Learn

Prompt Engineering for Lawyers

The difference between a useless AI response and a brilliant one is almost always the prompt. Prompt engineering is the skill of communicating effectively with AI systems — and for lawyers, it is quickly becoming as essential as legal research itself.

The CRAFT Framework

A structured approach to building effective legal prompts. Every great prompt addresses these five elements — Context, Role, Action, Format, and Tone.

C — Context

Set the scene. Tell the AI who you are, what area of law you practice, and what jurisdiction applies. The more context you provide, the more relevant the output.

You are a senior corporate attorney in New York reviewing a commercial lease agreement governed by New York law.

R — Role

Assign the AI a specific role or persona. This shapes the tone, depth, and perspective of the response. A "junior associate" produces different output than a "supervising partner."

Act as a litigation strategist advising on summary judgment motions in federal court.

A — Action

State exactly what you want the AI to do. Be specific about the task, the format, and the level of detail. Vague instructions produce vague results.

Analyze the following contract clause and identify all provisions that could expose my client to unlimited liability. Present your findings as a numbered list with risk severity ratings.

F — Format

Specify how you want the output structured. Bullet points, numbered lists, headings, tables, executive summaries — define the format explicitly.

Structure your response with: (1) Executive Summary (3 sentences), (2) Key Findings (numbered list), (3) Recommended Actions (bullet points), (4) Risk Assessment (table).

T — Tone

Define the audience and appropriate tone. A client memo reads differently than an internal research note or a judicial brief. Tell the AI who will read this.

Write in a professional but accessible tone suitable for a corporate client who is not a lawyer. Avoid legal jargon where possible; define it where necessary.

CRAFT in Action

Here is a complete prompt using all five CRAFT elements together:

[Context] I am a commercial litigator in Texas representing the defendant in a breach of contract dispute. The contract is governed by Texas law and involves a software licensing agreement valued at $2.4 million. [Role] Act as an experienced litigation strategist who specializes in commercial contract disputes in Texas state courts. [Action] Analyze whether the doctrine of substantial performance applies to my client's situation. My client delivered 90% of the contracted software modules on time, but the final module was delayed by 6 weeks. The plaintiff is seeking full contract damages. [Format] Structure your analysis as: (1) Elements of Substantial Performance under Texas law, (2) Application to These Facts, (3) Strongest Counterarguments, (4) Recommended Strategy. Use bullet points within each section. [Tone] Write for an internal strategy memo that will be reviewed by the senior partner. Be direct, analytical, and candid about weaknesses in our position.

Techniques in Practice

See the difference between weak and effective prompts. Each technique shows what to avoid, what works, and why.

Assign a Professional Role

Instruct the AI to adopt a specific professional perspective. This shapes the tone, depth, and focus of the response.

Weak Prompt

Tell me about lease agreements.

Effective Prompt

Act as a real estate attorney in California with 15 years of experience. I need you to analyze the clauses of a commercial lease agreement under California Civil Code and relevant case law from the California Court of Appeal.

Why it works: The role contextualizes the response. A "California real estate attorney" will produce different terminology, legal references, and approaches than a "UK solicitor" or an "Australian barrister."

Lock the Jurisdiction

Always specify the country, state, or region and the applicable law. AI trained on multiple jurisdictions may mix legal frameworks without warning.

Weak Prompt

Draft a non-compete clause.

Effective Prompt

Draft a non-compete clause for an employment agreement that complies with California Business and Professions Code Section 16600, considering recent California Supreme Court rulings on the enforceability of employee non-compete restrictions following the 2024 amendments.

Why it works: A non-compete clause valid in Texas may be unenforceable in California or void in North Dakota. Jurisdiction determines the legal boundaries of what you can agree to.

Define the Output Format

Specify exactly how you want to receive the information: table, numbered list, memorandum, legal opinion, comparison chart.

Weak Prompt

Compare business entities in Delaware and California.

