Prompt Engineering for Lawyers: A Guide to AI Efficiency

Prompt engineering for lawyers is the practice of strategically crafting instructions to guide AI models like ChatGPT, Claude, and Gemini toward producing accurate, relevant, and useful outputs for legal work. It’s far more than asking a simple question; it’s a developing competency that allows legal professionals to control the AI’s focus, tone, and format. In a field where precision is paramount, this skill is becoming indispensable. It enables attorneys to augment their expertise, delegating time-consuming tasks like initial document review, summarization, and drafting. This shift allows lawyers to reclaim valuable hours for strategic thinking, client counsel, and high-value analysis, transforming generative AI from a novelty into a powerful, efficiency-driving paralegal.

What is Prompt Engineering in a Legal Context? (And What It Isn't)

In the legal field, prompt engineering is the critical bridge between a powerful technology and its practical, responsible application. It is the process of designing inputs for a Large Language Model (LLM) that are specific enough to generate legally relevant and properly formatted content. This isn't about finding a magic button to produce a flawless legal brief. Instead, it’s about treating the AI as an incredibly fast but inexperienced junior associate who requires explicit, detailed instructions.

A weak prompt like, “analyze this contract,” will yield a generic, often useless summary. A well-engineered prompt, however, will assign the AI a role (“Act as a senior partner specializing in M&A”), provide context (“Our client is the seller in this transaction and is risk-averse”), define the task (“Identify all clauses that deviate from a standard 'market' indemnification provision”), and specify the output format (“Present your findings in a table with columns for Clause, Issue, and Recommended Change”). This methodical approach minimizes ambiguity and guides the model toward a useful work product. Prompt engineering is not a replacement for legal judgment but a tool to accelerate the creation of first drafts and initial analyses, which a qualified attorney must then refine and validate.

The Cardinal Rule: Confidentiality and Client Data

Before ever inputting a single word related to a client matter, the foremost consideration must be confidentiality and data security. Breaching client confidentiality is a catastrophic, practice-ending error, and using public-facing AI tools without safeguards is a direct path to that risk. Standard versions of models like ChatGPT often use user inputs to train their future models, which is an unacceptable risk for any privileged or sensitive information.

The non-negotiable first step is to use an enterprise-grade AI solution. Services like ChatGPT Enterprise and the Claude Team plan offer critical security features, most notably a zero-retention policy, meaning your firm's data is not used for training the model. You must confirm this with the provider and ensure it's part of your service-level agreement.

Even with a secure platform, the best practice is to anonymize data rigorously. Before inputting any document, systematically remove or replace all Personally Identifiable Information (PII), company names, financial figures, and specific location details. Use placeholders like [CLIENT_NAME], [OPPOSING_PARTY], [AGREEMENT_DATE], and [PROJECT_ID]. This principle of data minimization reduces risk at every level. You can even use the AI itself to help with this process on a non-sensitive draft.

I am providing you with a case summary paragraph. Please replace all names of individuals with '[PERSON]', all company names with '[COMPANY]', all specific dates with '[DATE]', and all monetary values with '[AMOUNT]'. Do not alter the underlying facts of the summary.

Original Text: 'On May 5, 2023, John Smith of Acme Inc. filed a complaint against Bob Johnson of Widgets LLC, claiming damages of $1.2 million for a breach of contract signed on January 10, 2022.'

Core Prompting Techniques for Legal Precision

To move from basic questions to sophisticated legal tasks, you need a toolkit of prompting techniques. These methods provide the structure and context AI models need to deliver precise, relevant results. Mastering them is the key to unlocking consistent performance.

1. Role-Playing (Persona): This is perhaps the most powerful technique. By assigning the AI a specific role, you frame its entire response. Instead of a generic AI, you're interacting with a specialist. Specify the role, experience level, and even the perspective (e.g., in-house counsel for a startup vs. senior litigator at a large firm). This helps the model adopt the appropriate tone, vocabulary, and analytical lens.

