Claude Prompts for Legal Research (2026)

There is one rule that makes Claude genuinely useful for legal work and one that gets people sanctioned, and they're the same rule: Claude reads the law you give it — it does not find law for you. Paste an opinion, a rule, a transcript, or a discovery index and it will pull holdings, build timelines, and catch contradictions faster than you can. Ask it to remember a case and it may invent one. Below are 14 copy-ready prompts built around that distinction, from a working investigator who uses them on capital cases.

Start here — set the ground rules and load the record (prompts 1–2)

Prompt 1 is the single most important thing on this page. It costs you fifteen seconds and it is the difference between a research assistant and a liability. Run it at the top of every legal conversation.

Ground rules for this entire conversation. Do not cite any case, statute, rule, or regulation unless I have pasted its text into this chat. If you believe relevant authority exists that I haven't given you, say “you should look for authority on X” and describe it — do not produce a citation, a case name, or a quotation. If I ask you a question you can only answer from memory, tell me that instead of answering. Confirm you understand before we start.

I'm going to give you the record in a case. Here are the documents: [PASTE OR LIST]. Build me an index as a table: document name, date, author or source, and one line on the role it plays in the case. Then list every document that is referenced inside these documents but that I haven't given you — exhibits, attachments, prior transcripts, reports. I want to know what's missing before I start reading.

Reading opinions and case law (prompts 3–5)

This is where Claude earns its keep. You've pulled the case yourself, so there's nothing to hallucinate — the text is right there. Now make it do the tedious part.

Here is the full text of an opinion: [PASTE]. Break it down for me: procedural posture, standard of review, the precise holding, and the reasoning in three sentences. Then separate what is holding from what is dicta, and tell me explicitly what this opinion does not decide. Quote the language you're relying on for each answer.

Here are the facts of my case: [PASTE]. Here is a case I want to rely on: [PASTE OPINION]. First, give me the strongest version of my argument that this case controls. Then switch sides completely: you are opposing counsel — write the sharpest distinction you can draw between these facts and that case. Tell me honestly which side of that exchange is stronger.

This is a 60-page opinion: [PASTE]. Find every passage that addresses [ISSUE]. Quote each one verbatim with its page or paragraph number, and give me one line of context for each. If the opinion never actually addresses that issue, say so plainly rather than giving me the closest thing you can find.

Statutes, rules, and deadlines (prompts 6–7)

Rules change and briefs get recycled. I have personally caught a filing that argued a ten-day deadline under a rule that had been amended to twenty days five years earlier — the entire fallback argument in it was backwards. Prompt 6 is that check.

Here is the current text of [RULE / STATUTE] as I pulled it today from the official source: [PASTE]. Here is how a brief describes and relies on it: [PASTE THE RELEVANT SECTIONS]. List every substantive difference between what the rule says and what the brief says it says. Then tell me which specific arguments in that brief depend on the inaccurate version, and whether each one survives under the correct text.

Here is the text of the rule governing this deadline: [PASTE], and here are the relevant dates: [LIST]. Build me a table of every deadline this triggers: the date, the subdivision it comes from, and whether it runs from filing or from service. Flag anything where filing and service would produce different dates, and anything that falls on a weekend or holiday. Show your arithmetic.

Working a record or a discovery production (prompts 8–10)

A large production is the strongest use case there is. Claude will hold a full deposition transcript and several reports at once and cross-reference them — work that would take a person days.

Here is a transcript and a set of reports: [PASTE]. Build a chronological timeline of every event with a date or time attached. Columns: date, time, event, and the document and page it came from. Where two documents give different times for the same event, put both in and mark the row as a conflict. Don't smooth over disagreements.

Here are three statements by the same witness taken on different dates: [PASTE A], [PASTE B], [PASTE C]. Build a table with one row per topic and one column per statement, quoting what they said each time. Add a final column marking whether the accounts are consistent, a difference in emphasis, or a genuine contradiction. Quote exactly — do not paraphrase, because I'm going to use these on cross.

Here is the index of everything produced to me: [PASTE]. Based on what these documents reference internally — exhibits, lab submissions, property receipts, prior statements, body-camera or call recordings — list every item that should exist but does not appear in the production. Rank them by how much they'd matter, and draft the specific request language for the top five.

Opposing experts and witnesses (prompts 11–12)

When the other side lists an expert you don't know, you have two jobs: understand what they'll say, and find out how they've held up before. Claude handles the first well and helps you organize the second — but the actual search for prior testimony happens in a real database.

Here is the opposing expert's CV and report: [PASTE]. Give me two lists. First, every conclusion that rests on a test result or measurement. Second, every conclusion that rests on an assumption, a judgment call, or another person's work — and for each one, name the assumption. Then give me the five questions I should ask to find out whether each assumption holds.

