Generative AI takes its name from the fact that AI models like ChatGPT and Claude generate answers in response to user prompts. Although using these models may seem straightforward at first (akin to typing a string of words into a Google search without much thought), truly taking advantage of what generative AI models can offer requires some skill in crafting effective prompts. Use the following takeaways to assist in writing effective prompts for AI to ensure you get your desired answers.
Write Clear and Concise Prompts
A good prompt tells the AI model what you want. Leave out unnecessary words in the prompt, ensuring the AI model will deliver the most on-point results.
For example, if you want a summary of the felony murder rule, write: “Summarize the felony murder rule.” Doing so in ChatGPT provides a concise paragraph summary of the rule: “The felony murder rule holds that if a death occurs during the commission of a serious felony (like robbery, arson, or burglary), the defendant can be charged with murder, even if they did not intend to kill anyone. The rule applies to those involved in the felony, including accomplices, and the death must happen during or in the immediate aftermath of the felony. The key aspect of the rule is that intent to kill is not necessary for a murder charge.”
Use Context and Examples
Generative AI seems more effective when given context, much like human conversations. Providing a prompt without context or examples may eventually get the AI model to where the user wants it, but using examples with AI early makes the process smoother and more efficient.
To gauge ChatGPT’s effectiveness, if you were writing an article about AI’s usefulness for young lawyers, you might submit this prompt: “Write an article to young lawyers about how to write effective prompts for AI.” One answer the AI provided that would seem especially helpful was its instruction to “use examples” to guide the AI model to a particular type of response. ChatGPT’s response suggested that users ask the AI model, for example, to “provide a sample clause for a non-disclosure agreement that includes both duration and jurisdiction.” Doing so helps guide the model in understanding the format or specific type of response the user requires.
Tell the AI Model How to Act
Don’t hesitate to give the model a role—a prompt like: “You are a trusts and estates lawyer in New York City,” which tells the model how to act when it formulates its result. Then, after you have given it that instruction, follow up with a request such as “provide the benefits and disadvantages of including a no-contest clause in a will.” Generally speaking, a no-contest or “in terrorem” clause is a provision in a will that provides that if a beneficiary challenges the validity of the will, the beneficiary will forfeit their entire inheritance; the clauses are designed to discourage beneficiaries from bringing will challenges that delay the probate process.
Doing this gives the model a level of specificity to your request and, in this case, prompted Claude to add as a disadvantage that no-contest clauses have limited enforceability in New York.
Continue the Conversation by Building on the Request
Using AI models can resemble having a conversation with another person. MIT Sloan Teaching & Learning Technologies, part of the technology services team at MIT Sloan, recommends that users build on these conversations because “the chat-based systems are capable of remembering what happened earlier in your conversation without re-establishing context.” Giving Claude further instruction to “revise this answer so your client is the audience” resulted in mostly reworded answers in the lists of benefits and disadvantages, but one notable change was that the first benefit switched from “deterrence of challenges: the primary benefit is discouraging beneficiaries from contesting the will, as they risk losing their inheritance if they challenge and lose” to “protects your wishes: a no-contest clause helps ensure your estate is distributed exactly as you've outlined, reducing the chance that anyone will successfully override your intentions.”
Georgetown University’s Library resources suggest including words from the first answer in a second prompt. For example, asking, “Why doesn’t Georgetown have a Metro station?” may result in an answer that includes “geological obstacles.” A helpful follow-up prompt would be: “What are the engineering challenges and geological obstacles to constructing a Metro station in Georgetown?”
Correct AI’s Mistakes
AI models are still evolving, and they are far from perfect. (Just ask the two New York lawyers who were sanctioned in June 2023 for submitting a brief with six fictitious case citations generated by ChatGPT.) To improve these models, users can and should provide feedback to eliminate erroneous responses. For example, after a PC Mag writer asked ChatGPT to name all of the films of one of his favorite actors from the 1950s and 1960s, the list it returned omitted one film from 1953, Terminal Station. The writer responded by telling ChatGPT it had left Terminal Station off the list. ChatGPT responded: “You’re correct! Montgomery Clift did appear in the film Terminal Station, also known as Indiscretion of an American Wife in the US.” The AI model then provided an updated list with that inclusion. When the writer asked ChatGPT his initial question again, it included Terminal Station (Indiscretion of an American Wife) in the list.
Make the Most of AI
As AI becomes more popular in the legal field, lawyers who know how to craft useful prompts will be able to use models like ChatGPT and Claude most effectively. By writing clear prompts, using examples, instructing the AI model on how to act, building on conversations, and correcting mistakes when they occur, AI can be a successful tool in a lawyer’s practice.
