Prompt engineering for lawyers

Type “summarise this contract” into ChatGPT and you’ll get something back in seconds. It’ll read fluently. It’ll also be close to useless: a generic recap that any layperson could have written, missing the two clauses that actually carry the risk, silent on the jurisdiction, and blind to which side you act for. Same tool, same contract, thirty seconds later. The difference between that and a genuinely useful answer isn’t the model. It’s the prompt.

Here’s the engineered version of the same request. “Act as a commercial lawyer reviewing this software-as-a-service agreement under Indian law, on behalf of the customer. List the five clauses that carry the most risk to my client. For each, quote the exact clause text, explain the risk in one sentence, and flag whether it’s market-standard or unusual. If the agreement is silent on indemnity caps or data protection, say so explicitly rather than guessing.” Feed that the same document and you get a working first pass a junior would take an afternoon to produce.

That gap is what prompt engineering for lawyers is really about. Not clever tricks or secret keywords. Just the discipline of telling the tool who it is, what law applies, what facts it’s working with, and exactly what you want back. Get that right and generative AI behaves like a fast, tireless associate. Get it wrong and it behaves like a confident intern who never admits they’re unsure.

And in India, the stakes for “confidently wrong” have moved from embarrassing to disciplinary. In July 2026, the Supreme Court held in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 SCC OnLine SC 1258) that a decision built on AI-hallucinated material is “no decision in the eyes of law,” and directed the Bar Council of India to frame disciplinary consequences for advocates who file fabricated citations. So a better prompt isn’t just about quality. It’s part of how you stay out of trouble.

Before we go deeper, here’s a template you can copy straight into any AI tool and adapt. Keep it somewhere handy.

Role: Act as a [practice-area] lawyer practising under [jurisdiction] law, advising the [party you act for]. Context: Here are the facts and the document: [paste text or facts]. Task: [the exact thing you want, e.g. “identify the top five risks to my client”]. Output format: [e.g. “a table with clause reference, risk, and market-standard vs unusual”]. Constraints: Cite the exact clause or provision for every point. If the document is silent on something, write “not stated” instead of guessing. Do not invent case law or section numbers.

Why generic prompts fail on legal work

Legal work is precision work, and a vague prompt is precision’s enemy. When you ask a general question, the model answers for a general audience: an average contract, an unnamed country’s law, no particular client interest. That average is exactly what you don’t need. Your client isn’t average, the governing law isn’t optional, and the side you act for changes which clauses count as risks in the first place.

There’s a second failure mode that’s sharper for lawyers than for anyone else. Generative models hallucinate, which means they’ll produce a citation or a section number that looks perfect and refers to nothing real. Ask loosely and you invite it. The Thomson Reuters 2025 Generative AI in Professional Services Report found legal document review and research are now the top AI use cases, at 77% and 74% of firms respectively. Those are precisely the tasks where an invented authority does the most damage.

So what actually fixes it? Structure. A prompt that names the role, the jurisdiction, the facts, the output shape, and the constraints removes the room the model uses to drift. This is the single highest-impact habit in the whole discipline, and it’s the subject of the rest of this guide. Everything else is refinement on top of getting the structure right.

Want to turn AI from a novelty into a reliable part of your legal workflow? SkillArbitrage’s Generative AI & Prompt Engineering course teaches the safety-first prompting patterns this guide is built on, designed for working professionals with no coding required. Explore the Generative AI & Prompt Engineering programme to see how it maps to real legal work.

Anatomy of a legal prompt

Strip away the jargon and a strong legal prompt has five parts. Learn these once and you’ll never stare at a blank chat box again. The parts are role, jurisdiction, context, output format, and constraints, and each one closes a gap the model would otherwise fill with a guess.

Role tells the model whose shoes to stand in. “Act as a commercial contracts lawyer” produces a different answer from “act as a litigator” or “explain this to a first-year student.” Jurisdiction is the one lawyers skip most and regret most, because a model trained largely on US and UK material will happily apply the wrong country’s doctrine unless you pin it to Indian law (or whichever law actually governs). Context is the raw material: the clause, the facts, the prior correspondence, the deal background. The more relevant context you give, the less the model invents.