2. Chain-of-Thought (CoT) Prompting: Legal analysis often requires showing your work. CoT prompting instructs the AI to break down its reasoning step-by-step. By adding phrases like “Think step-by-step” or “First, identify the relevant legal principle. Second, apply it to the facts provided. Third, state your conclusion,” you force the model to follow a logical sequence. This makes its output easier to verify and helps you spot flawed reasoning or hallucinations quickly.

3. Few-Shot Learning: AI models learn from examples. In few-shot prompting, you provide one or more examples of the desired input-output pair before giving the actual task. This is incredibly effective for formatting tasks or matching a specific analytical style. For instance, if you need deposition summaries in a specific format, you would provide two or three short, correctly formatted examples before pasting the full transcript and asking it to continue.

Act as a litigation associate. I will provide a short fact pattern. Your task is to perform a simple legal analysis using the IRAC (Issue, Rule, Application, Conclusion) framework. Think step-by-step to structure your response.

Fact Pattern: A customer purchased a latte, telling the barista they had a severe dairy allergy. The barista mistakenly used whole milk instead of oat milk. The customer consumed the latte and had a severe allergic reaction requiring hospitalization.

Begin your analysis.

Mastering Legal Research and Case Summarization

Generative AI presents both a huge opportunity and a significant risk in legal research. Used correctly, it's a powerful tool for accelerating understanding. Used carelessly, it can lead to citing non-existent cases and committing professional malpractice. The cardinal rule is this: AI is for understanding and summarizing provided text, not for finding new, verifiable legal authority.

Models like ChatGPT are notorious for “hallucinating” legal citations. They are designed to predict plausible-sounding text, and a plausible-sounding legal argument includes case names and citation numbers. The AI will invent them with complete confidence. Therefore, never ask a public AI, “What cases support the argument for X?”

Instead, use AI for tasks where you provide the source material. Once you’ve located a relevant judicial opinion from a reliable service like Westlaw, LexisNexis, or a court website, you can use AI to dramatically speed up your review. Paste the full text of the opinion into a model with a large context window (like Claude) and ask for a targeted summary. Go beyond a generic summary by asking for specific elements.

I am providing you with the full text of the court opinion for [Case Name, e.g., Palsgraf v. Long Island Railroad Co.]. Please perform the following tasks:

1. Write a one-paragraph summary of the majority opinion, focusing on the legal reasoning for the holding.
2. Write a one-paragraph summary of the dissenting opinion, highlighting its key points of disagreement.
3. Extract and quote the single sentence that best defines the legal principle of 'proximate cause' according to the majority.
4. List the key facts of the case in a bulleted list.

This approach leverages the AI's language processing strengths while keeping you in full control of the sources and the ultimate legal verification.

Accelerating Contract Review and Analysis

Contract review is one of the most immediate and high-ROI applications of prompt engineering for lawyers. The process of manually reading through dozens or hundreds of pages to spot problematic clauses is tedious and prone to human error. AI can act as a tireless first-pass filter, flagging items for your expert review.

The key is specificity. Don't ask the AI to “review the contract for problems.” Instead, direct its attention to specific risks based on your client's position and the nature of the deal. For example, you can instruct the AI to analyze an agreement's indemnification clause and compare it to a standard mutual indemnification provision you provide. This turns the AI into a powerful comparison engine.

You can also task the AI with hunting for specific types of clauses or missing provisions. For instance, “Scan the attached Master Services Agreement and confirm if it contains a clause addressing data breach notification. If so, extract the clause. If not, state that it is missing.” This is far more efficient than a manual word search, as the AI understands context and can identify conceptually related language even if specific keywords are absent.

I am providing you with two versions of a Limitation of Liability clause. Clause A is from our firm's standard template. Clause B is from a counterparty's proposed agreement.

**Clause A (Our Template):**
'[Paste your standard clause text here]'

**Clause B (Counterparty Proposal):**
'[Paste the counterparty's clause text here]'

Your task is to compare Clause B against Clause A. Identify all substantive differences and explain the legal and financial risk each difference poses to our client, who is the service provider. Present your findings in a simple bulleted list.