Here is the transcript of this witness's prior testimony: [PASTE], and here is what I need to establish: [LIST]. Draft a cross-examination outline of short, closed, leading questions, grouped by objective. For each question, cite the page and line of the transcript that locks the answer in. Flag any question where the witness has a good answer available, and tell me what it is.

Drafting and pressure-testing (prompts 13–14)

Do not ask Claude to write your brief. Ask it to attack your brief. That is where the value is, and it's the one use with no citation risk at all, because everything it's working from is yours.

Here is my argument: [PASTE]. You are opposing counsel and you are good at this. Write the strongest response brief you can — not a strawman. Then step back and tell me which of your own points I have no real answer to, and what I would need to add to my argument to close each gap.

Here is my motion: [PASTE]. You are the judge who has to rule against me. Write the two paragraphs of the order denying it, in the voice a trial judge would actually use. Then tell me which sentence of my motion made that order easiest to write.

How to get better results

Legal work punishes the vague prompt harder than most fields, because a plausible wrong answer looks exactly like a right one. Six habits do most of the work:

• Paste the authority, never request it. The whole risk profile changes when the text is in front of it. This is the difference between a research tool and a sanctions order.
• Demand the quote and the pin cite. “Quote the language you're relying on, with the page number” makes every claim checkable in seconds — and makes fabrication obvious immediately.
• Give it permission to find nothing. “If the opinion doesn't address this, say so” removes the pressure to produce an answer, which is what generates the closest-thing-available response.
• Ask for the other side first. The most valuable output isn't your argument written prettier; it's the distinction opposing counsel is going to draw.
• Render scanned pages; don't trust the text layer. Much of what you get in discovery is an image-only scan, and character recognition on a bad photocopy silently changes numbers. If a finding turns on whether something is absent from a document, recognition software can't prove that. Look at the page.
• Keep the verification step human. Claude drafts the analysis. A citator confirms the cite. That division of labor never moves, on any case, for any deadline.

Mistakes to avoid

• Filing a citation you have not personally pulled up. The one that ends careers. Reading the case is not optional because the summary was convincing.
• Pasting privileged material without checking first. ABA Formal Opinion 512, issued July 29, 2024, addresses confidentiality under Model Rule 1.6 among other duties. Know your jurisdiction's guidance, your firm's policy, and the tool's data-retention terms before client information goes anywhere.
• Asking a broad question instead of loading a document. “What's the law on X in Florida?” is the prompt that produces fiction. “Here is the statute — what does it require?” is the prompt that produces work.
• Accepting a paraphrase where you need a quote. On cross, a paraphrase is worthless. Ask for verbatim text every time, then confirm it against the transcript.
• Assuming the rule hasn't changed. Pull the current version from the official source. A brief that quotes a superseded rule is worse than one that cites nothing.
• Letting it write the argument. The output reads well and says little. Use it to find holes in an argument you wrote.

If you work in this area specifically, we have deeper guides for lawyers, paralegals, private investigators, public defenders, and criminal defense. New to Claude entirely? Start with the 10-minute install guide, then browse the full Learn hub.

Frequently asked questions

Can Claude do legal research?

Claude is excellent at reading and structuring law you give it, and unreliable at finding law you don't. Paste an opinion, a rule, a transcript, or a production index and Claude will pull holdings, build timelines, and surface inconsistencies faster than any human can. But asking it to name cases from memory invites fabricated citations. Use a real database to find authority, then use Claude to read it.

Does Claude make up case citations?

It can, like every large language model, when you ask it to recall authority instead of read it. In Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023), two lawyers were sanctioned $5,000 for filing a brief containing fabricated cases. The fix is procedural, not technical: tell Claude up front never to cite anything you haven't pasted, and verify every citation in a citator before it goes into a filing.

Is it ethical for a lawyer to use Claude for legal research?

The ABA addressed this in Formal Opinion 512, issued July 29, 2024, covering competence, confidentiality, client communication, candor to the tribunal, supervision, and fees. Nothing in it prohibits generative AI; it requires you to understand the tool, protect client information, and verify output. Client confidentiality under Model Rule 1.6 is the sharpest issue — check your jurisdiction and your firm policy before pasting privileged material anywhere.

Can Claude replace Westlaw or Lexis?

No, and treating it as a replacement is how people get sanctioned. Westlaw and Lexis do two things Claude cannot: they retrieve authority that actually exists, and their citators tell you whether it's still good law. Claude does something they do poorly — reading a 200-page transcript or a 3,000-page production and telling you what's in it. They're complements, not substitutes.

Is Claude or ChatGPT better for legal research?

Claude has two practical advantages for document-heavy legal work: a very large context window, so a full deposition transcript or several opinions fit in one conversation, and a tendency to follow negative constraints faithfully — if you tell it not to cite anything you haven't pasted, it generally won't. Both models hallucinate authority when asked to recall it. The verification habit matters far more than the model.