Output format is where a lot of wasted time hides. Ask for “a table with columns for clause, risk, and recommendation” and you get something you can drop into a note to the client. Leave it open and you get a wall of prose you then have to reorganise yourself. Constraints are the guardrails: cite the exact clause, flag anything unusual, and say “not stated” when the document is silent instead of filling the gap. That last instruction alone eliminates a surprising share of hallucinations.

Does every prompt need all five, every time? No. A quick plain-English explainer needs role and context and little else. But for anything that touches a real matter, running through the five parts takes ten seconds and saves you an hour of correcting a lazy answer. Think of it as the legal drafting instinct you already have, pointed at the machine.

Core prompting techniques for legal work

Once the anatomy is second nature, three techniques do most of the heavy lifting on harder tasks. None of them require any technical skill. They’re just structured ways of asking.

The first is few-shot prompting, which is a fancy name for “show, don’t just tell.” Instead of describing the clause you want, paste one or two examples of clauses you’ve drafted before and ask the model to draft a new one in the same style, tone, and risk allocation. Give it two of your own indemnity clauses and it’ll match your house style far more closely than any adjective could. Lawyers already work from precedents. This is the same instinct, handed to the tool.

The second is chain-of-thought prompting for issue-spotting. Ask the model to “work through this step by step: first list every obligation the customer takes on, then flag which ones lack a corresponding remedy, then rank the gaps by risk.” Forcing the reasoning into visible steps tends to surface issues that a one-shot “what are the risks?” glosses over. It also makes the output easier to check, because you can see where the logic went if it went wrong.

The third is persona prompting, the close cousin of role. “Act as a senior Indian contract lawyer with fifteen years in cross-border SaaS deals, known for spotting one-sided limitation-of-liability clauses” primes the model toward the exact lens you want. Is the persona literally true? It doesn’t matter. What matters is that a specific persona pulls a sharper, more specialised answer than a generic one. Which of these three you reach for depends on the task, and part of the skill is knowing that.

This is exactly the kind of structured prompting covered in depth in SkillArbitrage’s Generative AI & Prompt Engineering course. You’ll learn how to build few-shot examples from your own precedents, how to design chain-of-thought prompts for review work, and how to package these skills into services you can sell to global clients. See the full course curriculum.

10 ready-to-use legal prompts

Here’s the working toolkit, and this is the part worth bookmarking. Each of these is a full, copy-paste prompt built the way the anatomy section describes, not a one-line request you’ll have to babysit. Fill in the bracketed placeholders, paste in your document or facts, and always run the verification step from the next section before anything reaches a client or a court. They’re deliberately long, and that’s the point: in legal work, the instruction you leave out is the loophole the model walks through.

One habit before you start. Where a prompt tells the model to flag uncertainty or mark something “[verify],” keep that line in. It’s tempting to trim it for a tidier answer, but that single instruction is what turns silent invention into a visible flag you can catch.

1. Contract review. This is where a lazy prompt does the most damage, because the risks that matter most are usually buried in cross-references, defined terms, and the interaction between two clauses that each look harmless alone. The prompt below forces a clause-by-clause pass and closes the gaps a quick “what are the risks?” leaves wide open.

You are a senior commercial contracts lawyer qualified in [jurisdiction, e.g. India], reviewing the attached [contract type, e.g. SaaS master services agreement] on behalf of the [party you act for, e.g. the customer]. Apply only [jurisdiction] law. Assume nothing about the governing law: if any clause selects a different governing law or forum, make that your first observation, because it changes how every other clause is read.

Work through the agreement clause by clause, in the order it appears. For each clause that creates a material risk, obligation, or cost for my client, quote the exact clause text and its number as it appears in the document. Do not paraphrase a clause when you quote it, and never cite a clause number that is not actually in the text. When you refer to a defined term, use the definition given in the contract, not the general meaning of the word, and tell me if a term the contract relies on is used but never actually defined.

