Requirements: Node.js 18+ ## Architecture ### Extraction Engine The extraction engine uses a two-pass approach with Cheerio: 1. **Noise Removal** — Strips `<nav>`, `<footer>`, `<script>`, and elements with noise-related classes/IDs (sidebar, cookie, popup, etc.) 2. **Content Scoring** — Scores remaining block elements by: - Text density (legal text is text-heavy, not link-heavy) - Legal keyword frequency ("terms", "agreement", "liability", etc.) - Link density penalty (too many links = navigation, not legal text) - Structural hints (`<main>`, `<article>`, legal-related IDs/classes) ### Snapshot Storage Snapshots are stored as JSON files in `snapshots/`: ``` snapshots/ registry.json # Tracked URLs metadata example-com-terms-2026-04-11T17-00.json # Timestamped snapshot ``` Each snapshot includes the full extracted text, SHA-256 hash, and fetch metadata. The hash enables instant "no changes" detection without invoking the LLM. ### Analysis Prompting The skill builds a structured prompt that delegates semantic analysis to the Claude Code runtime. The prompt instructs the LLM to: - Ignore cosmetic changes (typos, formatting, reordering) - Ignore clarifying language that doesn't change legal meaning - Flag removals of user protections as higher severity - Quote specific old/new text for each change ## License MIT