TL;DR: Reliable legal AI output comes from instructions with five elements: context (the matter and parties in two sentences), jurisdiction and scope (the element lawyers skip most), task (one precise verb — review, extract, compare, draft), format (memo, grid, clause list, tracked changes), and constraints (what to rely on, what to flag, what not to do). Brief the AI the way you would brief a capable, literal junior on their first day — assume no shared context. Iterate by narrowing rather than restarting, and remember what prompting cannot fix: no instruction makes an ungrounded tool verifiable. The endgame is standardisation — the best prompts become team templates.
The best prompting advice for lawyers is hiding in plain sight: you already know how to do this. Every partner who has briefed a junior associate — precisely, with context, scope, and the shape of the deliverable — has the entire skill. What trips lawyers up is mis-calibrating the audience: the AI is simultaneously more capable than a first-year (it has read everything) and more literal (it knows nothing about your matter, your client, or your intent beyond what you type). This guide translates instruction-giving into that register: five elements, worked examples, iteration patterns, and the honest limits.
The mental model: brief it like a junior lawyer
Picture delegating to a brilliant, tireless junior who started this morning: encyclopaedic knowledge, zero context, takes everything literally, and never pushes back on a vague instruction — just produces something plausible that matches its best guess at what you meant. Every prompting failure maps onto that picture. "Summarise this agreement" produces a generic summary because you did not say for what purpose. "Is this clause standard?" produces hedge-prose because you did not say against what benchmark, for which party, in which market. The fix is never cleverness; it is the same specificity you would owe the human junior — plus one adjustment: the AI does not accumulate context between matters unless you provide it, so the briefing discipline applies every time.
The five elements of a reliable instruction
| Element | What it supplies | Example fragment |
|---|---|---|
| 1. Context | The matter, the parties, your side | "We act for the tenant, a retail chain, negotiating a 10-year anchor lease." |
| 2. Jurisdiction & scope | Which law, what time frame, what posture | "Under English law, current as of mid-2026, pre-signature review." |
| 3. Task | One precise verb with an object | "Identify every clause that shifts repair obligations to the tenant." |
| 4. Format | The shape of the deliverable | "A table: clause reference, quoted text, risk level, suggested redline." |
| 5. Constraints | Boundaries and honesty rules | "Quote the document verbatim; flag anything ambiguous rather than guessing; do not advise on tax." |
Not every task needs all five spelled out — a quick clause explanation needs less than a research memo — but when output disappoints, the diagnosis is almost always a missing element, and the fastest fix is to supply it rather than rephrase what was already there.
Jurisdiction and scope: the element lawyers skip most
Law is jurisdictional; a question without a jurisdiction invites an answer blended from several systems — fluent, plausible, and attached to nowhere. Always state the governing system ("under Delaware law", "under the Indian Contract Act"), the temporal frame ("as of mid-2026", "at the time of signing in 2023"), and the posture ("advising pre-signature" vs "assessing a live dispute"), because each changes what a correct answer looks like. Well-designed platforms backstop this — Judicio detects the likely jurisdiction and confirms it before answering, and asks a clarifying question when scope is ambiguous rather than guessing — but the backstop is for the cases you miss, not a substitute for stating it. Cross-border questions deserve their own discipline; see our guide to multi-jurisdiction research.
Before and after: three worked examples
Research. Before: "What's the law on non-solicitation clauses?" After: "We act for an employer in Singapore. A departing account manager signed a 12-month non-solicitation covering clients she serviced. As of mid-2026, how have Singapore courts treated enforceability of client non-solicits of this duration? Cite the leading authorities with pinpoint references; note any divergence from the position on non-competes; flag if the law is unsettled." The after-version scopes system, facts, task, and honesty constraints — and produces an answer you can verify authority by authority.
Review. Before: "Check this MSA for problems." After: "Review this MSA batch as counsel for the customer. Flag: uncapped liability, indemnity broader than mutual, auto-renewal longer than 12 months, unilateral price changes, and data-use rights beyond service delivery. For each finding: quote the clause, cite page, rate severity, propose a redline we could send. List anything material outside these criteria separately." Structured criteria turn review from vibes into a checklist — which is why platforms let you save exactly this as a reusable template.
Drafting. Before: "Draft a termination clause." After: "Draft a termination-for-convenience clause for the customer side of a SaaS agreement, Indian law, 30 days' notice, pro-rata refund of prepaid fees, survival of confidentiality and accrued rights. Match the defined terms in the attached draft; deliver as a tracked-change insertion after clause 11." Constraints plus format produce a draft that drops into the document instead of a paragraph that needs translating.
Iteration patterns: refining without starting over
First outputs are first drafts of the instruction as much as the answer. The productive moves are narrowing, not restarting: tighten one element ("same analysis, but only clauses 8-14", "reformat as a two-column table"); escalate specificity where the output hedged ("assume the employee is senior management — does the answer change?"); ask for the counter-position ("now argue the licensor's best response to each finding") — a genuinely underused pattern that surfaces weaknesses before opposing counsel does; and interrogate provenance ("which source supports the second point? quote it"). On grounded platforms that last move is a click rather than a prompt — every claim already links to its page and passage — which is the workflow difference between asking a tool to be trustworthy and using one built to show its work.
What prompting cannot fix
Honest limits, so the skill is not oversold. Prompting cannot make an ungrounded tool verifiable — no instruction forces a consumer chatbot to retrieve real sources; that is architecture, chosen when you choose the tool (see why generic chatbots fall short). It cannot replace verification — a beautifully briefed answer still gets its citations opened and checked before reliance, as our verification guide details. It cannot supply confidentiality — what you may paste into which tool is a data-governance question settled by firm policy, not phrasing. And it cannot turn output into legal advice — the judgment that turns research into counsel remains the lawyer's. Prompting optimises within those boundaries; it does not move them.
From personal skill to team standard
The compounding gains arrive when good instructions stop being personal craft and become team infrastructure. The mechanism is templates: the review criteria that caught everything last quarter become the firm's standard checklist; the research instruction that produced the clean memo becomes a playbook; the twenty questions that abstract a lease become a shared matrix set; the drafting brief that matched house style becomes an outline. New joiners inherit the firm's accumulated instruction-quality on day one, and improvements propagate by editing the template rather than re-teaching everyone. This is also the honest answer to "should we hire a prompt engineer?" — no; encode the senior lawyers' judgment into shared templates instead.
How Judicio reduces the prompting burden
Judicio is designed so that good instructions are the default rather than a craft skill. The five elements are built into the interfaces: Document Review asks for criteria explicitly — write your own, start from an expert checklist, or let Judicio suggest them from the contract; the Review Matrix turns a plain-English brief or even a column header like "Governing Law" into tuned questions with typed answers; Research detects and confirms jurisdiction and asks clarifying questions instead of guessing; Drafting starts from a brief, a precedent, or one of 100 outlines and delivers tracked changes. And the standardisation layer ships with the platform: 500+ expert-built templates across review, research, drafting, timeline, and matrix in Templates & Workflows — use them as-is, build your own three ways, and share them across the organisation. Every output cites page and passage, so the "interrogate provenance" step is a click. Test your best instruction against it on a 7-day free trial — 500 credits, no credit card.
Examples are illustrative and jurisdiction-neutral where possible. Judicio outputs are for research and informational purposes and are not legal advice — verify before relying.