How to Write AI Prompts for Legal Work Without Getting Dangerously Generic Answers

Legal prompts fail in a specific, predictable way that other prompt categories mostly don’t: they come back sounding confident and complete while being quietly wrong for your situation. Ask ChatGPT to "explain this clause" or "draft a demand letter" with no other input, and it will produce something fluent, well-formatted, and generic enough to apply to almost any contract or dispute in the country — which also means it applies precisely to none of them. The output reads like it was written by someone who understands law in general and knows nothing about your lease, your state, or your specific fact pattern, because that is exactly what happened.

Why "draft me a contract" is the wrong prompt

The problem isn’t that AI tools are bad at legal language — they’re often fluent in it, sometimes too fluent, producing boilerplate that sounds more authoritative than it should. The problem is that legal answers are almost entirely dependent on variables a vague prompt never supplies: which state or country’s law applies, what kind of agreement or dispute this actually is, what has already happened, and what you’re trying to accomplish. A contract clause that’s standard and enforceable in one jurisdiction can be unenforceable or entirely absent from the law in another. A one-line prompt forces the model to either pick a jurisdiction at random (usually defaulting to generic US common-law assumptions) or hedge so heavily the answer says nothing useful. Neither outcome is what you asked for, and neither is obviously wrong-looking, which is what makes it risky rather than just unhelpful.

The five layers a legal-grade prompt needs

Role and jurisdiction — tell the model what lens to use and, critically, under which state, country, or governing law, since the same clause can mean opposite things elsewhere ("acting as a contracts-literate assistant, analyzing this under California law"). Document or matter type — a residential lease is not a commercial lease is not an NDA is not a demand letter, and each has different standard clauses and different red flags to check for. The specific ask — "explain what this means in plain language," "flag anything unusual compared to a standard version," and "draft a first-draft letter" are three different tasks that need three different prompts, not one prompt trying to do all three. Constraints — explicitly state that this is not a substitute for a licensed attorney and should not be treated as final legal advice, and say what to avoid (invented case law, absolute claims like "this is definitely enforceable"). Output format — plain-language summary, a clause-by-clause breakdown, a redline-style comparison, or a formatted draft letter are all different shapes, and the model will guess the wrong one if you don’t specify.

Four legal prompt tasks worth knowing how to structure

Clause explanation: give the model the exact clause text, the document type, and the jurisdiction, and ask specifically what it obligates each party to do and what happens if it’s breached — a vague "what does this mean" invites a dictionary-style answer instead of one grounded in the actual contract. Plain-language translation of a lease or terms of service: the useful prompt isn’t "summarize this," it’s "identify anything in here that is unusual or more restrictive than a standard [document type] in [jurisdiction], and explain each one in plain language" — that framing gets you the parts that actually matter instead of a compressed restatement of the whole document. First-draft correspondence (a demand letter, a cease-and-desist, a dispute notice): specify the facts in order, the outcome you want, and the tone (firm but professional, not adversarial), and explicitly ask for a draft to be reviewed and finalized by a licensed attorney before sending — not a document to send as-is. Side-by-side comparison: when you have two versions of an agreement (a redline, or an old lease versus a new one), naming both documents and asking specifically "what changed, and what does each change favor" produces a far more useful answer than pasting both in and asking to "compare" with no direction.

Jurisdiction is not a footnote — it’s an input

Most people treat jurisdiction as an afterthought, something to mention if it comes up. In a legal prompt it needs to be one of the first things stated, because it changes what the correct answer even is, not just how it’s phrased. "Is a 30-day notice period enforceable" has a different answer in different states and different countries, and a model given no jurisdiction will answer as if one exists — usually a generalized, US-leaning default — without flagging that the answer would change elsewhere. Naming the governing jurisdiction up front (the state on the lease, the country a contract is signed under, the venue clause in an existing agreement) turns a guess into an actual answer, and naming it explicitly also makes it obvious to you, reading the output, whether the model is missing a piece of the puzzle it should have asked for.

What a legal prompt should never ask an AI to do

Replace a licensed attorney on anything with real financial, contractual, or legal consequences — the honest framing for every one of the tasks above is "produce a first draft or explanation to review with counsel," not a final answer to act on directly. Generate or rely on case citations without independently verifying they exist — language models can produce citation-shaped text for cases that were never decided, and this has already caused real, publicly documented sanctions for attorneys who filed briefs without checking. Receive privileged or genuinely sensitive client documents pasted into a general-purpose consumer AI tool without first checking your organization’s data-handling policy — confidentiality obligations don’t pause because the recipient is a model instead of a person. A well-built legal prompt makes all three of these boundaries part of the instruction itself, not an afterthought you remember to add later.

A worked example, before and after

Weak prompt: "explain this lease clause to me." No jurisdiction, no document type beyond "lease," no stated purpose, and no format — the model has to guess whether you want a legal analysis, a plain-language summary, or a check for enforceability, and it will likely produce a little of each without fully committing to any. Structured prompt: "Acting as a contracts-literate assistant (not a substitute for a licensed attorney), analyze the following clause from a residential lease governed by New York State law: [clause text]. Explain in plain language what it obligates the tenant to do, flag whether the notice period is shorter than what New York typically requires for this lease type, and note explicitly if this is something to raise with a local attorney before signing. Format as a short plain-language paragraph followed by a one-line flag if anything is unusual." The second version tells the model exactly which state’s norms to check the clause against, exactly what to explain, exactly what to flag, and exactly how long the answer should be — there’s very little left for the model to invent, which is the entire point.

Common mistakes

Skipping jurisdiction and getting an answer that quietly assumes a default that may not apply to you at all. Asking the model to "just draft the contract" with no starting terms, which produces a maximally generic template rather than something reflecting your actual deal. Treating a fluent, confident-sounding answer as equivalent to a verified one — fluency and correctness are not the same property, and legal text is exactly where that gap is most expensive to discover late. Pasting an entire real, sensitive document into a prompt for a quick check without first considering whether that document should be leaving your hands at all.