AI Prompts for Lawyers: 631-Prompt Library + 2026 Guide

Key facts

• The HAQQ Prompt Library ships 631 prompts across 11 use cases; the largest is Draft / Generate with 189 prompts, then Review / Redline with 80 (HAQQ prompt library data, last updated 19 Oct 2026).
• 382 of the 631 prompts are jurisdiction-neutral and 249 are tagged to a country or region, including 17 for the UAE, 11 for Saudi Arabia, 9 for Egypt and 4 for the EU (HAQQ prompt library data, 19 Oct 2026).
• Every library prompt is published in English, Arabic, Spanish, French and Romanian (HAQQ prompt library metadata, 19 Oct 2026).
• ABA Formal Opinion 512 on generative AI tools, issued 29 July 2024, ties a lawyer's use of AI to the duties of competence, confidentiality, client communication, supervision, candor and reasonable fees (ABA Standing Committee on Ethics and Professional Responsibility, 29 Jul 2024).
• In United States v. Heppner, Judge Jed Rakoff (S.D.N.Y.) held on 17 February 2026 that 31 documents a defendant generated with a consumer AI tool were protected by neither attorney-client privilege nor work product (Debevoise & Plimpton summary, Feb 2026).

What should a legal prompt include?

A prompt is the brief you give the model. Write it the way you would brief a capable junior who has never seen the file: who you act for, which law applies, what happened, what you need, and what the deliverable should look like. Each element removes a guess the model would otherwise make for you.

• Role and client: 'I act for the employer.' The same clause reads differently from each side of the table.
• Jurisdiction and forum: 'UAE mainland, Federal Decree-Law No. 33 of 2021.' Name the statute if you know it, and the court or free zone if it matters.
• Facts: the dates, amounts, roles and documents that change the answer. Placeholders such as [Party A] keep client names out.
• Task: one verb per prompt. Review, draft, compare, summarize, or research.
• Output format: 'A table with columns for clause, risk, authority and proposed wording, then a three-line summary for the client.'

How do I write a legal prompt step by step?

• 1. Write one sentence stating who you act for and what decision the output will support.
• 2. Name the governing law, the forum and, where you can, the statute and article.
• 3. Paste or attach the operative documents, with client identifiers replaced by placeholders.
• 4. List the facts that matter as short numbered lines.
• 5. State the deliverable, its audience (client, partner, court) and its length.
• 6. Add the rules: cite the source for every legal proposition, quote the clause behind every risk, and say 'not determinable from the documents' instead of guessing.
• 7. Read the output against the sources, then refine with follow-up prompts.

Why does the wording of a prompt change the answer?

A large language model generates text by predicting what comes next from its training and from everything in your prompt. It does not look up a stored answer. Three practical consequences follow for lawyers.

• The model weighs everything you give it. Paste a badly drafted precedent as an example and it may copy the bad parts. Show only drafting you would sign.
• Output varies from run to run. A structured prompt with a fixed format narrows that variation, which matters when two associates run the same review.
• Long, multi-part requests degrade. A sequence of focused prompts usually beats one prompt that asks for research, analysis and a draft at once.

A shorter version of the same structure, used in several vendor training guides, is intent, context, instruction and format: say what you are trying to decide, give the background, then give the instruction and the shape of the answer.

Which prompting techniques work best for lawyers?

These seven techniques show up in every prompt we ship in the library. None of them needs technical knowledge.

1. Assign a role

Telling the model to act as a specific kind of lawyer narrows its answer to that specialist's frame. 'Act as a senior employment lawyer in Dubai reviewing this offer letter for the employee' produces a different, more useful review than 'review this offer letter'.

2. Give the context the model cannot know

Case type, jurisdiction, parties, governing law, deadlines and commercial constraints. 'A cross-border supply agreement between a US manufacturer and a German distributor, governed by German law, with a renewal date in 60 days.' The less the model has to assume, the less it invents.

3. Break complex work into steps

Instead of 'draft a full board resolution', ask first for the outline, then the recitals, then the operative clauses. You review each stage before the next one builds on it.

4. Specify the output format

Table, memo, numbered list, redline or plain-language summary. Show the pattern you expect, for example '[dd/mm/yyyy]: [event] ([source document, page])' for a chronology. A defined format is often the difference between text you reformat for an hour and a deliverable you can send.