This level of directed analysis allows you to focus your limited time on negotiating the points that truly matter, with the confidence that a preliminary check has been thoroughly completed.

Drafting Client Communications and Internal Memos

An attorney's value isn't just in their legal knowledge, but in their ability to communicate complex ideas clearly to different audiences. Generative AI can be a valuable assistant in drafting these communications, helping to overcome writer's block and ensuring the right tone and level of detail.

Whether you are updating a client on a case's progress, explaining the implications of a contract provision to a business stakeholder, or writing an internal memo for a supervising partner, the context you provide in the prompt is everything. You must explicitly define the target audience and the goal of the communication.

For a client email, you might specify a tone that is “reassuring but realistic.” For an internal memo, you might ask for a tone that is “formal and highly technical.” By defining these parameters, you guide the AI to produce a draft that is much closer to a finished product. Always provide the core facts or legal conclusions you want to convey. The AI’s job is to package them effectively, not to invent them. After generating a draft, you must review and edit it to add your personal voice and ensure it perfectly aligns with your strategy and the client relationship. Never copy and paste a client-facing email without a thorough human review.

Act as a senior associate communicating with a long-term client, the CEO of [Client Company Name]. Your tone should be professional, concise, and confident. 

Draft an email updating the client that we have received the initial discovery requests from [Opposing Party Name]. State that the requests are extensive but generally within the scope of what we anticipated. Mention that our team is beginning the process of review and will set up a meeting next week to discuss our strategy for collection and production. Do not promise specific deadlines. Reassure the client that this is a standard phase of litigation and we are well-prepared to handle it. Keep the email under 200 words.

Streamlining Discovery: Summarizing Depositions and Documents

Document-intensive phases of litigation and transactional due diligence are prime candidates for AI-powered efficiency gains. The sheer volume of material in discovery—from deposition transcripts to email caches—can be overwhelming. LLMs, particularly those with large context windows like Claude 3, can process and summarize hundreds of pages of text in seconds.

A key use case is deposition summary. Instead of manually creating a time-consuming digest, you can paste an entire anonymized transcript into the AI and issue precise commands. You can ask for a chronological summary, a list of all instances where a specific person or topic was mentioned, or an extraction of all questions and answers related to a key event. This allows you to quickly get the gist of a deposition or locate critical testimony without hours of reading.

Beyond summarization, you can use AI for preliminary document categorization. After performing an initial privilege and relevance review, you can feed batches of anonymized documents to the AI and ask it to perform tasks like creating a timeline of events based on the documents' content, identifying all documents that mention a specific project code, or flagging documents containing language that suggests employee misconduct. This AI-assisted review must always be supervised and validated by an attorney, but it can drastically reduce the number of hours spent on manual, low-level review tasks.

I am providing excerpts from several project status emails. The documents are not in chronological order. Your task is to extract the key event and the date of that event from each excerpt and then assemble them into a chronological timeline.

[Paste 5-10 anonymized email excerpts here, each with a date and a brief description of an event]

Output a markdown list in chronological order, with each item formatted as: 'YYYY-MM-DD: [Brief description of the event].'

Choosing the Right AI Model for the Task

Not all AI models are created equal. While ChatGPT, Claude, and Gemini have overlapping capabilities, they possess different strengths that make them better suited for specific legal tasks. Understanding these differences allows you to choose the right tool for the job.

Do vs. Don't for Legal AI Prompting

• Do: Use an enterprise-grade AI with a zero-data retention policy.

• Don't: Paste confidential or privileged client information into a public AI tool.

• Do: Provide copious context, including role, audience, and desired format.

• Don't: Ask vague, one-sentence questions and expect a useful answer.

• Do: Use AI to summarize, reformat, and analyze text you provide.

• Don't: Ask the AI to find case law or provide legal advice from its own knowledge base.

