Best AI Agent Skills for Legal Research in 2026

Legal research is 90% reading and 10% judgment, and the reading part is where a good agent skill saves you hours. Not the judgment. You still need a lawyer for that, and nothing below changes it. But pulling the relevant clause out of a 60-page contract, cross-checking three versions of a policy, drafting a first-pass memo you'll then tear apart? That's exactly the grunt work Claude Code skills handle now.

A quick definition, because the terms get muddled. An agent skill is a folder with a SKILL.md file that teaches Claude a repeatable procedure: how to research a case, how to review a contract, how to format a memo. Claude loads it only when the task matches, so you can install a dozen and none of them clog your context until they're needed. For legal work, a handful of these turn Claude Code into a competent research assistant that never gets bored on page 40.

Here are the ones worth installing, with the security caveats that matter more in legal than almost anywhere else.

The skills, at a glance

The trust column is not decoration. I'll come back to it, because installing a skill means running its instructions with your file access, and legal documents are exactly the kind of thing you don't want leaking.

1. deep-research: the backbone

Most legal research questions boil down to "find everything relevant, then tell me what contradicts what." That's what deep-research does: it fans out multiple searches, fetches the sources, adversarially checks claims against each other, and hands back a cited report. The adversarial step is the reason to prefer it over asking Claude to "research X" cold, because it actively looks for the counter-argument instead of confirming your first hypothesis.

Install it via the plugin marketplace:

# Add the marketplace, then install
/plugin marketplace add anthropics/skills
/plugin install deep-research

Or drop it in manually if you're working from a cloned repo:

git clone https://github.com/anthropics/skills.git
cp -r skills/deep-research ~/.claude/skills/deep-research

Then in Claude Code: "Research the enforceability of non-compete clauses for remote employees across California, New York, and Texas, with sources." It'll come back with a structured brief and citations you can verify. Verify them; the security section explains why.

2. pdf-tools: because everything is a PDF

Legal runs on PDFs, and half of them are scanned images with no selectable text. pdf-tools extracts text, searches across a folder of documents, splits a 200-page filing into the exhibits you care about, and runs OCR on the scanned ones. This is the unglamorous skill you'll use most.

/plugin install pdf-tools

A real task it handles well: "OCR every PDF in this folder, then find every document that mentions 'indemnification' and give me the page number." That's an afternoon of manual Ctrl+F across un-searchable scans, done in a couple of minutes. On a due-diligence data room with 400 files, that difference is your whole afternoon back.

3. docx-word: first-draft memos and redlines

When the research is done, someone has to write it up in Word with tracked changes, because that's what the partner expects. docx-word reads existing .docx files, produces new ones with proper headings and formatting, and works with tracked changes and comments. Pair it with deep-research and you get "research this question, then draft a memo in our house format" as a single flow.

/plugin install docx-word

Keep your expectations calibrated: it drafts, you edit. The formatting will be clean and the citations will be placed, but the argument is yours to sharpen. It also respects a template if you feed it one, so the output matches your firm's house style instead of some generic memo shape.

4. Contract-focused skills: useful, but audit first

opencouncil-contract-inspector and commercial-legal-pl target the contract side: flagging unusual clauses, applying a review checklist, using commercial-legal templates. They're useful for a first pass over an NDA or a vendor agreement. They're also community-published, which means the instructions inside them were written by someone you don't know, and you're about to run those instructions with access to confidential contracts.

So audit before installing. Open the SKILL.md and read it end to end:

git clone https://github.com/<publisher>/<skill-repo>.git
cat <skill-repo>/SKILL.md   # read every line before you install

You're checking for one thing above all: does it try to send document content anywhere, to an external API, a webhook, a "logging" endpoint? A legit contract skill processes text locally and returns analysis. If it phones home with your document, that's a dealbreaker, not a feature. Our full list of AI tools for lawyers covers the standalone contract-review products if you'd rather not run community skills at all.

Security: the part that's non-negotiable in legal

Everywhere else on this site I say "check the skill before installing." In legal, it's the whole ballgame, because the documents are privileged and the downside of a leak is a bar complaint, not an oops.

Read the SKILL.md before every install. A skill is instructions Claude will follow with your file access. Community skills can be perfectly good, and plenty are, but "I trust the author" is not a security model. Reading 200 lines of Markdown takes five minutes and is the entire defense.

Watch for network calls. The single highest-risk pattern is a skill that transmits document content externally. Grep the skill folder:

grep -rEi "curl|http|fetch|webhook|api\." ~/.claude/skills/<name>/

Hits aren't automatically bad, since a research skill needs to fetch web pages, but every one should have an obvious, legitimate reason. A contract-review skill making outbound calls with your contract text does not.

Keep privileged documents off third-party tools. Even with a clean skill, know where the underlying model processes your data and whether it's retained. For anything privileged, use an enterprise agreement with data-retention turned off, and strip client-identifying details when the task doesn't need them.

Never file what you didn't verify. AI hallucinates citations. This has already gotten real lawyers sanctioned for submitting briefs citing cases that don't exist. Every citation an agent produces gets checked against the actual source before it goes anywhere near a filing. No exceptions, no "it's probably fine."

The honest limit

These skills make you faster at the research, extraction, and drafting that eats your day. They do not practice law, and they get citations wrong often enough that unverified output is a liability, not an asset. Treat the output as a sharp first draft from a tireless junior who occasionally makes things up with total confidence. Useful, and never trusted blind.

And they don't know your jurisdiction unless you tell them, so a memo drafted without you specifying the governing law will confidently cite the wrong state's rule. If you install two things from this list, make them pdf-tools and deep-research. They cover the reading-heavy 80% of legal research, they're from a source you can trust, and they'll save you the most hours per week.

FAQ

Can AI agent skills replace a paralegal or lawyer? No. They handle the mechanical parts of research (extracting text, searching documents, drafting first-pass memos), but legal judgment, strategy, and any advice remain human work. Think of them as tools that make a paralegal faster, not a replacement for one, and never as a substitute for a licensed attorney.

How do I install a Claude Code skill for legal research? Add the skills marketplace with /plugin marketplace add anthropics/skills, then /plugin install <skill-name>. For community skills, clone the repo and read the SKILL.md before copying it into ~/.claude/skills/. Always audit community skills before installing. You're running their instructions with your file access.

Is it safe to use AI on confidential legal documents? Only with precautions. Read every skill's source before installing, check for outbound network calls, use a model with data retention disabled for privileged material, and strip identifying details where possible. The bigger risk is usually the underlying model's data policy, not the skill itself. Check both.

Will AI make up case citations? Yes, and it's already gotten lawyers sanctioned. Language models generate plausible-looking citations that don't exist. Verify every single citation against the actual source before using it in any filing or client communication. This is the one rule with no exceptions.

Which legal research skill should I start with? deep-research and pdf-tools. deep-research handles multi-source research with fact-checking built in; pdf-tools handles the constant job of extracting and searching document text, including scanned files. Both come from a trusted source and cover the bulk of day-to-day legal research work.