5. Set positive guardrails

Say what the model should do. 'Only analyse the indemnity and limitation of liability clauses under English law' works better than 'ignore irrelevant clauses'. Tell it what to do when unsure: 'If a clause is ambiguous, flag it for human review and do not interpret it.'

6. Iterate

Treat the first answer as a draft. Follow up: 'Expand the data protection section.' 'Rewrite paragraph 3 for a CFO.' 'Which of these points depends on a fact I have not given you?' Each round costs seconds.

7. Repeat the instruction that matters most

Models pay less attention to the middle of long prompts. Put the critical instruction at the start and again at the end, for example 'cite the article for every statement of law'.

What else improves a legal prompt? Five principles from our prompt architecture notes

Our earlier guide on prompt architecture, now merged into this page, listed seven principles. Two of them (role and iteration) are covered above. These five add something the techniques do not.

• Work from the documents, not from memory. Upload the contract, judgment or statutory text and tell the model to answer from it. A model answering from training data alone is where invented citations come from.
• Use few-shot examples. Before asking for a clause, show one or two clauses drafted to your standard. The model matches the pattern far more closely than it follows a description of the pattern.
• Define the audience. A client letter, an internal memo and a court submission need different register, length and caveats. Say which one you are writing.
• Ask for sources, then check them. Instruct the model to cite the authority behind each proposition, and verify each one yourself before relying on it.
• Verify before anything leaves your hands. ABA Formal Opinion 512 (29 July 2024) treats review of AI output as part of the duty of competence, and the level of review depends on the tool and the task. The lawyer signs the document.

What do jurisdiction-specific legal prompts look like?

A prompt that names the statute gets a narrower, more checkable answer than one that says 'under local law'. The four templates below follow the structure the library uses: who you act for, the facts as fill-in fields, numbered deliverables, and mandatory rules. Each one points to the matching entry in the HAQQ Prompt Library by its library ID; filter by jurisdiction to find it.

UAE: non-compete validity under the Labour Law

I act for [the employer / the employee]. Assess whether this non-compete is enforceable under Article 10 of UAE Federal Decree-Law No. 33 of 2021 on the Regulation of Labour Relations.

Facts: role [Role]; restriction period [Period]; geographic scope [Scope]; activities covered [Activities]; how employment ended [resignation / termination / expiry]; mainland or free zone [Mainland / DIFC / ADGM / other free zone].

Deliverables:
1. Whether the clause meets the statutory limits on duration (no more than two years after the contract ends), place and type of work.
2. Whether the role gave access to clients or confidential information, and what the employer must prove.
3. Whether the way employment ended makes the restriction fall away.
4. If the employer is in the DIFC or ADGM, say that their own employment rules apply and do not apply the mainland analysis.

Rules: cite the article for every statement of law. If a fact I have not given changes the answer, say which fact.

Library entry: AE-02, Non-Compete Validity Under UAE Labour Law. The UAE bucket holds 17 prompts, from liability caps under UAE law (AE-01) to commercial agency termination (AE-07).

Saudi Arabia: transferring personal data outside the Kingdom

I act for [the Saudi controller]. We want to host [type of personal data] for [number] data subjects with a provider in [country]. Advise under Article 29 of the Saudi Personal Data Protection Law and SDAIA's Regulation on Personal Data Transfer Outside the Kingdom (updated September 2024).

Facts: purpose of the transfer [Purpose]; whether the data includes sensitive data [Yes/No]; safeguards available [standard contractual clauses / binding common rules / certification / none].

Deliverables:
1. The legal basis that permits this transfer, or why none does.
2. Whether a transfer risk assessment is required before the transfer, and what it must cover.
3. The contract terms the provider must accept.
4. A one-paragraph answer the CEO can read.

Rules: distinguish what the Law says from what the Regulation adds. Flag any point where SDAIA guidance is pending or unclear.

Related library entries: P44-30, Saudi AI Services Agreement Negotiation Memo, and P47-08, Saudi Company Ongoing Compliance Checklist. The KSA bucket holds 11 prompts.

Egypt: asking the court to appoint an arbitrator

I act for the claimant. The respondent has not appointed its arbitrator. Set out how to obtain a court appointment under Article 17 of Egyptian Law No. 27 of 1994 on Arbitration in Civil and Commercial Matters.

Facts: arbitration clause [quote it]; number of arbitrators [Number]; international or domestic arbitration [International / Domestic]; date of our request to appoint [Date].

