Prompt Engineering for Lawyers: The One Framework Behind CLAIM, RICE and the Rest

On 19 March 2026 a non-lawyer named Nav Toor posted twelve prompts on X under the headline “12 prompts that replace $15,000 in legal bills”. Each began with a persona: “You are a senior corporate attorney at Skadden Arps who drafts NDAs for Fortune 500 companies”. The thread drew 3,919 likes, a backlash from lawyers, and one sharp question from Artificial Lawyer: “It’s possible that Claude simply read ‘Wachtell’ and thought ‘OK, this means do the contract in the style of any large commercial law firm’.”

That question is the whole subject of prompt engineering for lawyers: what in a prompt changes the output, and what is costume? The industry answers with acronyms. SAL and Microsoft teach Goal, Context, Expectations, Source. Harvey teaches CLAIM. The North Carolina Bar teaches RICE. Thomson Reuters teaches Intent, Context, Instruction. Clio teaches six components. They are all the same framework, and once you see the seven parts underneath you can stop collecting acronyms and start briefing.

Why every framework is the same seven parts

Lay the guides side by side and the overlap is total. Each asks whom the model works for, what it knows about the matter, what to do, what to use, what shape the answer takes, what not to do, and what happens next.

Harvey’s summary is the best of the lot: “Vague prompts produce generic answers. Structured prompts produce reviewable work product.” Reviewable is the word to hold on to: the point of structure is a draft you can check.

Role, context, task, sources, format, constraints, iteration: the anatomy

Harvey’s reusable template shows all seven in one breath: “You are assisting a [type of lawyer/team] working on a [matter type] in [jurisdiction]. Using [documents/sources], please [task]. Focus on [issues]. Do not [limitations]. Format the output as [table/memo/checklist/outline]. Include [citations/source references/factual assumptions/open questions]. Flag any uncertainty or areas requiring attorney review.”

Perspective, not persona. The useful content of a role is the side you act for, the reader and the standard. Anthropic’s prompting guidance says “modern models are sophisticated enough that heavy-handed role prompting is often unnecessary” and that being explicit about the perspective you want is more effective. “Review for the customer, who has little bargaining power and wants to sign in two weeks” changes the output. “You are a Skadden partner” changes the tone.

Context. Facts, parties, posture, commercial pressure, and above all jurisdiction (next section). Nico Kuhlmann of Hogan Lovells calls the context window the “digitale Handakte” and adds the caution most guides omit: “Mehr Kontext ist nicht automatisch besser.” Relevant file, not whole file.

Task. One precise verb. Thomson Reuters’ CoCounsel team: “CoCounsel takes things literally, so extraneous descriptors can be confusing or misleading.” Their before/after: “What state pension funds are prohibited from considering ESG factors?” became “Identify each state that prohibits pension funds from considering environmental, social, and governance (ESG) factors.”

Sources. The SAL guide ends every sample prompt with a fence: “Use only these materials: [list]. Where attachments are referenced, restrict analysis and drafting to those materials and provide pinpoint references where applicable.” Pinpoints turn review into spot-checks.

Format. Name the columns (section below).

Constraints. Affirmative, not negative. Legal Genie’s Australian prompt guide advises that “Only cite Australian legislation in your answer” beats “Do not cite US legislation”; Anthropic’s rule is “Tell Claude what to do instead of what not to do.” A prohibition says what to avoid, not what to do instead, so the model improvises.

Iteration. SAL’s DO list includes “Start a new chat for each task”; its DON’T list, “Ask the generative AI system to do too many things at one go”. Justia: “Treat your chat sessions like individual case files.”

The piece lawyers skip: jurisdiction

Clio’s guide contains the single most useful sentence in the genre: “Jurisdiction is the piece most lawyers skip because it’s obvious to them. You’ve been thinking about Texas employment law all morning. The AI hasn’t.” Its gut check follows: “if the same prompt could apply to a range of different matters without changing a word, it’s too broad.”

