10 Guidelines for Generative AI Prompting in Contract Drafting: Practice Points

Start small and stop worrying about breaking anything. Sterling's first piece of advice was that you cannot break it, so have fun with it. You do not need a twelve page prompt to get value, and the fastest way to get comfortable was finding small useful tasks first and building toward the complicated ones. Plenty of good prompts already exist in guides and blogs, and running someone else's prompt still returns something interesting. The people who are behind are not impossibly behind, and that train was leaving the station.

Be specific about what you want, how you want it formatted, and what tone it should carry. Asking a tool to explain GDPR returns a very good explanation of GDPR, most of which you did not want. Asking it to explain the differences between a data processor and a data controller returns the thing you actually needed. Tell it the shape too, because absent instruction you get a memo-ish pile of bullets, and asking for each provision set out in a table by numbered section gets you a table. Tone works the same way, so ask for persuasive when you are pitching the CFO and friendly when you are explaining your red lines.

Give the tool a persona and name who will read the answer. Telling it that it is an experienced commercial agreements lawyer practicing in the United States pushes it toward the material it learned from lawyer work product rather than the general pile. Naming the audience does the other half of the job. The same GDPR explanation written for an average business executive reads nothing like the one written for an experienced lawyer, and toggling between the two shows you the difference immediately. Translating complex legal answers for the business is one of the most valuable things we do, and this takes the first pass off your plate.

Feed it the context you would give a new associate. Sterling's whole frame was that prompting works like handing an assignment to a summer associate, which means briefing it, giving background, and clarifying as you go. His favorite version of this was the negotiation prompt, where you tell it you are the customer, hand it the vendor's contract, and ask for a table of each section, why it helps or hurts you, how you would change the wording, and how you would justify the edit to the other side's lawyers. Then ask what arguments the vendor will make back and how you would answer them. You will already know most of what comes back, and the one or two things you had not thought of are the point.

Break complicated work into ordered steps. A prompt works as much like a checklist as a recipe, and that matters most when you are loading several agreements and asking for something built out of all of them. Telling it to create a playbook produces something slow and shapeless. Thinking through the steps you would take yourself and writing the prompt in that order produces something you can use. Sequence the output too, so the introduction comes first, then the benefits, then the challenges, then the savings, then the call to action.

Keep iterating and stack new work onto the same conversation. Sterling called the follow-up the part he found most powerful. You can ask it to expand the rights granted section, then ask it to redraft the whole contract in a manner most favorable to the vendor, and watch the results move. The iteration does not have to stay on the document either. Summarize Section 2 of the Sherman Act, then ask for ten examples of manufacturers running afoul of it, then ask it to turn the whole thing into an email a senior executive without a legal background can read.

Point it at the pre-signature work nobody enjoys. Sterling loaded a full agreement and asked for an execution readiness review, and it came back with a chart of signature blockers, the required fix, and why each one mattered. It caught a missing Schedule A, a blank annual fee, a capitalized Customer Data term that was never defined, and empty names in the signature block. He ran the same idea against a playbook, asking whether a limitation of liability provision complied, deviated, or was unclear, with the quoted language and a proposed drafting fix for each gap. The instruction that made it work was telling it not to treat silence in the playbook as approval.

Make it show you what it left out. The output looks confident and polished, which is exactly what makes it dangerous. Sterling built a standing follow-up prompt that asks what uncertainty sits in the analysis, what assumptions went undisclosed, and where the information is incomplete. Sometimes it comes back with a list of assumptions that tells you the answer was wrong from the start. Laura ran a version of the same idea, grading output against her requirements and sending it back until every grade came back an A.

Keep a human in the middle of every output. AI was the assistant in Sterling's framing and never the decision maker. The people who got in trouble with it skipped the basics they learned in law school, which were cite checking and reviewing and confirming that what they were looking at was correct. You would never file a first-year associate's draft without reading it or hand it to the board untouched, and the polish on AI output creates a false sense of security that invites exactly that. Ask for cites you can click through and check them.

Protect privilege and confidential information before you paste anything in. Putting a privileged document into what amounts to a public database generally waives the privilege, and our confidentiality obligations do not soften because the tool is convenient. An enterprise license or a tool that only draws on data you provided changes that analysis, since nothing leaves the company. A large share of useful prompts need no confidential information at all, so the constraint is narrower than it feels. Sterling also advised opening prompts by stating that you are an in-house lawyer preparing legal analysis for the business, because prompts are potentially discoverable and labeling them helps.