The prompts worth using force the model to work only from documents you supplied and to show where each statement came from. The ones that cause trouble ask a model what the law is.
Key takeaways
A dangerous prompt asks the model what the law is. A useful prompt gives the model a document and asks it to do something structured with that document. The second kind can be checked in minutes against the source; the first kind cannot be checked at all without redoing the work.
Every prompt below follows the same five-part shape: the role, the source material, the task, the output format, and an instruction about uncertainty. That last part matters most. A model told to flag what it cannot find will flag it; a model told nothing will fill the gap with something plausible.
"You are assisting a licensed Ontario paralegal. Using only the attached documents, produce a table with three columns: date, event, source document and page. Include only events supported by the documents. Where a date is ambiguous or missing, write UNCLEAR and say why. Do not infer events that are not recorded."
This is the highest-value prompt in a paralegal practice because verifying a chronology is fast and building one is slow. The source-and-page column is what makes verification possible.
"Here is my file list for a residential tenancy matter and the client’s account of events. List the documents referred to in the account that are not in the file list, and for each, say which part of the account refers to it. Do not suggest documents that are not mentioned."
It turns a vague "we need more from the client" into a specific request you can send the same day.
"Using the attached chronology, which I have verified, draft the facts section of a plaintiff’s claim in numbered short paragraphs, one fact per paragraph, in chronological order, in plain language, using only the facts in the chronology. Do not state legal conclusions, do not name causes of action, and do not add amounts that are not in the chronology."
Note what the prompt withholds: the model is not asked to pick the cause of action. That is the paralegal’s call and it is bounded by the permissible scope rule in Rule 3.02.
"From the attached pleadings and my verified chronology, produce: (a) the three factual points the parties agree on, (b) the three they dispute, with the document reference for each side, and (c) ten questions I should be ready to answer. Mark anything you cannot support with a document reference."
For the process this is preparing for, our sister firm sets out what happens at a Small Claims settlement conference.
"Here is my time and disbursement record for this matter. Group it by task type, total each group, and produce a plain table with a one-line description per entry. Do not calculate any entitlement, apply any tariff, or state what should be awarded."
The arithmetic is delegable; the entitlement is not. What can be claimed, and on what basis, is covered in claiming legal costs in Ontario Small Claims Court.
"Write a 150-word email to a client at a Grade 8 reading level explaining what a default judgment is and what happens next in their matter, based only on the attached summary of their file. Neutral tone, no promises about outcome, no legal advice beyond the process description."
If the matter reaches enforcement, the process itself is set out in how to enforce a Small Claims judgment in Ontario.
"Read my draft submissions and the attached chronology. List every statement in the draft that is not supported by the chronology. Do not rewrite anything. Do not add arguments."
Used this way the model is a checker rather than an author, which is the role it is best suited to and the one that fits most comfortably inside a paralegal’s obligation to assume complete professional responsibility for all business entrusted to them.
"What is the limitation period for this?" — a legal question, answered from memory, unverifiable. "What is the Small Claims limit?" — a model trained before October 2025 will say $35,000, when Ontario’s own guidance now sets a $50,000 limit effective 1 October 2025. "Find me cases that say X" — the failure mode here is invented citations, and the accountability for a filed document sits with the person who filed it. "Should my client accept this offer?" — advice, which is licensed activity.
Whatever the prompt, the same three checks apply before anything leaves the office. Every factual statement traces to a document in the file. Every date, amount and name is confirmed against the source. Nothing in the output states a legal conclusion the paralegal has not independently reached. Firms that write those three lines into a checklist have far fewer problems than firms that hold them in their heads.
Should I keep the prompts I use?
Yes. A short prompt library, stored with the firm’s precedents, is what makes output consistent between staff and reviewable by the responsible paralegal.
Can I paste a whole client file into a chatbot?
Only into a tool whose data handling you have checked, because the confidentiality duty in Rule 3.03 applies to the whole file and continues after the retainer ends. The Privacy Commissioner’s AI guidance sets out the questions to ask a vendor.
Do longer prompts work better?
Specific ones do. The gains come from naming the source, the format and what to do when the answer is not in the source — not from length.
A 30-minute call is enough to tell you whether AI pays for itself here.