Pay particular attention to the clauses that carry hidden risk even when they look standard: limitation of liability and its carve-outs, indemnity scope and how it interacts with the liability cap, termination (including termination for convenience, notice periods, and auto-renewal), governing law and dispute resolution, assignment and change of control, confidentiality, intellectual property ownership, data protection, warranties, and force majeure. Where two clauses interact, for example an indemnity that is expressly excluded from the liability cap, explain the combined effect, not just each clause in isolation.

For every risk, state whether the position is market-standard, borderline, or unusual for this type of agreement, and give a one-sentence reason. Where the contract is silent on something a [party]-side lawyer would expect to see, for example a liability cap, a data-processing addendum, or a service-credit regime, write “the agreement is silent on this” rather than assuming a position. Do not invent industry norms you are not sure of; if you’re unsure whether something is standard, say so plainly instead of guessing.

Present your output as a table with four columns: clause reference, what it says, the risk to my client, and a suggested redline or fallback position. After the table, list separately any annexures, schedules, or exhibits the contract refers to but that were not provided to you, because their absence may hide further obligations. End with the three items I should negotiate first and why.

2. Clause drafting. The danger in AI drafting isn’t a clumsy clause, it’s a confident one that’s unenforceable in your jurisdiction or quietly contradicts the rest of the contract. This prompt makes the model draft to your house style, stay inside the law, and own up to what it’s unsure of.

You are a [jurisdiction] commercial drafting lawyer. Draft a [clause type, e.g. limitation-of-liability] clause for a [contract type] governed by [jurisdiction] law, drafted to favour the [party], but only as far as is actually enforceable in that jurisdiction. If [jurisdiction] law limits how far this clause can go, for example restrictions on excluding liability for fraud, gross negligence, or death and personal injury, or the treatment of penalty versus liquidated-damages clauses, draft to the edge of what will hold up and flag the limit rather than drafting something a court would strike down.

Match my house style. Here are one or two clauses I have drafted before: [paste examples]. Mirror their structure, defined-term conventions, numbering, and level of formality. If my examples use a defined term such as “Losses” or “Liability Cap,” reuse the same defined terms and do not silently introduce new undefined ones. If your draft needs a term my examples don’t define, define it explicitly and tell me you’ve added it.

Draft the clause in full, including every sub-clause needed for it to actually operate: the cap amount or formula, the categories of loss excluded such as indirect or consequential loss, any carve-outs from the cap, and how it interacts with any indemnity. Where a figure or period is needed, insert a clearly marked placeholder like [CAP AMOUNT] or [NOTICE PERIOD] rather than inventing a number.

Do not cite any statute, section number, or case unless you are certain it exists and says what you claim. If a statutory reference would strengthen the clause, flag it as “[verify: possible reference to …]” so I can confirm it rather than relying on your memory. Inventing a plausible-looking section number is worse than leaving a gap.

After the clause, list in plain English every assumption you made, every choice that favours my client, and every point the other side is likely to resist, with a fallback position for each. Then check your own draft for internal consistency: confirm it doesn’t contradict the indemnity, warranty, or termination clauses a [contract type] would normally contain, and name any clause I should review alongside this one to avoid a conflict.

3. Case summary. The failure mode here is the model quietly adding a case, a statute, or a holding that isn’t in the judgment you handed it, and blurring what’s binding with what’s just commentary. This prompt boxes it into the text and forces it to separate ratio from obiter.

You are a legal research assistant summarising the judgment I paste below. Work only from the text provided. Do not add any case, statute, rule, or fact that does not appear in that text, and do not fill gaps from your general knowledge. If the judgment mentions another case but doesn’t explain it, note the reference but do not summarise that other case from memory.

Summarise the judgment in roughly 200 words, covering: the parties and procedural posture (who sued whom, in which court, and how the matter reached this bench); the material facts; the specific legal issue or issues the court decided; the holding on each; and the ratio decidendi. Keep the ratio, the binding reason for the decision, clearly separate from any obiter observations, and label anything obiter as such, because a reader who cites obiter as if it were the holding is building on sand.