Apply the test to the prompt most lawyers type first.

Analyse the [non-compete / limitation of liability / indemnity] in Section [X] of the attached agreement under [jurisdiction] law as of [today's date].

Facts: [party, role, location, length of employment or relationship, what they have access to]. The clause: [duration, territory, scope, consideration].

Identify the controlling test, apply each factor to these facts, and note any developments since [date] that change the analysis. Cite only [jurisdiction] authority; where none exists, write NO VERIFIABLE LOCAL AUTHORITY FOUND. Tag every authority [VERIFY]. Finish with the three facts I have not given you that would most change your answer.

The last line matters. Brooke Loesby’s “Curiosity Prompt” in the ABA Journal is the one-line version: “Ask me what else you need to know to give me the most accurate response.” It works, she writes, “very much like a supervising attorney questioning a junior associate”, and “Because the model interviews the user, the attorney remains the source of all material facts.” The curiosity prompt guide has worked examples.

Brief it like a junior associate (and why the intern analogy is retiring)

Every guide reaches for the same comparison. Catherine Sanders Reach at the NC Bar: “In the same way you would not ask a first-year associate to generate a pretrial motion without significant guidance and instruction, you must guide the GAI to get the best output.”

The analogy is right about briefing and wrong about memory. Ethan Mollick’s refinement is the one to keep: “Treat AI like an infinitely patient new coworker who forgets everything you tell them each new conversation, one that comes highly recommended but whose actual abilities are not that clear.” It forgets, so every prompt carries the full brief or points to a stored one (a Claude Project, a custom GPT, a playbook file; see Claude Projects and custom GPTs for law firms). It is infinitely patient, so iteration is free. And its abilities are unclear, so you test it on work where you know the answer before trusting it on work where you do not.

Show, don’t tell: few-shot examples that flipped a wrong answer

The best documented effect of a single prompting technique in legal work comes from Thomson Reuters’ CoCounsel team. They asked GPT-4 whether a clause restricts a party from contesting IP ownership. The clause was a covenant not to sue: “RemainCo hereby covenants not to sue SpinCo under any Licensed RemainCo Know-How…”. The model answered “No”. Wrong: a covenant not to sue restricts without using the word “contest”.

They then added labelled examples, three “Yes” clauses such as “Company agrees that it will not at any time contest the ownership or validity of any Reed’s Intellectual Property” and one “No” clause about audit rights, and GPT-4 answered correctly. The write-up is on Thomson Reuters’ blog. Anthropic’s general form: “Include 3–5 examples for best results”, relevant, diverse and wrapped in tags; “You can also ask Claude to evaluate your examples for relevance and diversity, or to generate additional ones.” Examples teach a boundary the way a supervising lawyer would, by showing the cases on either side of it.

I am classifying clauses as RESTRICTS (a party may not contest the ownership or validity of the other side's IP) or PERMITS.

<example label="RESTRICTS">Company agrees that it will not at any time contest the ownership or validity of any of Licensor's Intellectual Property.</example>
<example label="RESTRICTS">Licensee covenants not to sue Licensor for any claim that the Licensed Know-How infringes Licensee's rights.</example>
<example label="RESTRICTS">Distributor acknowledges Supplier's exclusive ownership of the Marks and shall not challenge, or assist any third party in challenging, that ownership.</example>
<example label="PERMITS">Either party may audit the other's books and records on 30 days' notice.</example>

Classify the following clause and quote the words that drive the classification:
<clause>[paste clause]</clause>

Vals’ redlining benchmark reported the same lesson: “Both AI tools performed much better when clauses were provided as clearly labeled plain text.” Labels and examples are cheap; use them whenever the boundary is not obvious from the instruction.