Effective Prompt

Create a comparison table with these columns: (1) Entity Type, (2) Formation Requirements, (3) Minimum Members/Shareholders, (4) Liability Protection, (5) Tax Treatment, (6) Governing Statute. Include: LLC, S-Corp, C-Corp for both Delaware and California. Cite the specific statutory provisions from the Delaware General Corporation Law and the California Corporations Code.

Why it works: Without a defined format, the AI chooses its own, which may not serve your purpose. A comparison chart for a client is different from an internal research memo.

Provide Sufficient Context

Include the relevant facts, parties involved, deadlines, and any constraints or preferences. The more context, the more precise the response.

Weak Prompt

I need a services agreement.

Effective Prompt

I need a draft professional services agreement between a technology company (provider) incorporated in Delaware as an LLC and a law firm (client) in New York. The scope is development of a document management system. Duration: 12 months. Value: $250,000 payable in quarterly milestones. Must include confidentiality clauses, intellectual property ownership over the developed software, service level agreements (SLA), and dispute resolution by arbitration under AAA Commercial Rules in New York.

Why it works: A generic contract is useless. Specific details allow the AI to generate a draft that is a viable starting point, not a document requiring total rewriting.

Use Chain-of-Thought Reasoning

Ask the AI to reason step by step before giving its conclusion. This reduces errors and makes the reasoning verifiable.

Weak Prompt

Can my client sue?

Effective Prompt

My client suffered damages from a defective product in Florida. Analyze step by step: (1) What law applies? Consider Florida's product liability statute (Fla. Stat. §768.81) and strict liability under Restatement (Third) of Torts. (2) Are the elements of strict product liability met? (3) What is the applicable statute of limitations? (4) Against whom does the action lie? (5) Which court has jurisdiction? (6) Conclusion with recommendation on viability.

Why it works: Step-by-step reasoning produces more structured analyses and allows you to identify the exact point where the AI makes an error, facilitating verification.

Iterate and Refine

Do not expect the perfect response on the first attempt. Use initial responses as a starting point to go deeper.

Weak Prompt

[Accept the first response without questioning]

Effective Prompt

Thank you for the analysis. Now go deeper on point 3 regarding the statute of limitations. How does Florida's discovery rule affect our case if the injury manifested two years after purchase? Is there Florida Supreme Court precedent on when the limitations period begins to run for latent product defects?

Why it works: The best AI responses come from iterative conversations. The first response is a draft; follow-up questions turn it into a useful analysis.

Prompt Patterns for Legal Work

Beyond the CRAFT framework, these proven patterns address common challenges in legal AI interactions. Copy and adapt them to your practice.

The Jurisdiction Lock

Force the AI to stay within your jurisdiction by making it explicit and asking for confirmation.

All analysis must be based exclusively on [State/Federal] law. If you are unsure whether a principle applies in this jurisdiction, say so explicitly rather than guessing.

The Verification Demand

Instruct the AI to flag uncertainty and separate verified facts from generated analysis.

For each case or statute you cite, indicate your confidence level: VERIFIED (you are certain this exists) or UNVERIFIED (this may need human verification). I will check all citations independently.

The Structured Output

Request specific sections and formatting to get consistently usable results.

Organize your response under these exact headings: Summary, Legal Analysis, Risk Factors, Recommended Next Steps. Under each heading, use bullet points. Keep the total response under 500 words.

The Devil's Advocate

Ask the AI to argue against your position to stress-test your reasoning.

Now take the opposing counsel's perspective. What are the three strongest arguments against my position? For each, cite the legal principle or precedent that supports it.

The Iterative Refinement

Build complex outputs through multiple focused prompts instead of one massive request.

Step 1: List all potentially applicable legal theories. Step 2: For each theory, assess the strength of evidence. Step 3: Recommend the top two theories to pursue and explain why.

Practice Makes Proficiency

Prompt engineering is a skill that improves with practice. Start with our Quick Wins — each one includes a carefully crafted prompt you can study, modify, and learn from.

Try a Quick Win Browse Prompt Library

Ready for structured learning? Explore the Learning Program →

Comments

Loading comments...