How I Use Claude as a Legal Research Assistant (Without Risking Malpractice)

Every lawyer I talk to has the same worry about AI: the hallucinated-citation horror stories. A model invents a case, a filing cites it, a judge notices, and suddenly it's a news story about a lawyer getting sanctioned. That risk is real. But the answer isn't to avoid Claude entirely — it's to use it for the parts of research it's actually good at, and never for the part it isn't.

What Claude is genuinely useful for

Claude is excellent at the unglamorous first 80% of research: getting oriented in an unfamiliar area of law, turning a messy fact pattern into a list of legal issues, drafting a research memo skeleton, and explaining doctrine in plain English so you can sanity-check what you already half-remember from law school.

• Turning a client's rambling description of events into a clean issue-spotting outline.
• Explaining an unfamiliar area of law (say, a state-specific wage-and-hour rule) at a level you can then verify against primary sources.
• Drafting the structure of a research memo — headings, the questions presented, the framework — before you fill it in with verified authority.
• Summarizing a case or statute you paste in directly, so you're working from the actual text rather than the model's memory of it.

The one rule I never break

I never let Claude tell me a case exists that I haven't independently verified in Westlaw, Lexis, or a court's own database. If Claude names a case, I treat the name and citation as a lead to check, not a fact. This single habit eliminates almost the entire malpractice risk people worry about, while keeping all of the speed benefit.

Claude is a research assistant, not a source of truth. Treat every citation it gives you exactly the way you'd treat a citation from a first-year associate you've never worked with before — promising, unverified, and worth checking.

The prompt I actually use

Here's the prompt template I paste into a fresh Claude conversation whenever a new matter lands on my desk. It forces Claude to separate what it's confident about from what needs verification, instead of blending both into one confident-sounding paragraph.

I'm a lawyer researching [area of law] for a client matter involving [one-sentence fact pattern]. I am in [jurisdiction].

1. List the legal issues this fact pattern raises, in order of importance.
2. For each issue, explain the general legal framework in plain English.
3. Flag anything jurisdiction-specific that I should double-check, since rules vary significantly by state.
4. Do NOT cite specific case names or statute numbers unless I have given them to you. If you're not certain of an exact citation, say so explicitly instead of guessing.
5. End with a short list of search terms I should use in Westlaw/Lexis to verify each point.

That last instruction matters more than it looks. Explicitly telling Claude not to guess at citations dramatically reduces the chance it produces something that looks authoritative but isn't — and gives you a research plan instead of a false sense of completion.

Where this saves the most time

The biggest win isn't any single research task — it's the intake call. Fifteen minutes after a client describes their situation, I can have a structured issue list and a jurisdiction-aware framework ready, instead of starting from a blank page. That turns a client conversation into an actual work product faster, which matters more than people expect when you're billing by value rather than by the hour.