Confidentiality as a prompting technique: placeholders and anonymised facts

The most important prompting decision is made before the first word is typed: what goes into the box. Loesby’s example is the cleanest statement of the technique. Not: “My client Sarah is suing her business partner for embezzling $400,000.” But: “I am working on a partnership dispute involving allegations of financial misconduct.” The legal analysis is identical. The client is not in the prompt.

Harvey’s templates use “[Client]”, “[Counterparty]”, “[Witness A]”; Sterling Miller’s in-house prompts find-and-replace the company name with “Big Co.”. Swap names for consistent placeholders, keep a key table offline, re-hydrate on output, and strip metadata first. Why this matters, including the February 2026 ruling that a client’s consumer-Claude chats were not privileged, is in is ChatGPT confidential for lawyers?.

Context: I act for [PARTY_A], a [mid-market manufacturer] in a dispute with [PARTY_B], its former [distributor], under [jurisdiction] law. The claim concerns [alleged breach of exclusivity and unpaid commissions of roughly [AMOUNT_1]]. Key dates: [DATE_1] contract signed, [DATE_2] termination notice, [DATE_3] proceedings threatened.

Task: list the elements [PARTY_A] must prove, the facts here that support and cut against each, and the documents I should obtain. Do not cite cases. Cite statutes only where confident, tagged [VERIFY]. Do not guess at any placeholder.

Verification instructions: what to add and, for reasoning models, what to remove

Here the genre has gone out of date. The verification lines that measurably helped older models now hurt on some newer ones.

The evidence for adding them is real: the CoCounsel team showed “Before you answer, think through your reasoning step-by-step” turning a wrong LSAT-style answer into a right one, and “Double-check your answer and fix any problems you find” correcting GPT-4. Anthropic’s general advice agrees: “Append something like ‘Before you finish, verify your answer against [test criteria].’ This catches errors reliably”.

Then the exception. Anthropic’s guide to Claude Opus 5 says the model “verifies its own work without being told to. If your prompt contains explicit verification instructions… remove them: instructions like these cause over-verification”, naming “double-check your answer” and “re-verify before responding”. OpenAI’s reasoning best practices say of its o-series models: “Since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.” Instead: “Keep prompts simple and direct”, “Use delimiters for clarity”, “Be very specific about your end goal.”

The last row never changes. Whatever the model, an instruction to tag authorities and to abstain rather than invent costs nothing and gives you a checklist; Anthropic’s troubleshooting entry for “AI makes up information” is simply “Explicitly give permission to say ‘I don’t know’ when uncertain.” And no verification line replaces verification. SAL’s guide puts it in first-year terms, “do not rely on an authority that you have not read”, and sums up the approach as “Copilot, not autopilot.” Prompting reasoning models for legal work has the full decision table, including why reasoning models can hallucinate more, not less.

Goal: a complete list of every provision in this MSA that allocates risk away from my client, the customer, so nothing is missed at tomorrow's negotiation call. Completeness matters more than brevity: I would rather see twenty candidate issues with three false positives than ten with one omission.

For each provision: quote the operative words, give the clause reference, and say in one line why it shifts risk. Finish with the three you would raise first. Use only the attached agreement; cite no case or statute.

<agreement>
[paste]
</agreement>

Format instructions that produce reviewable work product

A format instruction is a verification instruction in disguise: name the columns and every row becomes checkable. SAL’s witness-credibility prompt asks for “Item Number; Affidavit (Para/Section); Transcript (Page/Line Number); Description of inconsistency; Impact on Credibility”. Harvey’s vendor-agreement review asks for “clause reference, issue, business impact, suggested revision, and priority level”.

Two mechanical points. Anthropic says to put long documents at the top and the question at the end: “Queries at the end can improve response quality by up to 30 percent in tests.” Thomson Reuters warns of “primacy and recency bias”: “An LLM is more likely to forget or fail to consider information contained in the middle of a prompt.” Justia’s “Boundary Walls” tag the source, the standard and the template so the model does not mistake “The licensee shall notify…” for an instruction to itself. Prompting long contracts covers the 80-page document.

