Prompt engineering for lawyers

Ask a tool like ChatGPT a question and the answer feels like magic. It is not. Under the hood, the model does one thing over and over: it predicts the next word. Modern generative AI runs on transformer architecture, and the innovation that changed everything is called self-attention. It lets the model weigh how every word in your prompt relates to every other word, much the way you do when you read.

Walk through what happens when you type "What are the consequences of copyright infringement?" First the model chops your question into tokens, small chunks of text it can work with. Those tokens pass through layers that size up what matters: one layer flags "consequences" as the payload, another ties it to "copyright infringement." Then the model writes back one token at a time, choosing each next word by probability. "The" wins the opening slot. In this context, words like "penalties," "fines," and "damages" rank high. Token by token, out comes a complete answer: fines, damages, injunctions, maybe criminal charges.

Why should a lawyer care? Because a prediction engine has no built-in commitment to the truth. It commits to the plausible. Every skill on this page exists to close that gap. You are learning to shape the probabilities. That is all prompting is.

A few distinctions worth keeping straight. General chatbots (ChatGPT, Gemini, Copilot) draw on broad knowledge: versatile, but not specialized. Tools like Westlaw AI work from structured legal data and are more reliable inside their lane. And a tool is either cloud-based, meaning your prompts travel to someone else’s servers, or local, meaning they stay on hardware you control. That difference matters every time client information is involved.

These tools sometimes get legal questions wrong, and the reason is usually a mismatch: they do not think the way we think. Left alone, a model does not always understand that you want the truth, or that "I do not know" beats an invented answer. Creativity is its default setting. Accuracy takes guidance.

So play to the strengths. AI is terrific at improving writing style, because style asks for options, not facts: it has processed more sentences than any writer alive, so let it brainstorm ten ways to recast yours. It summarizes defined text fast and well: the cases, briefs, and statutes you paste in. It generates themes, storylines, and headlines. It drafts starter text that beats a blank page. It extracts and organizes information into clean tables. And it explains and translates complex material for any audience you name.

Now the weak spots. It can misstate the law and cite cases that do not exist, in prose that sounds completely sure of itself. It handles jurisdiction-specific nuance badly. And it has no client instincts: no empathy, no judgment, no feel for what this judge, on this record, needs to hear. It is a tool, not a wand.

Generative AI works so well at first glance that it tempts you to put your pen down. Resist. You cannot use this tool, or any tool, well if you do not know what the final product should look like. Some call this the pilot era of lawyering: technology handles more of the small details while the lawyer sits at the controls, adjusting, steering, correcting. That only works if you can still fly the plane yourself.

I like a homelier picture: the bread machine. I make a lot of bread, and the machine does the heavy lifting. No kneading by hand, no waiting on the rise. But loaves still come out bad, and when they do, I know how to adjust, because I know how to make bread. If I did not, I would get the same loaf every time, good or bad, and I would be taking orders from an appliance. So decide now: do you want to be a bread machine assistant, or do you want the bread machine to assist you?

That question is the reason this product exists. Every tune in the menu below keeps the judgment calls on your desk: the AI drafts and suggests; you direct and decide.

You owe clients competence and diligence, and you owe courts candor. Use AI to draft a paragraph and you are responsible for that paragraph as if you wrote it yourself. Competence now includes knowing how your tools work and staying reasonably current as they change. Check the rules that govern you, too: California bar guidance says lawyers should get client approval before using generative AI on a matter, and some courts require you to disclose AI use in your briefing.

Confidentiality has two exposure paths, and you need to check both. First, transmission and storage: like any cloud tool, an AI service streams your data out, holds it while working, and streams a response back, and it can be compromised at any point along that path. Second, and unique to AI: training data. Your prompts are gold to AI companies, which pay people to do all day what you do for free. Confirm your inputs will not be folded into a model’s training in a way others could extract.

Keep bias on the radar as well. These models learn from enormous datasets that include discriminatory content, and garbage in means garbage out. Read outputs with a critical eye for stereotypes and slanted framing no careful lawyer would knowingly sign.

AI output can be wrong in ways that read as right. It can miss the depth of a novel issue, mangle a doctrine, and produce a citation that does not exist, all in confident, fluent prose. So build verification into your process instead of your intentions: every authority checked by a human, every quotation confirmed against the source, every legal statement traced to something you read yourself.

And make the AI help. Tell it to flag what it is unsure of. Give it standing permission to say "I do not know." Give it an exit so it never needs to invent one:

Specific and narrow beats broad and vague, every time. "Draft a contract" hands the model the whole world to parse. Start with a concrete verb (summarize, brainstorm, rewrite, extract), deliver one point at a time, and keep the task tight. You can always go deeper on the next turn.

Then wall your text off. Delimiters separate your instructions from the text you want worked on and from your examples, so the model never confuses the document with the directions. Anything works as a delimiter if it cannot be mistaken for a word, and per the research, *** works very well.

