The prompts that work ask for extraction, quotation and comparison. The prompts that fail ask for a legal conclusion — and get a confident one.
Key takeaways
The prompts that work in contract review share one property — they ask for something checkable. Extract this clause and quote it. List every date. Compare this wording against this standard and mark each as matching, weaker or absent. Every one of those outputs can be verified against the document in seconds, which is what makes it useful rather than merely impressive.
The prompts that fail ask for judgment: is this contract reasonable, should my client sign, what is market for this clause. Those produce fluent answers that cannot be checked without redoing the work, and where the model has no reliable basis it will still answer.
The five prompts below are written for a document you have supplied to the tool. Adapt the bracketed parts. They assume you have already settled whether that document may be sent to that tool at all — a strict confidence question under rule 3.3-1 of the Model Code of Professional Conduct, not a preference.
Clause inventory
“From the attached agreement, produce a table with one row per clause in this list: term and renewal, termination for convenience, termination for cause, assignment, change of control, limitation of liability, indemnity, insurance, confidentiality, governing law, dispute resolution. For each row give the clause number, the heading as written, and the operative text quoted verbatim. If a clause in the list does not appear in the agreement, write NOT FOUND. Do not summarise, paraphrase or comment.”
This is the workhorse. It converts an 80-page document into a page you can read, and every cell is verifiable. The NOT FOUND instruction matters more than the rest: absence is usually the finding that costs money, and without an explicit instruction a model will often supply a plausible clause that is not in the document.
Playbook comparison
“Here is our standard position for each clause: [paste your positions]. For each clause in the attached agreement, quote the actual wording, then classify it as AT STANDARD, WEAKER THAN STANDARD, STRONGER THAN STANDARD or ABSENT, and state in one sentence which words drive that classification. Do not recommend a course of action.”
This is the prompt that turns a general tool into your firm’s tool. It requires you to have written the positions down, which is the real work. The instruction not to recommend is deliberate: the classification is checkable, the recommendation is advice, and advice is the lawyer’s.
Obligation schedule
“List every obligation in the attached agreement that is tied to a date, a notice period or a triggering event. For each, give the clause number, the party who owes the obligation, the trigger, the deadline as written, and the consequence of failure if the agreement states one. Quote the deadline language exactly. Do not calculate dates.”
“Do not calculate dates” is there because date arithmetic against business days, holidays and receipt rules is exactly where quiet errors appear. Extract the language; let a person calculate against the applicable rules. Where an early exit is the question, the underlying law is its own topic — see Treadstone Law on ending a fixed-term contract early in Ontario.
Consistency check
“Check the attached agreement for internal inconsistencies only. Report: defined terms that are used but never defined; defined terms that are defined but never used; cross-references that point to a clause number that does not exist or does not match the subject matter; numbering breaks; and any two provisions that contradict each other. Quote both provisions in the last case. Report nothing else.”
This is the prompt with the highest hit rate on documents that have been amended more than once, and it is close to risk-free because every finding is a fact about the document rather than a view about the law.
Client summary
“Using only the clauses quoted in the table above, write a plain-language summary for a business reader of what this agreement requires our client to do, what it allows the other party to do, and how it ends. Maximum 400 words. Do not add any obligation, right or risk that is not in the quoted text. Do not give an opinion.”
Note the constraint to material already extracted and verified. Summarising from the verified table rather than from the raw document is what keeps invention out of the client-facing document, and the summary still goes out over a lawyer’s name after a lawyer has read it.
Three instructions to put in every prompt
“Find case law supporting our position on this clause.” A generative tool asked for authorities can return citations that do not exist, in a format that looks exactly like ones that do. The Office of the Privacy Commissioner of Canada’s generative AI principles make the general point in plain terms: organisations should evaluate the validity and reliability of a tool for the intended purpose, and providers should disclose known or reasonably expected failure cases, including contexts in which a system may produce incorrect information. If you ask for authorities, verify every one against the reported decision before it goes anywhere, and check the notices of the forum you are filing in.
“Is this indemnity market standard?” There is no reliable corpus behind that answer, and the response will sound authoritative anyway. Ask instead for the indemnity to be quoted and compared against your own recorded positions — which is Prompt 2. Whether an indemnity cap should track the whole purchase price or part of it is a negotiation question with real consequences, as Treadstone Law’s note on indemnity caps shows.
“Draft the clause we should propose.” This one is fine, with a caveat: draft from your precedent, supplied in the prompt, rather than from the model’s general knowledge. A clause invented from nothing reads well and carries none of your firm’s accumulated positions.
Do longer prompts work better?
More specific ones do. The gains come from naming the output format, naming the clauses, and stating what to do when something is missing — not from length.
Can we save these as templates?
Yes, and you should. A shared prompt library with the firm’s playbook baked in is the difference between consistent output and every lawyer improvising. Version it like a precedent.
Does any of this remove the review step?
No. It makes the review step fast, because quoted text with clause numbers can be checked against the source in seconds. That is the whole design goal.
A 30-minute call is enough to tell you whether AI pays for itself here.