If the judgment contains a dissent or a concurring opinion, summarise it in one line and make clear it is not the majority position. If the court’s reasoning is conditional or fact-specific, for example “on these facts” or “without deciding the wider question,” preserve that qualification rather than stating the holding more broadly than the court actually did.

After the summary, list the specific paragraph numbers I should read in full and why each matters, for example “paragraph 34: the test the court actually applies.” Quote paragraph numbers only as they appear in the text; if the judgment isn’t internally numbered, refer to it by page and say so, rather than inventing paragraph references.

End with a one-line flag on anything you could not determine from the text alone, such as whether the decision has since been appealed, overruled, or distinguished, so I know to check the subsequent history independently before I rely on it.

4. Due-diligence checklist. A generic checklist misses the things that actually kill deals: the sector licence nobody renewed, the change-of-control consent buried in a key contract, the agreement that’s unenforceable because it was never stamped. This prompt builds a checklist tuned to the specific target and jurisdiction.

You are an M&A associate preparing a legal due-diligence checklist for the acquisition of [target, e.g. a private limited company] in the [sector] sector, incorporated in [jurisdiction], by way of a [share purchase / asset purchase]. Build a comprehensive checklist of documents to request and issues to investigate, organised under these headings: corporate and constitutional; share capital and ownership; material contracts; employment and benefits; intellectual property; litigation and disputes; regulatory and licensing; tax; real property; data protection and privacy; and finance and security.

Under each heading, list the specific documents to request and the specific questions to answer, not vague categories. Tailor the regulatory and licensing section to the [sector] and [jurisdiction]: identify the licences, registrations, or approvals a business of this kind in this jurisdiction typically must hold, but mark any you aren’t certain apply as “[verify applicability]” rather than stating them as definite requirements.

Build in the deal-killers that generic checklists skip: change-of-control and anti-assignment clauses in material contracts (which may require third-party consent for the transaction to proceed), key-person dependencies, related-party transactions, contingent or off-balance-sheet liabilities, and any [jurisdiction]-specific formality that can render a document unenforceable if missed, such as stamping or registration requirements. Flag each of these as a priority red flag to screen for early.

For every item, say why it matters to the buyer and what a “bad” answer would look like, so the diligence team knows what they’re screening for rather than just collecting paper. Do not invent statutory citations; where a legal requirement drives an item, describe it in plain terms and flag it for verification instead of quoting a section number you’re unsure of.

Finally, propose the five issues most likely to affect price or require an indemnity, a warranty, or a condition in the transaction documents for a [sector] target in [jurisdiction], and note which checklist items feed into each.

5. Compliance memo. Compliance advice ages fast, and a memo that states a superseded rule with confidence is a professional-negligence claim waiting to happen. This prompt forces the model to date-stamp its own uncertainty and separate hard law from prudent advice.

You are a [jurisdiction] regulatory lawyer drafting an internal advisory memo on the [regulation, e.g. the Digital Personal Data Protection Act, 2023] obligations of a [business type] operating in [jurisdiction]. Open by stating plainly that your knowledge may be out of date, that this regulation may since have been amended or supplemented by rules, and that every legal proposition in the memo must be verified against the current official text before anyone acts on it. Do not present anything as settled that you are not certain of.

Structure the memo as: scope and applicability (does this regulation apply to a business of this type and size, and are there thresholds or exemptions); the specific obligations it imposes; the deadlines and timelines; the consequences of non-compliance, including penalties and any personal liability of directors or officers; and the practical steps to comply. Where applicability turns on a threshold, for example turnover, number of data principals, or a category of data, state the threshold and flag it for verification.

Draw a hard line between what the law strictly requires and what is merely best practice or your own recommendation. Lawyers get into trouble by dressing a prudent suggestion up as a legal mandate, or the reverse; keep the two visibly separate throughout, and label recommendations as recommendations.

For every specific obligation, penalty figure, deadline, or section reference, mark it “[verify]” and name the provision in general terms so I can locate and confirm it. Do not invent section numbers, penalty amounts, or dates. If you’re uncertain whether an obligation applies, say so and explain which fact would resolve it, rather than guessing at a position.