Extract the following from the attached lease into a table with exactly these columns: Field | Extracted value | Clause reference (section and page) | Confidence (High/Medium/Low) | Note.

Fields: Tenant; Landlord; Premises; Commencement date; Expiry date; Renewal options (number, length, notice window); Base rent and escalation; Security deposit; Permitted use; Assignment and subletting; Break rights.

If a field is not in the lease, write NOT FOUND rather than inferring it. After the table, list every amendment or side letter the lease refers to.

<lease>
[paste]
</lease>

The reusable template (copy it)

Everything above collapses into one block. Save it as a text snippet, fill the brackets once per matter type, and keep the safety lines at the top: “Lead your custom instructions with the safety rules, not your bio”, as the claudeforlawyers.com Project workflow puts it.

WORKING RULES
1. Jurisdiction: [jurisdiction]. Apply only [jurisdiction] law. Use [British/American] spelling.
2. Use only the materials I provide. Tag every citation from your own knowledge [VERIFY]. If you cannot support a proposition, write NO VERIFIABLE AUTHORITY FOUND. Do not extrapolate from other jurisdictions.
3. Never invent facts, dates, amounts or names. Leave a [BRACKET] where information is missing.
4. Everything you produce is a draft for review by a licensed lawyer.

PERSPECTIVE: I act for [client role, e.g. the customer / the seller / the respondent]. The reader is [a partner deciding what to fight for / a client with no legal training / the court].
CONTEXT: [matter type, anonymised facts, parties as [PARTY_A]/[PARTY_B], posture, commercial pressure, deadline].
TASK: [one precise verb and object].
SOURCES: <materials>[paste or attach]</materials>
FORMAT: [table with named columns / memo with these headings / email under 200 words].
CONSTRAINTS: [length, tone, what to include, stated affirmatively].
ITERATION: Ask me up to [3] clarifying questions if the request could be read in materially different ways; otherwise make routine judgement calls and list them at the end.

Keep it short. Anthropic’s corrective for anyone tempted to add a page of instructions: “The best prompt isn’t the longest or most complex. It’s the one that achieves your goals reliably with the minimum necessary structure.” For task-specific versions, the ChatGPT and Claude prompts for lawyers collection and the prompt library have forty-plus ready to paste.

Prompting is changing: matter-aware tools, agents and skills

Now the contrarian point, made by the people selling prompting guides. Catherine Sanders Reach wrote in 2024 that “Reliance on complex prompting will be diminished, just like the need for Boolean search was reduced by natural language searches.” Clio reports that 84% of queries in Clio Work are already freeform goal statements. Vendors say their tools are becoming matter-aware: they know the jurisdiction, the documents and the playbook before you type.

Both are true; the brief has moved, not disappeared. Anthropic’s open-source Claude for Legal plugins start every practice area with a cold-start interview that writes a practice profile, and the README warns that “Skipping setup is the single most common reason a skill produces generic output.” Zack Shapiro, whose February 2026 post on running a two-person firm on Claude drew millions of views, put it most directly: “The entire gap between ‘AI is a toy’ and ‘AI changed my practice’ lives in the quality of your instructions.”

So the seven parts survive: typed into a chat today, stored in a Project or a skill tomorrow, handed to an agent next year. Prompt engineering for lawyers is instructing a capable, forgetful colleague precisely enough that the work comes back reviewable. The adversarial half of the skill, making the model attack its own draft, is in adversarial prompting and self-critique.

The frameworks fit on one slide. The skill is in the iteration: write the brief, read the output as the most sceptical partner in the firm, fix the brief, run it again. That loop is what the four live two-hour sessions of AI Lab for Lawyers are built around, prompt engineering being one of the six modules.

Where to go next: the prompting guides cover each technique in depth, and the twelve prompting mistakes shows, from court records, what the template above prevents.