Pregătește înainte să întrebi: un brief al dosarului pentru promptul tău de analiză a contractelor cu AI

The review is decided before you write a prompt

Picture a familiar start. You upload the draft and type one line: "Review this contract and flag the risks."

The model now has the document. It does not have the matter. Nothing in that line tells it which side you are on, what was agreed on the phone last week, which clause your client will walk away over, or that Schedule 2 is still sitting in someone's inbox. So it falls back on a generic review of everything that could matter to anyone.

That limit is not a secret. ABA Formal Opinion 512, the American Bar Association's July 2024 opinion on generative AI, says these tools "lack the ability to understand the meaning of the text they generate or evaluate its context." Claude's own prompting guide tells users to think of Claude "as a brilliant but new employee who lacks context on your norms and workflows." Hand a new employee a contract with "any thoughts?" written on it and expect a long memo. A model is no different.

One of the prompt guides ranking for this topic puts the failure plainly: people give the model "a one-line instruction like 'review this NDA and flag issues' and act surprised when the output is a generic risk list that ignores the playbook, the deal, and the client."

We have already written about how to phrase the ask. Prompt architecture for lawyers covers structure, and our legal prompting guide covers the Role + Context + Task + Format formula. This post is about the step before both: the matter brief, the facts about this deal that exist whether or not you use AI, written down for a model.

What the top-ranking prompt pages ask for, counted

We wanted to know whether the advice already out there covers the matter or only the contract. So we counted.

Method. On 1 October 2026 we ran one US web search for "AI contract review prompt" and took the nine results it returned, mostly prompt libraries, including two vendor guides and one community forum post. One page showed no prompt text when we fetched it, so we scored eight. We read every prompt and every instruction on each page and marked an item "yes" if anything on the page asked the user to supply it or told the model to act on it. The rule is generous on purpose: one mention anywhere on a long page counts.

The pattern is clear. The advice covers the contract and the reader's role well. It barely touches the state of the negotiation. Six of the eight pages cap the output somewhere, which sounds like the fix. But a top five chosen without knowing what you conceded or where your red lines are is a shorter version of the same flat list. The cap is only as good as the inputs behind it.

This is one search on one day, scored by us. Treat it as a snapshot of the advice that ranks, not a survey of how lawyers work.

The matter brief, ready to copy

Fill this in before you upload anything. Use roles, not names (more on that below). Paste it above the contract in whatever tool you use.

MATTER BRIEF (paste above the contract)

1. We act for: [party and role, e.g. "the Distributor"]
2. Governing law and arbitration seat we expect: [e.g. "UAE law, ADGM-seated arbitration"]
   If the contract says something different, flag it first, before any other finding.
3. Deal context and leverage: [what the deal is for, its rough size, who needs it more, deadline]
4. Already conceded or agreed: [points settled in heads of terms or earlier rounds; do not reopen]
5. Red lines: [terms the client will not accept, in one line each]
6. Documents in scope: [list what is attached]
   Known to be missing: [e.g. "Schedule 2 (pricing) not yet received"]
   If a finding depends on a missing document, say so and give no overall risk score.
7. Output: the top 3 to 5 issues for our side, ranked. For each: clause number,
   one line on why it matters for this deal, proposed action. Then one line listing
   everything else you reviewed. No more than one page.
8. Ignore: market-standard boilerplate (notices, counterparts, severability,
   entire agreement) unless it conflicts with points 4 or 5.
9. Proposed replacement wording: label it "opening position" and give one fallback.
If any of points 1 to 6 is blank, ask me before you start.

Each line earns its place.

1. Which side you act for

The same clause can help one party and hurt the other. Without a side, the model has to review for both, which lengthens the list and blurs every recommendation on it.

2. Governing law and arbitration seat, and what the contract actually says

Governing law is the contract clause that decides which law applies if there is a dispute. The seat of an arbitration is its legal home, and it decides which arbitration law governs the proceedings: the main provisions of England's Arbitration Act 1996, for example, apply where the seat is in England and Wales or Northern Ireland. The two are separate choices and can point in different directions. In Abu Dhabi Global Market, for example, the Application of English Law Regulations 2015 make English common law directly applicable. A model that assumes a different baseline may call the wrong clauses unusual. The instruction to flag a mismatch matters as much as the field: the draft you were sent may not say what was agreed.

