Prompt Engineering for Legal Teams: A Practical Guide

Quick-Start (Copy This Right Now)

Can AI actually help with legal work, or is it too risky to touch anything that might end up in a contract? That's the question every in-house counsel asks the first time someone on their team suggests using AI for legal drafting. The honest answer: it helps a lot with drafting speed and pattern-spotting, and it should never be the final word on anything binding. Here's a starting template for prompt engineering for legal teams that keeps that line clear:

You are assisting with a FIRST DRAFT only. This is not legal advice and requires review by a licensed attorney before use.

Document type: [e.g. "mutual NDA between two software companies"]
Jurisdiction: [STATE/COUNTRY]
Key terms to include: [LIST SPECIFIC TERMS, e.g. "2-year term, mutual confidentiality, standard carve-outs for public information"]

Draft this section: [SPECIFIC CLAUSE OR SECTION]

Flag any term that varies significantly by jurisdiction with [JURISDICTION-DEPENDENT].
Do not invent case law, statutes, or citations. If you're uncertain about a legal standard, say so explicitly.

What this does: it sets the boundary immediately — draft, not advice — and the "do not invent citations" line matters a lot because AI models can sometimes produce citations that look completely legitimate but reference cases that simply don't exist.

⚠️ Common mistake: Treating AI-drafted clauses as final language. Every output here needs attorney review before it touches an actual contract, no exceptions.

Understanding the Variables

Jurisdiction is the variable that trips up the most legal teams using AI. A contracts manager at a mid-size manufacturing company learned this the hard way after an AI draft used a non-compete clause structure that's enforceable in one state and essentially void in another. The model wasn't "wrong" exactly — it produced a common clause pattern. It just didn't know which state's rules actually applied because she hadn't told it.

Key terms matter just as much. Vague instructions like "make it protect us" produce generic risk-averse language that might not match your actual negotiating position. Specific terms — "we need to retain IP rights on any derivative work created during the engagement" — give the model something concrete to draft toward, rather than a defensive posture that might actually work against you in the negotiation.

⚡ Pro tip: Always specify which party's interests the draft should favor. "Draft from the vendor's perspective" and "draft from the client's perspective" produce meaningfully different default terms, even for the same clause.

⚡ Pro tip: Keep a short glossary of your jurisdiction's specific quirks — terms or clause structures that behave differently there than in the "generic" version the model defaults to — and paste it into prompts for that jurisdiction. It saves you from re-explaining the same local exception every single time.

Step-by-Step: Drafting a Contract Clause

Start by defining the document type and jurisdiction before anything else — this is the context the model needs to avoid defaulting to generic boilerplate.

Document type: Consulting agreement
Jurisdiction: Delaware
Party perspective: Drafting for the consultant (not the hiring company)

Next, name the specific clause you need, along with any non-negotiable terms:

Draft a payment terms clause. Non-negotiables: net-15 payment, late fee of 1.5% monthly on overdue balances, right to pause work if payment is more than 30 days late.

What this does: separates the "what" (payment terms clause) from the "how" (specific numbers and rights), which keeps the model from guessing at terms you already know you want.

Then, ask for a plain-language summary alongside the legal draft:

After the clause, add a 2-sentence plain-language summary of what this means in practice for the consultant.

A solo practitioner handling contract review for small business clients said this step alone saves her a meaningful chunk of client-facing time each week, since she can forward the summary directly instead of translating legalese in every single follow-up email.

⚡ Pro tip: Ask the model to list the 2-3 most negotiable elements of the clause it drafted. This gives you a head start on what to expect the other party to push back on.

⚡ Pro tip: For any clause involving liability, indemnification, or IP, explicitly request the model flag anything unusually aggressive or unusually weak compared to market-standard terms — it won't always catch this, but it's a useful second check.

Pro-Level Variations

For contract review rather than drafting, flip the prompt: paste in the existing clause and ask the model to identify risks specific to your role in the deal. A paralegal at a busy commercial real estate firm uses this exact structure for lease review every week:

Here is a lease clause: [PASTE CLAUSE]
We represent the tenant. Identify any terms that are unusually unfavorable to the tenant compared to standard commercial lease terms, and explain why in one sentence each.

For summarizing long documents before a meeting, ask for a structured breakdown rather than a narrative summary — obligations, deadlines, and termination rights as separate lists, since that's usually the format attorneys actually scan for first when they've got five minutes before walking into the room.

⚠️ Common mistake: Pasting in confidential client documents into a general consumer AI tool without checking your firm's data handling policy first. This is a real risk, not a theoretical one, and it's worth confirming before you paste anything client-specific — many firms now require an enterprise-tier tool with contractual data protections before any client material touches an AI system at all.

Troubleshooting Common Issues

If the draft feels generic no matter how you phrase the request, the issue is almost always missing specificity in your key terms. Go back and add actual numbers, actual party names (even placeholder ones), and the actual jurisdiction — vague inputs produce vague clauses regardless of how the prompt is structured. A useful test: if you could hand your prompt to a first-year associate and they'd have to ask you three clarifying questions before drafting anything, the model will hit the same wall, just silently.

Another common issue is length creep — asking for "a clause about payment terms" without a length constraint can produce either a two-sentence fragment or an eight-paragraph overcorrection depending on how the model interprets the request. Specify a rough length, like "3-4 sentences" or "one paragraph," so you're not stuck trimming or padding after the fact.

If you're getting citations or references to specific statutes and you're not confident they're accurate, don't trust them without verification. This is one of the few areas where AI legal drafting genuinely fails in a way that matters — treat any cited case, statute, or regulation as a lead to verify, never as a fact to cite directly.

If the tone feels off for your document type — too casual for a formal filing, too formal for an internal policy memo — add an explicit tone instruction and, if possible, paste in one paragraph from a document you know is written in your firm's preferred style.

There's also a category of failure that's less about the AI and more about workflow: teams that get a solid first draft and then skip the review step entirely because the output "looked professional." Professional-sounding is not the same as legally sound, and a general counsel at a logistics company put it well when she said the risk isn't that AI drafts bad clauses — it's that good-looking drafts make people more likely to skip the review they'd never skip for something that looked obviously rough.

Your Turn

Take one recurring clause your team drafts often — an NDA confidentiality section, a standard indemnification clause, whatever comes up weekly — and build a template using the jurisdiction, party perspective, and key terms structure above. Run it, then have an attorney mark up what needs fixing. Note specifically what the attorney changed, not just that they changed it — was it a jurisdiction issue, a missing term, an overly aggressive default? That pattern tells you what to add to the prompt next time.

Do this for three or four of your most common document types, and you'll start to see the same categories of correction come up repeatedly. That's useful information: it means your prompt template has a specific, fixable gap rather than being generally unreliable. A contracts team at a SaaS company found that almost all of their corrections fell into one bucket — missing jurisdiction-specific carve-outs — and fixing that one gap in their template cut their attorney markup time by close to a third.

It's also worth building a short internal guide for anyone on your team using AI for legal drafting, even informally. Two or three rules — always specify jurisdiction, never treat citations as verified, always route through attorney review before anything is sent externally — go a long way toward preventing the kind of mistake that's expensive to fix after the fact rather than before it.

Save the corrected version as your working template, not the raw AI output, so next time you start from something an attorney has already reviewed rather than a blank prompt. A tool like PromptABCD is useful here for keeping that library of attorney-approved templates organized and versioned across your team, so the next associate who needs an NDA draft isn't starting from scratch either.