evolsb/claude-legal-skill: CUAD-based contract review inside your coding agent

The gap claude-legal-skill fills between enterprise legal AI and a blank chat window

The README states the problem directly: the author was reviewing NDAs, SaaS agreements, M&A documents and merchant agreements and wanted assistance inside a coding workflow. Commercial legal AI products are described there as enterprise-only with custom quotes and no API access for individual developers. Open source alternatives such as LexNLP, OpenContracts and LawGlance are described as incomplete and requiring integration work, and none of them are built for AI coding assistants. Generic contract checklists get dismissed for a specific reason: they do not differentiate an NDA from an M&A agreement, they give the same advice to buyers and sellers, and they say "negotiate this" without saying what to ask for.

That last complaint is the design brief. The skill is for someone who already has Claude Code, Codex, Cursor, GitHub Copilot or Gemini CLI open, has a contract in front of them, and wants a structured read rather than a chat reply. It is not a document management system, it does not store contracts, and it has no review queue. The unit of work is one contract and one party position.

How the skill works: skill.md, CUAD categories and position-aware checklists

The repository is small. The top level holds CHANGELOG.md, LICENSE, README.md, skill.md and an examples directory. There is no build step and no runtime service. The skill is the skill.md file plus the examples, loaded by whatever Agent Skills-compatible tool you point at the directory.

When you invoke it, you supply two things in the prompt: the contract and your side. The README shows the pattern in its usage examples, for instance "Review this NDA for red flags - I'm the receiving party" or "Analyze the indemnification in this MSA - I'm the vendor". Position changes what gets flagged, because the same clause is not equally risky for a vendor and a customer.

From there the output is structured rather than conversational. According to the README it includes a risk assessment with severity ratings, a red flags quick scan, a key terms table with section references, market standard benchmarks, negotiability ratings, specific redline language, missing provisions with suggested wording, and internal consistency checks for broken cross-references and undefined terms. The risk categories come from the CUAD dataset, which the README describes as 41 legal risk categories drawn from 510 real contracts. The benchmark table is the part that turns a flag into a number: a liability cap of 12 months is listed as standard, 6 to 11 months as yellow, under 6 months as red; auto-renewal notice of 90 or more days is standard, 60 to 89 yellow, under 60 red; non-competes of 1 to 2 years are standard, 3 to 4 years yellow, 5 or more red; M&A rep survival of 12 to 18 months is standard, 24 to 30 yellow, 36 or more red.

Negotiability ratings are a separate axis from severity, and the README is explicit that they are not the same thing. Mutual termination, cure periods and data export are rated high; liability cap increases and price caps medium; network rules and regulatory requirements low. That distinction is the most useful idea in the project, because a red flag you cannot move is a different problem from a red flag you can.

Installing claude-legal-skill and running a first NDA review

Installation is a clone into your tool's skills directory. The README gives the exact paths for Claude Code and Codex, and says to clone into your tool's skills directory for anything else that follows the Agent Skills standard.

git clone https://github.com/evolsb/claude-legal-skill ~/.claude/skills/contract-review

For Codex the destination changes to ~/.codex/skills/contract-review. If you want to keep a working copy and link it, the README documents this development layout:

git clone https://github.com/evolsb/claude-legal-skill ~/Developer/claude-legal-skill
ln -s ~/Developer/claude-legal-skill ~/.claude/skills/contract-review

After the clone, start your tool and invoke the skill by describing the contract and your position. The README's first example is a single line:

Review this NDA - I'm the receiving party

What you should see is the structured output described above: a red flags scan, a key terms table with section references, benchmark comparisons against the thresholds in the table, and proposed replacement language. The examples directory contains worked outputs for an NDA, a SaaS agreement, an M&A agreement and a balanced agreement, so you can compare your result against those before trusting it.

If you need a tracked-changes Word file or a redline PDF, that is a second install. The README points to legal-redline-tools and shows the command that consumes the JSON the skill produces:

pip install git+https://github.com/evolsb/legal-redline-tools.git
legal-redline apply contract.docx redlined.docx \
    --from-json redlines.json \
    --pdf redline.pdf \
    --memo-pdf internal-memo.pdf

Note the ordering: the skill generates redlines.json first, then legal-redline applies it. The skill alone does not produce the Word document.