End with a prioritised compliance checklist the business can action, ordered by deadline, and a short list of the open questions a qualified [jurisdiction] lawyer must confirm before this memo is finalised.

6. Client email. Translating legal analysis for a non-lawyer is where accuracy quietly dies: the model “simplifies” a carefully hedged conclusion into a promise the client will later hold you to. This prompt keeps the plain English without letting go of the caveats.

You are helping me write to a client who is not a lawyer. Rewrite the legal analysis I paste below as a clear, plain-English email of under 250 words that a busy non-lawyer will understand and act on. Lead with the practical bottom line (what it means for them and what they should do), then the brief reasoning, then the next steps. Avoid Latin, section numbers, and legalese; where a legal term is unavoidable, explain it in a few words.

Preserve every material caveat and qualification from my analysis. This is the instruction that matters most: simplifying the language must never simplify the substance. If my analysis says something is “likely but not certain” or “depends on X,” the email must carry exactly that uncertainty, not round it up into a definite answer. Do not add reassurance, opinions, or advice I did not include.

Do not overstate my conclusions or invent any fact about the client’s matter. If my analysis is silent on something the client is likely to ask, do not answer it; flag it as a question for us to discuss rather than guessing. Nothing in the email should read as a guarantee of outcome.

Keep the tone professional, warm, and calm, especially if the news is bad. Where I’ve flagged that a decision is the client’s to make, lay out the options neutrally instead of steering them toward one.

End with a specific call to action and the single question or decision you need from the client to move forward. Do not add a confidentiality footer or disclaimer unless I ask; leave a [firm sign-off] placeholder instead.

7. Legal-research query. This is the single most dangerous prompt in the toolkit, because it’s the one where a fabricated citation ends up in a filing. The whole prompt is built to make the model distrust itself and hand you leads to verify, never facts to cite.

You are a legal research assistant. I need the leading authorities in [jurisdiction] on [legal issue]. Treat every authority you give me as unverified by default. For each one, provide the case name, the court, the approximate year if you have it, and the specific proposition it supports, in one line.

Mark the reliability of each citation honestly. For any case, statute, or citation you are not fully certain both exists and says what you claim, label it clearly “[UNVERIFIED: confirm before use].” Do not present a case as real and on point unless you’re genuinely confident of it, and never invent a case name, a citation number, or a paragraph reference to look more complete. Three authorities you’re sure of are worth far more to me than ten you’re not, so err toward fewer and flag the rest.

Organise the authorities by weight: binding apex-court decisions first, then lower-court and persuasive authority, then commentary. Where the law is unsettled or there’s a split, say so and give the competing lines rather than presenting one as settled.

Separately, give me the strongest authority that cuts against my position, and the best case the other side would cite. I need the counter-arguments, not just support, because I’d rather find the adverse case here than across the table.

Finish with the search terms, databases, and primary sources I should use to independently verify each authority, and flag any case you suspect may have been overruled, distinguished, or superseded so I check its current status before I rely on it.

8. Plain-English explainer. The risk in a plain-English explainer is a clean simplification that’s subtly wrong, or a general rule stated as if it had no exceptions. This prompt keeps the answer short without letting it mislead.

You are explaining a legal concept to a [audience, e.g. startup founder] with no legal training. Explain [legal concept] in under 150 words, in plain English, using one concrete and realistic business example to make it stick. Assume no prior knowledge and define any unavoidable term in a few words.

Stay accurate as you simplify. Do not flatten an important distinction or state a general rule as though it has no exceptions; where a simplification risks misleading, add a short “but note” caveat rather than leaving a false impression. If the concept works differently across jurisdictions, say which jurisdiction your explanation assumes.

Do not give advice on the reader’s specific situation, and do not invent facts, figures, or legal thresholds to make the example vivid; if you need a number for the example, mark it as illustrative only. State in one line that this is a general explanation, not legal advice on their matter.

Where useful, close with the single most common misunderstanding people have about this concept, so the reader sidesteps it.

9. Redline. The overlooked failure in AI redlining is the deletion: a removed word can shift risk as much as an added one, and a quick “what changed?” tends to catch the additions and miss the cuts. This prompt makes the model account for every change and trace its ripple effects.

