Context Engineering for Attorneys
How to change the way you ask AI tools questions to achieve substantially better results and reduce frustrating back and forth.
QUESTION PRESENTED:
Does the way I ask an AI model a question impact the answer?
SHORT ANSWER:
Yes. The structure of a prompt substantially impacts the LLM’s response. The context that you provide within your question, the constraints you give the agent, and the guidance you give on what a good response looks like all affect the quality of the answers you receive.
DISCUSSION
I. What is Context in this…Context?
An LLM’s context window is how much information the model can access during your conversation. Context allows an LLM to interact with you specifically, addressing your questions with information needed to respond. More context usually leads to better results.
II. Building Context into Your Prompts
Let’s start with a common problem that you may be asking an AI tool to solve. You want a model to perform an initial review of some discovery.
It is important to remember that public AI tools such as Google Gemini and ChatGPT do not offer protection or safeguards for the data you upload to their services. If the information you are reviewing is privileged or contains PII, you cannot guarantee it’s custody unless you run your own AI tools at your office or home.
Common problem: Perform an initial review of Defendant’s discovery responses and identify boilerplate objections or missing documents.
Basic Prompt:
Perform an initial review of Defendant’s discovery responses look for boilerplate objections.
What’s the problem?
An LLM will respond generally to your prompt and likely miss some objections. What is boilerplate? How should it handle multiple objections? Prompting the model like a new employee helps.
Context Engineered Prompt:
Role: You are a senior litigation paralegal and discovery specialist at a premier plaintiffs’ personal injury firm. You are an expert at identifying evasive discovery responses and non-compliant objections under the Rules of Civil Procedure.
Task: Review the provided “Defendant’s Responses to Plaintiff’s First Request for Production” and generate a Discovery Deficiency Table.
Context: Our client was injured in a commercial trucking collision. We are seeking maintenance records and driver logs. The defendant is being obstructive.
Input Data: [PASTE DISCOVERY RESPONSES HERE]
Instructions: > 1. Analyze each response for “boilerplate” objections (e.g., “overly broad, unduly burdensome”) that lack specificity. 2. Check if the defendant agreed to produce documents but failed to provide a date certain for production. 3. Identify where the defendant claims documents do not exist for categories that are legally required to be maintained (e.g., DOT logs).
Output Format: Provide a Markdown table with the following columns: | RFP # | Summary of Request | Type of Deficiency | Analysis/Reasoning | | :--- | :--- | :--- | :--- |
Constraint: Do not offer legal advice. Limit your analysis strictly to whether the response meets the technical requirements of the Rules of Civil Procedure regarding specificity and completeness.
III. Breaking it Down
A context engineered prompt is more explicit, detailed, and constrained. The context engineering increases the likelihood of getting the best answer because you told the model how to respond.
1. Role Prompting
Most AI models are text based and respond in the context of their perceived role. By giving the AI a role, you prime it with a professional and detail-oriented focus. It narrows the “probability space” of the model’s vocabulary to legal terminology and civil procedure rather than general conversation.
2. Task & Context
AI models like Gemini and ChatGPT perform better when they understand the stakes of the problem. Telling the model, “this is a trucking case,” helps it prioritize certain documents (like maintenance logs) as high-value.
3. Granular Instructions
This is a form of step-by-step prompting. You give the model an example of what a problem looks like so it doesn’t make one up. This reduces hallucinations by giving it a checklist to follow.
4. Structured Output
Asking for your answer in a table forces the LLM’s responses into categories. The categorized data is easier to parse later on.
IV. The Power of Constraints
Telling your agent what not to do or what isn’t allowed in your prompt is a powerful tool in the legal context. Constraints give you power over how your answer is presented. Do you want a list of specific objections by interrogatory number? Just ask:
“Format the deficiency report as a table with columns for ‘Request Number,’ ‘Defendant Objection,’ and ‘Suggested Counter-argument.’”
Want to ensure the model doesn’t invent answers? Just ask:
“Answer using only the provided text. If the answer is not in the text, state ‘Information not found.’”
V. Thoughtful Prompting
Context engineering will increase your productivity using LLMs and reduce time spent correcting errors. Spending time building a prompt with proper context helps you better define the problem you wish to solve and more easily recognize issues.