A role tells the model how to approach the task and how to shape the answer. A research-assistant role produces neutral summary. An advocate role produces persuasion. A judge role shows you the pushback your argument will draw before a real judge supplies it. Few tasks fail to benefit from one.

Open your prompt with a short phrase that sets the role, and add a quality adjective ("a well regarded lawyer," "a renowned legal writer") to point the model at its best material. For style work, mix favorite legal and non-legal writers until the blend sounds right.

One distinction. These roles are personas you hand the AI. Prompt Pilot’s practice and role lenses point the other direction: they tune the work to your field and your seat. Use both.

These models have two gears: a fast gear that pattern-matches an answer, and a slower reasoning gear that works in steps. Complex legal tasks need the second gear, and chain-of-thought prompting is how you shift into it. The simplest trigger is a magic phrase. "Let’s think step by step" measurably improves results, and one wording performs even better in the research: "Let’s work this out step by step to be sure we have the right answer."

You can also list the steps yourself: first find the relevant provisions, second analyze how they apply, third conclude. Or make the model design its own plan and follow it. And when you want your work checked, do not let the AI grade first. Make it solve the problem independently, then compare:

The model will follow nearly any structural guardrail you give it, so give it several. Control depth with numbers: not "improve this verb" but "give me 10 versions, each using a different verb." Confine length the same way when it rambles. Set ground truths that overwrite its instincts: "All text within quotation marks is accurate and should not be modified." Give it a reading grade or a named audience. And name the format you want back, because the model is exceptional at filling patterns:

The AI world calls these few-shot prompts: you show the model a few exemplars of what you want, and it learns the pattern with startling speed. The research guidance: provide at least four examples when you can, make them diverse, randomize their order, structure them as prompt-and-answer pairs, and include wrong answers next to right ones so the model learns the difference.

Background changes everything, so supply it and label it. State "The following information is context," wall it off with delimiters, and load in your jurisdiction, the key cases, the bullet points from your draft. (Every tool has a context window, a ceiling on how much text it can hold at once, so spend the space on what matters.)

Then add hints. In recent studies, a short "Hint:" list of keywords significantly improved response quality. Summarizing a case? Tell the model which concepts the summary must center on:

You will rarely get the best answer on the first try, and you do not need to. The model remembers the conversation, so treat it like a junior colleague: say what it got right, what it got wrong, and what to change. Quote its own response back to it. If it seems unsure, tell it to ask you clarifying questions.

A few research-backed moves for the follow-up game. Phrase instructions affirmatively: say what to do, not what to avoid. Repeat key instructions at the end of a long prompt, because models weight the ending most (recency bias). Have the model generate background knowledge first, then answer on top of it. For a prompt that really matters, run five versions and compare; five is the number the research suggests. And stack your prompt types. Role, task, parameter, hints, four short lines:

AI is a machine for producing averages: the average structure, the average motion, the average sentence. Averages are a fine start and a terrible finish. When average will not cut it, push. "Provide the contrarian view on this topic." "Come up with three unique angles most lawyers wouldn’t use." "Brainstorm the most unusual approach for the verb in this sentence." Creativity prompts rarely deliver a finished product. They deliver the spark you could not find alone.

How far can this go? I was consulting with an elite appellate team on a pro bono case, impact litigation where every word counted. They had run dozens of versions of one heading and none was right. We tried a couple of generic prompts. Still flat. The case involved parody, so just for fun, with the team watching on Zoom, I typed: "Try rewriting this heading in the form of a unique poem." Silly, right? The model started producing verses, and one line carried a turn of phrase we all recognized instantly as the seed of the perfect heading. A little massaging later, we had our winner. I am not telling you to commission poetry for every brief. I am telling you to experiment, because you never know which prompt shakes the right answer loose.

Where does AI earn its keep in a legal writing practice? Style, first and best: "Pick better verbs in this sentence to convey [the emotion you are aiming for]." Grammar and mechanics: "Correct the punctuation in this sentence." Summarizing and extracting: "Summarize the statute and keep it to one paragraph." Starter drafts for rote sections and templates. Brainstorming: "Write a theme for two people fighting over a retirement account." Classifying and formatting: "List these cases in alphabetical order." And collaboration: run a presentation past it, ask for the weaknesses in your oral argument, or have it turn your margin notes into feedback a colleague can actually use.

One favorite deserves its own box. A few follow-ups after this prompt, you will have a writing playbook for yourself or your whole team:

In the end, generative AI can be a fantastic legal writing partner. It brainstorms. It prods you to see a sentence differently. It organizes and extracts. But the results come from trial and error with you in the pilot seat: questioning, guiding, editing, verifying. Know what it is good at. Know what it needs help with. Know what it is not for at all. Then go fly.