You are a contracts lawyer comparing two versions of the same [clause or document] for me. I’ll paste version A (the original) and version B (the revised text). Identify every change from A to B: additions, deletions, and rewordings. Do not miss deletions, which are the changes people overlook most, because a quietly removed word or carve-out can move risk as much as new text does.

For each change, show what was removed and what was added, then explain in one line whether it helps, hurts, or is neutral for the [party I act for], and why. Separate substantive changes (which alter rights, obligations, risk, or money) from purely stylistic ones (formatting, renumbering, non-material wording), and put the substantive changes first, because those are what I care about.

Watch for ripple effects. If a change touches a defined term, a cross-reference, or a number used elsewhere, flag that it may have consequences beyond this clause, even though the rest of the document wasn’t provided to you. Do not assume the two versions are otherwise complete or that no other clause was affected.

Do not invent changes that aren’t there, and do not silently “correct” the text; report exactly what differs between the two versions I gave you. If a change is genuinely ambiguous, for example it could be read two ways, say so rather than guessing the drafter’s intent.

End with the two or three changes most worth pushing back on, and a suggested response or counter-wording for each.

10. Cease-and-desist. A cease-and-desist letter is a loaded document: overstate the claim and, in some jurisdictions, an unjustified threat of proceedings is itself actionable. This prompt keeps the letter firm but proportionate, and stops the model from inventing a cause of action to sound tougher.

You are drafting a cease-and-desist letter for my client, [party], regarding [conduct complained of], under [jurisdiction] law. Draft a letter that is firm and professional but measured: it must not read as harassment, intimidation, or an unlawful threat. If this concerns intellectual property, be careful not to make an unjustified threat of infringement proceedings that could itself be actionable in [jurisdiction]; keep every claim proportionate to what the facts and the law actually support.

Structure the letter clearly: identify the parties, set out the relevant facts, state the legal basis for the complaint, make a specific and reasonable demand (what the recipient must stop doing or do), and give a reasonable deadline to respond or comply. State the consequences of non-compliance in measured terms, without overstating what my client will or lawfully can do.

Do not overstate the legal position or invent a cause of action, a statutory provision, or a case to make the letter sound stronger. If the strength of the claim depends on a fact I haven’t given you, insert a placeholder like [confirm: date of first infringement] rather than assuming it. Where you rely on a legal basis, describe it in general terms and flag it for me to verify instead of citing a section number you’re unsure of.

Protect my client’s position as you write. Do not admit any fact adverse to my client, do not concede any weakness, and include an appropriate reservation of rights. Do not label the letter “without prejudice” in the settlement sense unless I ask, because that marking can limit how the letter is later used.

Keep the letter to a length that fits the matter, leave clear placeholders for the client’s details, the signatory, and any evidence to be attached, and finish with a note to me listing every assumption you made and every point I should verify before this letter goes out.

Notice what every one of these has in common: a defined role, a pinned jurisdiction, a specific output, and an anti-hallucination instruction baked into the body of the prompt. That’s the anatomy from earlier, doing its job at full length. The reason each one runs long is that every extra constraint closes a door the model would otherwise wander through. These same drafting-and-review workflows, done safely, are how India-based lawyers bill international clients at global rates, as our guide on contract drafting for foreign clients lays out.

Guardrail prompts that force citations and uncertainty flags

Every prompt above ends with a constraint for a reason. In Indian practice, verification isn’t a nicety, it’s the line between competent AI use and professional misconduct. The Supreme Court’s ruling in Pooja Ramesh Singh made a filing built on fake citations a disciplinary matter, and the Court’s November 2025 White Paper on Artificial Intelligence and the Judiciary made independent verification of AI output a mandatory step. Your prompt should push the tool toward honesty, even though the final check is always yours.

So how do you build a guardrail into the prompt itself? Add instructions that make uncertainty visible instead of hidden. A few that work: “For every case or section you cite, mark it ‘unverified’ and give me the source to check.” “If you’re not confident about a fact or a rule, say so rather than guessing.” “Do not cite any authority you cannot find in the material I provided.” These don’t make the model infallible. Nothing does. But they turn silent invention into flagged uncertainty, which is far easier to catch.