Deliverables:
1. The competent court under Article 9 (for international commercial arbitration, the Cairo Court of Appeal unless the parties chose another appellate court).
2. The 30-day period that must run before we apply, and how to count it.
3. The form of the application and the documents to attach.
4. Whether the court's appointment decision can be challenged.

Rules: work from the clause I quoted. If the clause names an institution such as CRCICA, say whether its rules displace the court route.

Library entry: EG-03, Appointing an Arbitrator by Petition. The Egypt bucket holds 9 prompts, including a Court of Cassation research prompt on medical damages (EG-02).

EU: reviewing a processor agreement under the GDPR

I act for [the controller]. Review the attached data processing agreement against Article 28(3) of Regulation (EU) 2016/679 (GDPR).

Facts: the processor is [Vendor]; the service is [Service]; sub-processors are [listed / not listed]; data leaves the EEA [Yes/No, to which country].

Deliverables:
1. A table with one row per Article 28(3) requirement: present / missing / weak, the clause number, and the exact wording.
2. Proposed wording for every missing or weak term.
3. If data leaves the EEA, the transfer mechanism the agreement relies on and whether it is documented.

Rules: quote the clause behind every finding. Do not treat a reference to 'applicable law' as satisfying a specific requirement.

Related library entries: P02-02, Data Processing Agreement, and P43-25, EU Data Access Claim Strategy Playbook for Court. The EU bucket is small (4 prompts) because most GDPR prompts in the library are tagged jurisdiction-neutral.

What does a weak prompt look like after a rewrite?

Three rewrites we use in training sessions. The 'what changes' notes describe the typical difference in output we see when lawyers run both versions; your model and documents will vary.

Rewrite 1: contract review

BEFORE
Review this contract and tell me if it's OK.

AFTER
I act for the customer, a mid-size retailer. Review the attached SaaS agreement, governed by English law. List every clause that limits the vendor's liability, lets the vendor change price or service, or restricts our exit. For each: quote the clause, rate the risk high / medium / low for us, and propose replacement wording. End with the three points I should raise first in negotiation. If something depends on a fact not in the document, say so.

What changes: the first version returns a general summary that says the agreement 'appears standard' and lists headings. The second returns a clause-by-clause table anchored to quoted text, which you can check against the contract in minutes, and a negotiation order you can disagree with.

Rewrite 2: legal research

BEFORE
Is a non-compete enforceable in Dubai?

AFTER
I act for a Dubai mainland employer. A sales manager with access to our client list resigned and joined a competitor in the same emirate four months later. His contract has a 12-month, UAE-wide non-compete. Under Article 10 of Federal Decree-Law No. 33 of 2021, is the clause enforceable as drafted, what must we prove, and what is the deadline to bring a claim? Cite the article for each point and flag anything that differs if the employer were in the DIFC.

What changes: the short question gets a survey of UAE non-compete law that mixes mainland and free-zone rules. The rewrite gets an answer to this employee's case, organised around the statutory conditions, with the points that turn on evidence flagged for you to check.

Rewrite 3: client update

BEFORE
Summarize this judgment for the client.

AFTER
The client is the CFO of the defendant, not a lawyer. Using only the attached judgment, write a client update of no more than 250 words: what the court decided, what it means for the company's exposure in money terms, the deadline to appeal, and the two decisions the client must make this week. Plain English, no case citations in the body, and a one-line note of anything in the judgment you could not interpret.

What changes: the first version produces a legal summary written for another lawyer. The second produces something the CFO can act on, with the appeal deadline surfaced, and tells you where the model was unsure so you know what to check.

Ready-to-use legal prompts by practice area

Short starting points you can paste and then extend with the five elements above. Each has a fuller, structured version in the library.

Contract review and drafting

• 'You are a senior commercial lawyer acting for the buyer. Review this supplier agreement for risks in indemnity, liability, termination and exclusivity. Explain each in business language and quote the clause.' (Library: P11-01, Full Contract Risk Review.)
• 'Draft a limitation of liability clause for a SaaS agreement under English law, capping liability at 12 months of fees paid and excluding indirect and consequential loss.'
• 'Redline this NDA so it is mutual, prohibits reverse engineering, and protects trade secrets for as long as they remain secret.' (Library: P01-01, Standard NDA.)
• 'Summarize the key terms of this statement of work in plain English for the project manager, in a table of obligation, owner and deadline.'

