Generative AI is only as good as the instructions it receives. In legal practice, a vague prompt like "what is the law on negligence" returns generic boilerplate that no attorney would use. Extracting real value from these tools requires legal prompt engineering: a structured form of giving the model instructions that dictate how it should reason, what it should produce, and how the output should be formatted.1, 2

The Four Elements of a Strong Legal Prompt

A high-quality legal prompt has four elements: role, task, context, and format. Write clearly, use active voice, avoid double negatives, and strip away unnecessary lawyer-speak so the model can parse the instruction directly. The same writing discipline that makes a brief readable also makes a prompt effective.

Chain-of-Thought Prompting

For complex legal reasoning, one of the most useful techniques is chain-of-thought prompting.3, 4 Rather than asking the model for a final answer, you instruct it to produce a series of intermediate reasoning steps. The research literature shows that forcing the model to work through its reasoning aloud significantly improves accuracy on complex symbolic and arithmetic tasks. The practical benefit for lawyers is auditability: when the model shows its work, you can see exactly where the logic breaks down and stop relying on the answer.

Prompt engineering is becoming a required skill under the duty of technological competence. The mental model that works is treating the AI the way a senior partner treats a junior associate: give clear instructions, then edit the work that comes back.

Prompt Chaining and Iteration

The real productivity gains come from iteration and refinement. Prompt chaining lets a user tie multiple prompts together to produce a sequence of outputs, breaking a complex task into smaller sub-tasks. You might ask the tool to summarize a contract, then identify deviations from a standard template, then draft a response letter. At each step you act as the senior partner reviewing draft work: "make the tone more judicial," or "give me a list of the assumptions you made and which I should verify." The output gets sharper with each pass, and the verification load stays manageable because each step is small enough to check.

Sources

  1. White, J., et al. (Vanderbilt University). A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT. arXiv preprint. Systematic catalog of prompt engineering patterns applicable to professional and legal contexts.
  2. White, J., et al. A Prompt Pattern Catalog (supplementary pattern library).
  3. Wei, J., et al. Chain of Thought Prompting Elicits Reasoning in Large Language Models. Google Research / arXiv. Landmark paper demonstrating that step-by-step reasoning prompts significantly improve LLM accuracy on complex tasks.
  4. Wei, J., et al. Chain of Thought Prompting (supplementary experimental results).

Ready to bring responsible AI to your firm? Let's start with a conversation.

Book a Discovery Call