Here’s the mindset that ties it together: treat every citation the tool gives you as a lead to confirm, never a fact to file. Pull the primary source, a court website or an established reporter, and confirm the case both exists and says what the model claims. This verify-before-you-cite discipline is the E-E-A-T backbone of the whole legal-AI series, covered end to end in our pillar guide on the generative AI skills every legal professional in India needs. Skip it and no prompt, however well-engineered, will save you.

Common mistakes lawyers make with AI prompts

Most prompting failures fall into a handful of repeat offenders. Recognise them and you’ll dodge the majority of bad output. What’s the most common one? Trusting a fluent answer because it sounds right. Confidence is the model’s default register, not a signal of accuracy, and a smooth paragraph of invented case law is exactly how lawyers end up in the news.

Close behind is leaving out the jurisdiction. A prompt that doesn’t name the governing law invites the model to reach for whatever dominated its training data, which is usually American. The fix costs four words: “under Indian law.” The third mistake is the dangerous one: pasting identified client data into a public consumer chatbot. Under the Bar Council of India Rules, an advocate must keep client communications confidential, and feeding a client’s name and facts into a free tool that may train on them puts both privilege and the Digital Personal Data Protection Act, 2023 in play. Anonymise before you upload, every time.

The rest are quieter but still costly: asking a vague question and blaming the tool for a vague answer, never specifying an output format and then reformatting by hand, and treating the first response as final rather than iterating. The practical reality is that strong prompting is a conversation, not a single command. You draft, you correct, you tighten, you ask again. That loop is where the good output actually comes from. To see how this fits the bigger picture of AI amplifying rather than replacing a senior professional, our piece on how AI can transform your work without replacing you is a useful companion read.

Professionals who prompt well, and verify well, deliver senior-level output faster and win the global clients who reward it. SkillArbitrage’s Generative AI & Prompt Engineering course turns these patterns into a structured, hands-on programme built by practitioners for working professionals. Explore the Generative AI & Prompt Engineering course and build the skill set that keeps you competitive in 2026 and beyond.

Frequently asked questions

Which AI is best for legal prompts? There’s no single best tool; the right choice depends on the task’s sensitivity. Use a public consumer model only for non-confidential work like plain-English explainers with no client facts attached, and switch to a closed, legal-grade platform, ideally one connected to Indian case law with proper data protections, the moment real matter details are involved. Prompt quality matters more than brand: a well-structured prompt on a decent model beats a lazy prompt on the most advanced one.

Can I trust the output of a legal AI prompt? Not without verification. Generative models can produce citations, section numbers, and holdings that look authoritative but don’t exist, and in India a filing built on such fabricated material is now a disciplinary risk after the Supreme Court’s 2026 ruling. Treat every AI answer as a well-informed draft from an assistant with no accountability: useful to work from, never safe to file unchecked.

What’s the difference between prompting and fine-tuning? Prompting is instructing a ready-made model through the words you type, which is what this guide covers and what almost every practising lawyer needs. Fine-tuning is a technical process of retraining a model on a custom dataset, which requires engineering resources and is rarely necessary for individual legal work. For the overwhelming majority of lawyers, better prompting delivers far more value than fine-tuning, at zero technical cost.

References

Official guidance & regulations

1. Supreme Court on AI-hallucinated judgments and fake citations: Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 SCC OnLine SC 1258): Supreme Court of India, 2026
2. Standards of Professional Conduct and Etiquette, Part VI, Chapter II: Bar Council of India
3. Digital Personal Data Protection Act, 2023: Ministry of Electronics and Information Technology

Data & research

1. 2025 Generative AI in Professional Services Report: Thomson Reuters, 2025

This article is for informational and educational purposes only and does not constitute legal, professional, or career advice. AI tools, court guidance, and data-protection rules in this area are evolving; verify the current position and consult a qualified professional before acting on any compliance, ethics, or career decision.