Data privacy and compliance

• 'Create a checklist for responding to consumer access, deletion and opt-out requests under the CCPA as amended by the CPRA.'
• 'Explain the difference between a controller and a processor under Articles 4(7) and 4(8) GDPR, with examples for a SaaS company.'
• 'List what an in-house team should do in the first 48 hours of a suspected breach of customer data, noting the 72-hour notification window in Article 33 GDPR.'
• 'Draft a one-page internal policy on using AI tools with client and personal data.'

Legal research and analysis

• 'Under English law, summarise how courts interpret force majeure clauses in commercial contracts, with the leading cases and what each decided.'
• 'What statutes and regulations govern data portability in the European Union? Cite article numbers.'
• 'Draft the argument section of a brief that this non-compete is unenforceable, using only the authorities I attach.'

Corporate governance and M&A

• 'Write a one-page overview of directors' fiduciary duties under Delaware law: care, loyalty and the business judgment rule.'
• 'Draft a board resolution authorising the CEO to sign a partnership agreement with [Provider].'
• 'List legal due diligence questions for acquiring a software company, covering IP, material contracts, litigation and employment.'
• 'Explain what a no-shop clause is and how it protects each party in an acquisition.'

Can I paste client data into ChatGPT, Claude or Gemini?

Check the plan and the data terms first. Consumer tiers of general AI tools may store prompts and, depending on settings, use them to improve models. In United States v. Heppner (S.D.N.Y., 17 February 2026), Judge Rakoff held that documents a defendant generated with a consumer AI tool were not privileged, in part because the provider's privacy policy reserved the right to disclose user data. Commentators read the ruling as tied to consumer terms; enterprise agreements with confidentiality commitments put you in a stronger position, though no court has said they are enough.

• Replace names, dates, amounts and company details with placeholders before pasting anything into a consumer tool.
• Do not upload full contracts, privileged advice or regulated personal data to a tool whose terms you have not read.
• Prefer an enterprise plan or a legal platform with a data processing agreement, no training on your data, and a hosting region you can name.
• Keep a note of which tool processed which matter, in case a court or client asks.

How does a legal AI platform change the way you prompt?

In a general chatbot every session starts from zero, so you restate your role, jurisdiction and standards each time. A legal platform carries that context for you, so the prompt can be shorter and focus on the task. In HAQQ Legal AI Chat, the matter, its documents and your playbook stay attached to the conversation, answers come back as structured documents you can export, and data is hosted in the EU, the US or the Middle East, or on private cloud or on-premise. The interface runs in seven languages (English, Arabic with full right-to-left layout, French, Spanish, German, Italian and Portuguese).

The prompting discipline stays the same. A platform can supply context; it cannot decide which facts matter or check the result for you.

How do I build a prompt library for my team?

Save every prompt that produced work you used, with the matter type, the model and the date. Strip client details, turn the variables into [placeholders], and add the rules you had to add by hand on the second attempt. After a few weeks you have a set of tested prompts for your recurring work, and new joiners start from your standard instead of from a blank box.

To skip the first month, start from the HAQQ Prompt Library: 631 free prompts sorted by use case, practice area, jurisdiction, audience and style, each in English, Arabic, Spanish, French and Romanian. Copy one, adjust the facts, and keep your edited version as the firm's own.

Where should I start?

• Pick one high-volume, low-risk task this week: NDA review, a chronology from a document set, or a client update.
• Write the prompt with the five elements, run it on a closed matter where you know the right answer, and compare.
• Fix the prompt where the output went wrong, save it, and share it with one colleague.
• Move to higher-stakes work only after the prompt holds up on several closed matters.
• For testing a model or tool on your own questions, see our method in how to test legal AI.

General legal information, not legal advice; the statutory references in the templates above should be checked against the current text before you rely on them.

Sources

• UAE Federal Decree-Law No. 33 of 2021, Article 10 (non-compete), commentary by BSA Law
• SDAIA Regulation on Personal Data Transfer Outside the Kingdom (PDPL Article 29), summary by Securiti
• Egypt Law No. 27 of 1994 on Arbitration in Civil and Commercial Matters (WIPO Lex)
• Regulation (EU) 2016/679 (GDPR), EUR-Lex
• ABA Formal Opinion 512, Generative Artificial Intelligence Tools (29 July 2024)
• United States v. Heppner (S.D.N.Y. 2026), Debevoise & Plimpton summary

Related reading

• write a matter brief before your contract review prompt
• context engineering: what to give the model besides the prompt
• how large language models work, for lawyers
• which models we tested on legal work, and how they scored
• our audit of AI hallucinations in court filings