Prompt Engineering for Lawyers: A Practical Guide

The gap between lawyers who get useful output from AI and those who get expensive garbage is almost never the tool. It's the prompt. Most attorneys approach AI the way they approached their first summer associate: vague assignment, high expectations, profound disappointment. The difference is...

Learning to write good prompts is a billable-hour-adjacent skill now. Courts are already scrutinizing AI use — Mata v. Avianca (S.D.N.Y. 2023) made that brutally clear — and the ABA's Model Rule 1.1 commentary on competence has been interpreted to require understanding the tools you deploy. So let's get specific.

The Anatomy of a Legal Prompt

Every effective legal prompt has four components. Skip one and output quality degrades predictably.

Role tells the AI who it's being. Context gives it the case or matter background. Task specifies exactly what you want produced. Constraints define format, length, jurisdiction, and tone.

The skeletal structure looks like this:

You are [role]. Given [context], [task]. Your output should [constraints].

That's it. Everything else is elaboration on those four levers.

Five Patterns That Actually Work

1. The Jurisdiction Anchor (Research)

The most common research prompt mistake is forgetting that AI models train on law from everywhere, which means their default answer is jurisdictional soup.

Before: "What are the elements of promissory estoppel?"

After: "You are a California litigation associate. Under California law, including Kajima/Ray v. Los Angeles Unified School District (2000) and its progeny, list the elements of promissory estoppel, note any circuit split with federal courts sitting in California, and flag any post-2020 appellate decisions that have tightened or loosened the detrimental reliance requirement. Output as a numbered list with citations."

The second prompt constrains the jurisdiction, demands a specific doctrinal frame, flags a known tension point, and requests a usable format. You'll still verify every citation — that's non-negotiable — but you're starting from the right zip code.

2. The Persona Ladder (Drafting)

When drafting, the role instruction matters more than most lawyers realize. "Draft a non-compete clause" will produce something technically coherent and strategically useless. Instead, stack the persona.

Before: "Draft an NDA for my client."

After: "You are a senior M&A associate drafting for a Series B SaaS company in Delaware. The company is sharing proprietary machine learning training data with a potential acquirer. Draft a mutual NDA with a 24-month term. The definition of Confidential Information should be broad enough to cover model weights and synthetic data sets. Include a residuals clause but make it narrow. Flag where we should negotiate hard if the other side pushes back."

The negotiation flag instruction is what separates this from a template pull. You're not just getting a document; you're getting embedded strategy.

3. The Red Team Frame (Contract Review)

Lawyers reviewing contracts should use adversarial prompting. Tell the AI it's opposing counsel.

Before: "Review this indemnification clause."

After: "You are aggressive plaintiff's counsel in a commercial dispute. Read this indemnification clause and identify every ambiguity or gap that you would exploit in litigation. Then re-read it as defense counsel and identify which of those vulnerabilities could be cured by redline. Show both analyses side by side."

This dual-perspective structure is particularly useful for merger agreements or tech licensing deals where the drafting party assumes language is tighter than it is. The EU AI Act's liability provisions, now in force, make this kind of clause-level rigor more commercially urgent than it was two years ago.

4. The Stepwise Memo (Complex Analysis)

For anything involving multi-factor tests — Chevron deference replacements post-Loper Bright Enterprises v. Raimondo (2024), or the Dataphase preliminary injunction factors — force chain-of-thought reasoning explicitly.

Before: "Does my client have a strong preliminary injunction case?"

After: "Apply the Dataphase factors to the following fact pattern, step by step. Do not summarize until you have analyzed each factor individually. For each factor, state the applicable sub-rule, apply it to the facts below, and rate the factor as favorable, unfavorable, or contested. Then provide a summary paragraph with an overall assessment. Facts: [insert facts]."

Stepwise instructions prevent the AI from jumping to a conclusion and reverse-engineering the analysis, which is exactly what a bad law clerk does and exactly what you cannot afford to hand to a partner.

5. The Template Extractor (Document Automation)

If you're building clause libraries or automating intake documents, use extraction prompts before drafting prompts.

Before: "Write a force majeure clause."

After: "Extract every force majeure clause from the three contracts below. Create a comparison table showing: triggering events listed, notice requirements, duration of excuse, and whether pandemic/government action is explicitly included or excluded. Then draft a new clause that takes the most favorable elements for a commercial landlord in New York post-COVID."

This is how you build institutional knowledge systematically rather than prompting from scratch every time.

The Most Expensive Mistakes

Asking for conclusions without showing your reasoning chain. You get a confident answer with no audit trail. That's a malpractice exposure, not a memo.

Omitting the verification instruction. Every research prompt should end with "flag any citations you are uncertain about." It doesn't eliminate hallucination but it creates a documented checkpoint.

Treating AI like a search engine. Keyword queries produce keyword responses. Ask like you're briefing a junior attorney: background, objective, specific ask, format, deadline constraints.

Ignoring confidentiality. Know your firm's policy and your client's obligations before pasting a single paragraph into any AI interface. Samsung learned this the hard way in 2023; your client doesn't need a similar lesson.

The Bottom Line

Prompt engineering isn't a tech skill. It's a communication skill dressed in new clothes. Lawyers already know how to give precise instructions under constraint — that's what a brief is. Apply that same discipline to your AI inputs and the outputs become defensible, useful, and worth the subscription cost.

The lawyers who figure this out in the next eighteen months will have a structural advantage in research speed, drafting quality, and matter economics. The ones who keep typing "summarize this contract" into a chat box will keep wondering why the technology doesn't work.

It works. The prompt is the problem.