Contract Comparison Skill
Clause-by-clause comparison of two contract versions, flagging every change, who it favors, and hidden risks before you sign.
Pure prompt-only skill: reads two documents via Read/WebFetch, writes one CONTRACT-COMPARISON.md to the working directory; no destructive commands, no credential access, no covert exfiltration paths; narrow scope with transparent data flow. The mandatory legal disclaimer in the output template is a positive signal. Deductions: legal-domain output risks over-trust by users; no user confirmation required for bottom-line recommendations ('which version is more favorable'); no statement on handling of sensitive commercial terms; unknown repository license and unverifiable attribution preclude a high score.
Instructions are well-structured and self-consistent: five steps (read → structural comparison → clause comparison → dangerous patterns → output) with a complete output template; the happy path is plausible on static read. Deductions: no defined behavior for abnormal inputs — unreachable URLs, scanned PDFs (README implies PDF support but this skill has no OCR guidance), or cases where version ordering cannot be determined; no tests, no error-feedback guidance, shifting debugging cost to the user.
Trigger is explicit (/legal compare <file1> <file2>), input types (path/URL/pasted text) and output format are well defined, and the 'two versions vs two different contracts' branching is addressed. Deductions: no declared capability boundaries (non-English contracts, very long documents, jurisdictional limits), no non-fit disclosure; no consideration of Chinese-language contracts; WebFetch dependence on overseas reachability poses potential availability risk for mainland-China networks.
SKILL.md is well-organized and readable; frontmatter name/description match actual capability; the output template doubles as an example. Deductions: mild naming inconsistency (skill name 'contract-comparison' vs directory 'legal-compare' vs command '/legal compare'); no versioning, changelog, known-limitations disclosure, or FAQ; repository license unknown; maintenance ownership and update path unclear; README contains heavy unsourced marketing metrics ('60 seconds', '$10,000+' loss figures) that dilute documentation rigor.
The goal is clear (clause-by-clause comparison, favorability analysis, dangerous-change flagging), and the output structure is directly usable, offering real marginal value over manual line-by-line review. Deductions: no execution evidence under static review; correctness of exact-quote extraction, Major/Minor/Cosmetic classification, and favorability judgment depends heavily on model quality and is unverified; README's '60-second review' and business-value claims are unverifiable.
All skill claims (detecting sneaked-in clauses, favorability determination) are author statements with no test suite, no example input/output pairs, no third-party execution evidence, and no CI covering this skill path; the only auditable material is the SKILL.md source itself. Scored near the floor: no key claim is independently reproducible, and facts cannot be separated from inference.
- This is a static source review; nothing was executed and confidence is low; claims like '60-second review' are unverified.
- Output is informational analysis, not legal advice — always consult a licensed attorney before signing.
- URL inputs depend on WebFetch and may be unreachable from mainland-China networks; scanned PDFs or non-English contracts may fail silently or degrade quality.
- Repository license is unknown — confirm before commercial use; minor naming inconsistency exists among skill name, directory, and command.
- For contracts containing sensitive terms, note that document content will be sent to the model provider.
What does this skill do, and when should you use it?
This is one of 14 skills in the AI Legal Assistant collection (zubair-trabzada/ai-legal-claude), located at skills/legal-compare/. It compares two contracts — two versions of the same agreement or two entirely different contracts — classifying every difference (added, removed, substantive, cosmetic), assessing which party each change favors, and rating its significance and risk. It specifically flags dangerous patterns such as sneaked-in clauses, stripped liability protections, and unilateral amendment rights. The output is a structured CONTRACT-COMPARISON.md report with a change summary table, dangerous-change callouts, and a bottom-line recommendation. It is a pure prompt skill that relies on Claude Code's Read and WebFetch tools to ingest files or URLs, and its output is explicitly informational, not legal advice.
Triggered by /legal compare <file1> <file2>, accepting file paths, URLs, or pasted text. Workflow: 1) reads both documents and determines whether this is a version comparison or a comparison of two different contracts; 2) performs a structural comparison (sections shared, only in A, only in B, renumbering); 3) compares clause by clause, classifying change types with significance ratings (Major/Minor/Cosmetic) and favorability (Party A/B/Neutral); 4) flags nine dangerous patterns (sneaked-in clauses, stripped protections, scope expansion, financial term changes, termination changes, indemnification drift, IP rights shifts, governing law changes, unilateral amendment rights); 5) quotes exact text from both documents and generates CONTRACT-COMPARISON.md containing document overview, executive summary, change summary table, dangerous-change details, per-change analysis, and recommendations.
- A freelancer comparing the revised client contract against the original before signing, to see what quietly changed
- A small business checking a vendor's v2 contract to confirm liability caps or termination rights weren't removed
- A founder or in-house reviewer comparing pre- and post-negotiation drafts to build a list of clauses to push back on
- An agency comparing two contracts from different sources (e.g., two NDAs) on how equivalent sections stack up
- Anyone auditing boilerplate sections of a long contract for buried risk-allocation changes
What are this skill's strengths and limitations?
- Clear methodology: structured three-dimensional output of change type, favorability, and risk rating
- Mandates quoting exact source text rather than paraphrasing, making every difference verifiable
- Covers nine named dangerous-change patterns (sneaked-in clauses, uncapped indemnification, etc.) rather than a generic diff
- Self-contained Markdown report you can send to counterparties or archive
- Pure prompt implementation with no external dependencies — report generation needs no Python or libraries
- Explicitly disclaims legal advice; material contracts still require a licensed attorney
- No license declared for the repository, leaving commercial use and redistribution status unclear
- No test suite or benchmarks, so analysis quality across models/model versions is unverifiable
- Repository description contains marketing content (a Skool community pitch); separate claims from actual capability
- Dangerous-pattern detection relies on model judgment and may miss changes buried in boilerplate
How do you install this skill?
This skill is part of the ai-legal-claude collection, installed with the one-line installer: curl -fsSL https://raw.githubusercontent.com/zubair-trabzada/ai-legal-claude/main/install.sh | bash (installs all 14 skills, 5 agents, and the PDF scripts). The repo documents no standalone install for legal-compare alone; to use it individually, manually place the skills/legal-compare/ directory into your Claude Code skills directory.
How do you use this skill?
In Claude Code, run: /legal compare <file1> <file2>, where each argument is a local file path, a URL, or pasted contract text. The skill reads both documents and writes CONTRACT-COMPARISON.md to the current working directory. If the two inputs are entirely different contracts rather than versions, the output shifts to comparing equivalent sections instead of tracking changes.