• Do: Use placeholders and anonymize data before uploading.

• Don't: Trust the AI with any sensitive PII, ever.

• Do: Verify every single substantive output before it is used in any professional capacity.

• Don't: Copy and paste AI-generated text directly into a client email or court filing.

Overcoming Hallucinations and Ensuring Factual Accuracy

The most significant risk associated with using generative AI in a legal context is the phenomenon of “hallucination.” An AI hallucination is when the model generates information that is plausible-sounding but factually incorrect, or entirely fabricated. This can range from a minor factual error to inventing entire legal cases, complete with fake citations and judicial reasoning—as seen in several high-profile instances that led to court sanctions.

Understanding why this happens is key to mitigating it. LLMs are not databases of fact; they are sophisticated text predictors. Their goal is to generate the next most likely word in a sequence to form a coherent response. A coherent legal argument often contains citations, so the AI obligingly generates text that looks like a citation, without any underlying knowledge of whether that case exists.

To combat this, your prompting strategy must be built around verification and grounding the AI in facts you provide. The safest way to use an LLM is to limit its role to tasks that rely on a provided text (the “grounding context”). These include:

• Summarizing a document you paste into the prompt.
• Answering questions based only on the provided text.
• Reformatting information from one style to another.
• Comparing two pieces of text, like different contract clauses.

When you must venture beyond a provided text, build in cross-examination and skepticism directly into your prompt.

Brainstorm a list of potential legal arguments for a plaintiff in a breach of contract case where the defendant failed to deliver goods on time. For each argument, state the general legal principle it relies on. After the list, include a disclaimer paragraph starting with: 'The following is a speculative brainstorm and does not constitute legal advice. All legal principles and potential arguments must be independently researched and verified using primary legal sources.'

Ultimately, the human attorney is the final and only acceptable validator. Every fact, every assertion, and every citation generated by an AI must be independently verified using traditional, reliable legal research tools before it is even considered for use in a real work product.

Building Your Own Legal Prompt Library

As you begin to use AI more frequently in your practice, you'll find that you are often performing similar tasks repeatedly: summarizing depositions, reviewing NDAs, drafting status updates. Instead of rewriting prompts from scratch each time, the most efficient approach is to build a personal or firm-wide library of proven, effective prompts.

Start by creating a simple document or spreadsheet. When you craft a prompt that delivers an exceptionally good result, save it. Give it a clear title, like “NDA Review Prompt - Risk-Averse Client” or “Deposition Summary - Chronological.” Add a short description of what the prompt does and when to use it. You may also want to include a sample of the ideal output so you or your colleagues can see the target format.

Over time, you can refine and expand this library. You might create several variations of a contract review prompt, each tailored to a different client position (e.g., buyer vs. seller, licensor vs. licensee). This process of iterating on your prompts is where true mastery of the skill develops. A well-organized library not only saves immense time but also helps standardize the quality and consistency of AI-assisted work across your team. It turns prompt engineering from an individual ad-hoc effort into a scalable, institutional capability. For a head start, you can explore pre-built templates in our Lawyers prompt pack hub to see how structured prompts are built for various legal use cases.

Key Takeaways

• A New Competency: Effective prompt engineering is a core skill for the modern lawyer, enabling you to augment your legal expertise and drive efficiency, not replace your judgment.
• Confidentiality is Paramount: Never use public AI tools for client work. Only use enterprise-grade platforms with zero-data retention policies and always anonymize sensitive information.
• Context is King: The quality of your AI output is directly proportional to the quality of your input. Use specific techniques like role-playing, providing examples, and defining the output format to guide the model.
• Trust but Verify (Methodically): Treat the AI as a hyper-efficient but fallible assistant. It is excellent for summarizing and analyzing text you provide. It is dangerous for finding net-new case law or facts, which it frequently hallucinates. Always verify.
• Build a System: Create a personal or team library of your best prompts to save time, standardize quality, and scale the benefits of AI across recurring tasks in your practice.