3. Deal context and leverage

A small vendor contract and a strategic partnership deserve different tolerance for the same clause; one of the ranking guides makes exactly this point with a small vendor deal and a large partnership. Leverage decides what is worth asking for at all. One sentence on who needs the deal more keeps the model from proposing fights you cannot win.

4. What is already conceded

This is the line all eight pages we scored leave out. If payment terms were settled in an earlier round, a review that ranks them third has spent a third of your shortlist on them. Worse, it may recommend reopening a point your client gave away to win another.

5. Red lines

By red line we mean a term the client will not accept at any price. Stating it does two things. It moves any clause that crosses it to the top of the list. And it gives the model a test for everything else: does this clause come close?

6. Documents in scope, and what is missing

Contracts point at other documents: schedules, service levels, price lists, the master agreement an order form sits under. If the liability cap is "fees paid under Schedule 2" and Schedule 2 is not attached, the cap cannot be assessed. Nothing in a one-line request tells the model to stop at that gap rather than score around it. Tell it to stop. The Claude docs on reducing hallucinations make the same move in general terms: "Explicitly give Claude permission to admit uncertainty." A review that says "no score: Schedule 2 missing" is more useful than a confident number built on half the file.

7. Output format and length

Ask for a ranked top three to five, each tied to a clause number, a reason that refers to this deal, and an action. Then one line for everything else. The last line matters: it tells you what the model looked at and chose not to raise, which is what you will check first.

8. What to ignore

Boilerplate is stock language: the standardized clauses that appear in contract after contract of the same type, such as notices, counterparts, severability and entire agreement. Telling the model to skip it, unless it collides with a concession or a red line, leaves the shortlist for the clauses that are actually in play.

9. Label aggressive wording as an opening position

Replacement wording written for your side is an opening ask. The risk is that it reads like a reasonable middle and goes out as one. Asking for the label, plus one fallback, keeps you from sending a first offer that sounds like a final one.

What a good AI should ask you if you did not tell it

The last line of the template ("if any of points 1 to 6 is blank, ask me before you start") is the safety net. The Claude prompting guide has a test for this: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." A capable colleague handed a bare contract would ask questions before reading clause one. These are the ones worth getting:

• Which party are we acting for?
• Which law governs, where is the seat, and is that what the draft actually says?
• Is anything already agreed that I should not reopen?
• Is there anything the client will not accept?
• Are all the schedules and related documents attached? If not, which are missing?
• How long should the answer be, and who is going to read it?
• Should I skip the standard clauses?

A review that starts without any of these is a review of the document, not of your matter.

Before and after, on an illustrative contract

The setup: you act for a distributor reviewing a supplier's standard-form distribution agreement covering three Gulf markets. Your heads of terms (the short summary of the deal agreed before the full contract) expected onshore UAE law and an ADGM-seated arbitration, and settled the payment terms. Your client's red line: no termination for a volume shortfall in year one. The supplier has not yet sent Schedule 2, the price and rebate schedule.

Before. One line: "Review this contract and flag the risks." Our illustrative answer:

Risk review: Distribution Agreement
1. Clause 2.3  Exclusivity limited to the Territory. Medium risk.
2. Clause 4.1  Supplier may change prices on 60 days' notice. High risk.
3. Clause 5.2  Payment within 30 days of invoice. Medium risk.
4. Clause 9    Supplier liability capped. High risk.
5. Clause 17   Notices must be in writing. Low risk.
6. Clause 18   Entire agreement clause excludes prior statements. Medium risk.
7. Clause 21   Agreement may be signed in counterparts. Low risk.
[... the list continues clause by clause ...]
Overall risk rating: Medium.

Read it closely and the problems show up in the structure. It is in document order, not order of importance. Boilerplate sits next to the clauses that matter. Payment terms you settled in the heads of terms are flagged as a risk. It does not mention that the draft names English law, not the law your heads of terms expected. And it gives an overall rating for an agreement whose pricing schedule it has not seen.

After. The same contract, with the matter brief above it:

Acting for: Distributor.
Governing law: the draft names English law. Your brief expected onshore UAE law. Flagged first.
No overall rating: Schedule 2 (prices and rebates) not provided. Clause 6 depends on it.