Where the accuracy ceiling sits, and what the README does not promise

The README is unusually direct about accuracy. It cites ContractEval benchmarks and states that Claude achieves F1 of roughly 0.62 on clause extraction, then frames the tool as best for first-pass review and issue flagging rather than a replacement for attorney review on material deals. Take that number seriously. An F1 around 0.62 on clause extraction means both missed clauses and misidentified ones are expected, and a missed clause in an indemnity or a survival period is the kind of error that matters.

Three further limits are stated. The analysis is not legal advice and material terms should go to qualified counsel. The analysis defaults to US law, with the README noting that provisions vary by jurisdiction; the jurisdiction awareness feature flags that non-competes are void in CA, ND, OK and MN and raises Delaware versus New York versus California implications, but that is a flag, not a full choice-of-law analysis. Very long contracts may need section-by-section review because of the context window.

There is also a structural mismatch worth naming. The skill is built for a coding agent's context, and the deliverable a lawyer actually sends is a tracked-changes document. That handoff runs through a second repository, so the end-to-end path has two dependencies and a JSON intermediate file. If your workflow is a Word document in and a Word document out, this is not a single-step tool.

The wrong-tool cases follow from those limits. A non-US governing law contract is out of scope by the README's own statement. A deal where the liability cap or the earnout mechanics are the whole negotiation is a matter for counsel, not for a first-pass skill. And a high-volume review queue with hundreds of contracts and a need for audit trails is a document platform problem, which is what the enterprise tools the README lists actually sell.

claude-legal-skill against OpenContracts and LexNLP: checklists versus extraction pipelines

The README names LexNLP, OpenContracts and LawGlance as the open source alternatives it looked at, and describes them as incomplete projects requiring significant integration work that were not designed for AI coding assistants. The difference in approach is architectural rather than a matter of quality.

Those tools are extraction and annotation pipelines. You feed documents in, you get structured data and a review interface out, and you build the workflow around them. claude-legal-skill is the opposite shape: it is a prompt-level skill with a checklist and a benchmark table, and the reasoning happens inside the agent you already have open. There is no server, no database and no interface to maintain, which is why the install is a git clone.

That trade cuts both ways. You get no persistence, no corpus-level search across past reviews, and no audit log. You also inherit whatever model your agent runs, so output quality moves when the model moves. If your need is a searchable archive of every contract the company has signed, an annotation platform is the right category and this skill is not. If your need is a second opinion on one agreement before a call, the clone is faster than standing up a pipeline.

Maintenance, licence and the cost of keeping it current

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. The README credits the CUAD dataset to the Atticus Project. That is worth checking separately, because the MIT licence on this repository covers the skill's own files, and the CUAD dataset it draws its risk categories from carries its own terms. Nothing here is legal advice about your obligations; read both licences if you plan to redistribute.

The last push to the default branch was on 2026-07-23, and the most recent release is v3.0.0 from 2026-02-17, described as position-aware review with market benchmarks. There is no package registry entry to update and no version pin to bump, so upgrade cost is a git pull in the skills directory. The real maintenance cost is not the code. It is the benchmark table. Market standards for liability caps, rep survival periods and non-compete durations shift with deal practice, and the thresholds in the README are only as current as the last commit. If you rely on the yellow and red bands in a negotiation, check the CHANGELOG for when they were last touched rather than assuming they track the current market.

Editorial conclusion

Adopt it if you already work inside an Agent Skills-compatible tool and you want a first-pass read on an NDA, SaaS agreement, merchant agreement or M&A document, with the position of your side stated up front. Do not adopt it as a substitute for counsel on a material deal, and do not expect it to cover non-US law, since the README says analysis defaults to US. Before you rely on an output, verify the F1 ~0.62 clause-extraction figure against your own contract type and check whether your tool reads skill.md from the directory you cloned into. The tracked-changes deliverable is a separate install, legal-redline-tools, not part of this repository.

Frequently asked questions

Official sources

1. evolsb/claude-legal-skill on GitHub
2. License: MIT
3. Project website
4. README
5. Releases