The most effective prompt pattern for legal work is context-role-constraints: give the model the relevant facts and any governing document, tell it what role to write in, and constrain the output by jurisdiction, length, tone, and the propositions it must not assert without a source. Lawyers who prompt this way get drafts that need editing; lawyers who type one-line requests get generic text that needs rewriting — and then wrongly conclude the tool is useless.
Prompting technique changes output quality more than tool choice does, and it is a skill any lawyer can learn in an afternoon. This guide covers the pattern, ready-to-adapt structures for common legal tasks, and the confidentiality limits no prompt can override. It is part of our AI in Legal Practice library.
Related: AI in Legal Practice · Can Lawyers Use ChatGPT? · AI Contract Drafting · AI Hallucinations in Research · AI & Client Confidentiality · Practice-Area AI
The context-role-constraints pattern
Three elements do most of the work in a legal prompt. Context: paste the facts, the governing document, or the source text the model should work from — a model with your material produces something grounded in it, while a model working from a one-line description invents plausible generalities. Role: tell it who is writing ("outside counsel drafting for a sophisticated commercial client," "a plaintiff-side personal injury lawyer writing a demand letter"), which sets register and assumptions. Constraints: specify jurisdiction, length, tone, the provisions that must appear, and the propositions it must not assert without a cited source.
A weak prompt: "Write a demand letter for a car accident." A strong one: "You are plaintiff's counsel. Using the facts below, draft a demand letter to the defendant's insurer for a rear-end collision in Ontario. Keep it under 400 words, firm but professional, itemize the special damages listed, and do not state a settlement figure or cite any statute — I will add those. Facts: [paste]." The second returns something you edit; the first returns something you rewrite.
The confidentiality limit no prompt overrides
Before technique, the hard rule: what you paste as context is a disclosure to the vendor. Under Rule 1.6 and FLSC Rule 3.3-1, client-identifying facts, privileged material, and case documents do not go into a consumer-tier tool that may retain or train on them — no matter how good the prompt would be. Client-related context belongs only in a business or enterprise tier with training disabled and retention configured, and sometimes only with informed client consent per ABA Op. 512. Every prompt is also a record that can be retained, produced, or breached. The full analysis is in our guide to client confidentiality and AI tools.
Template structures for common tasks
Build a small library of reusable prompt structures for the firm's ten most common tasks. A few patterns:
- Summarize a document: "Summarize the attached [transcript/contract/memo] in [N] words for [audience]. List the key points as bullets, flag anything ambiguous, and do not add facts not in the document."
- Draft a clause: "Draft a [indemnification/limitation-of-liability] clause for a [jurisdiction] [contract type], favoring the [buyer/seller]. Keep it plain-language, and mark any place where a business decision is needed with [DECISION]."
- Issue-spot: "Acting as opposing counsel, list the strongest arguments against the position in the facts below. Do not soften them."
- Reformulate for a client: "Rewrite the passage below at a grade-8 reading level for a client with no legal background. Keep it accurate; do not add advice."
Templates standardize quality across everyone on the team and remove the blank-page problem. For transactional work, pair them with the clause-level review discipline in AI contract drafting for lawyers.
Techniques that reliably improve output
A handful of moves consistently raise quality:
- Ask for reasoning before the answer on analytical tasks — "work through the elements, then conclude" — which surfaces gaps you can catch.
- Give an example of the format you want (a prior letter with the client details stripped) so the model matches your house style.
- Iterate rather than restart: "tighten paragraph 2," "make the tone firmer," "add a without-prejudice line" beats rewriting the whole prompt.
- Ask it to flag its own uncertainty: "mark anything you are inferring rather than drawing from the text" turns silent guesses into visible ones you can verify.
- Constrain assertions: "do not state any legal rule as settled without noting it needs verification" reduces confident hallucination in the draft.
What prompting cannot fix
No prompt makes a general chatbot a reliable source of legal authority. Asking it to "only cite real cases" does not prevent fabrication — the model cannot check whether a case exists, so it generates plausible ones and, if asked, confirms they are real. This is exactly what happened in Mata v. Avianca. Prompting improves drafting, summarizing, and reasoning; it does not add a database the model does not have. For anything citable, use grounded research tools and verify, as detailed in AI hallucinations in legal research.
Prompting also cannot supply judgment. The constraints you set — jurisdiction, what to leave for the lawyer, what not to assert — are themselves legal judgments. A good prompt encodes a lawyer's thinking; it does not replace it, and the output still gets reviewed before it matters.
Managing long documents and context limits
Every model has a context window — a limit on how much text it can hold at once — and legal documents routinely exceed it. When they do, quality degrades silently: the model appears to process a 200-page contract but actually attends to the parts it can fit, dropping the rest without telling you. Two practical fixes. First, chunk the work: feed the document in labeled sections, summarize or analyze each, then have the tool assemble a master output — and check the seams, where models lose track of cross-references. Second, ask targeted questions against a specific section rather than "review this whole document," which forces the model to work on text it genuinely has in view.
Newer tools with very large context windows reduce but do not remove the problem, because performance on material buried in the middle of a long input tends to fall off even when it technically fits. For anything important, verifying against the source document — not the model's summary of it — remains the discipline.
Building the firm's prompt library
Treat prompts as shared work product. Collect the structures that work into a firm document, strip any client details from the examples, and update it as lawyers find better formulations. This is the small-firm equivalent of a knowledge-management system, and it does more to raise everyone's output than any single tool purchase — the technique compounds across the whole team. For a plan that connects internal AI skill to client-facing systems, talk to LexScale.ai.
Frequently Asked Questions
Grow your AI in Legal Practice practice with AI
LexScale.ai builds AI search visibility, websites, and intake systems for ai in legal practice firms across North America. Book a free strategy call to see what would move the needle for your practice.
Book a Free Strategy Call →