Top 3 for this deal
1. Clause 6    Minimum purchase volumes, with supplier termination for any shortfall.
               Crosses your red line (no termination for shortfall in year one).
               Action: delete the year-one termination right. Opening position.
               Fallback: termination only after two consecutive years of shortfall.
2. Clause 12.2 Supplier may appoint other distributors after 12 months.
               Undercuts the exclusivity the deal is priced on.
               Action: tie any new appointment to a volume test you control.
3. Clause 4.1  Unilateral price changes on 60 days' notice.
               Action: cap annual increases, or give the distributor an exit.

Not reopened (agreed in heads of terms): clause 5, payment.
Reviewed, market standard, no action: clauses 17 to 24.

The second answer is shorter and more useful because it was told what this matter is. In the illustration, nothing changes between the two except the brief.

Why a good brief also means fewer re-runs

Without a brief, the corrections arrive one at a time. "I act for the distributor." Run it again. "Ignore the boilerplate." Again. "Payment is agreed." Again. Each run is a fresh generation, and generation is not fully repeatable: The Claude API documentation notes that "even with temperature of 0.0, the results will not be fully deterministic." So a re-run can reshuffle findings you had already checked, and you end up verifying the same clauses twice.

Front-loading the brief replaces that loop with one fuller first pass. It also makes the checking cheaper, because a ranked list of five, each tied to a clause number, is something you can actually verify clause by clause before it goes anywhere.

The judgment in the brief stays with you

ABA Formal Opinion 512 is US guidance, based on the ABA Model Rules of Professional Conduct, but its reasoning travels. It names contract review among the tasks these tools may assist with. It warns that relying on output "without an appropriate degree of independent verification or review of its output" could breach the duty of competence. And it draws the line the brief sits on: lawyers "may not abdicate their responsibilities by relying solely on a GAI tool to perform tasks that call for the exercise of professional judgment," and "regardless of the level of review the lawyer selects, the lawyer is fully responsible for the work on behalf of the client."

Deciding what matters in a deal is that professional judgment. The matter brief is where you exercise it, in writing, before the model starts.

One more duty applies to the brief itself. The opinion says that before lawyers put client information into one of these tools, "they must evaluate the risks that the information will be disclosed to or accessed by others outside the firm." That is why the template uses roles, not names: "the Distributor" and "the Supplier" do the job without carrying the client's identity into the prompt. How far to go beyond that is the subject of our post on confidentiality, and the firm-level policies behind it are in our post on the legal AI implementation gap.

HAQQ's take

We made a version of this argument in that implementation-gap post, about a folder dragged wholesale into a chat tool: the tool was handed "a folder, when it needed a question and the two documents the question depends on." The matter brief is that question, written down.

It works in any tool, plain Claude included; if that is your setup, our post on why power lawyers stay on Claude is the companion read. In HAQQ, contract review checks each clause against your firm's playbook, which the product page describes as defining "your standard positions, acceptable alternatives, and deal-breakers for different contract types." A playbook is your standing brief: it holds what is true of every deal of a type. The matter brief holds what is true of this deal, this week. Paste it at the top either way.

Key takeaways

• The quality of an AI contract review is set before you write a prompt. A one-line request gets you a review of the document, not of your matter.
• Of the eight top-ranking prompt pages we scored, 7 ask which side you act for, 1 asks for your red lines, and 0 ask what you have already conceded.
• Use the nine-field matter brief: side, law and seat, deal context, concessions, red lines, documents and gaps, output, what to ignore, and an opening-position label.
• Tell the model to give no overall score when a document it depends on is missing.
• A short, ranked, clause-referenced list is one you can verify. ABA Formal Opinion 512 makes clear that the lawyer remains "fully responsible for the work on behalf of the client" whatever the tool.

Sources & further reading

• ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512, Generative Artificial Intelligence Tools, 29 July 2024 (PDF)
• Claude Platform Docs, Prompting best practices, "Be clear and direct"
• Claude Platform Docs, Reduce hallucinations
• Claude API Reference, Create a Message (temperature)
• ADGM, The English Common Law System
• Arbitration Act 1996, section 2 (scope) and section 3 (the seat), legislation.gov.uk
• Cornell LII Wex, governing law
• Cornell LII Wex, boilerplate
• Gannons, Heads of Terms explained
• How to phrase the ask
• The Role + Context + Task + Format formula
• Four policies and a verification habit