Everything You Need to Know About Prompting AI You Learned in Law School

Everything I really need to know about how to prompt AI, I learned in law school.

The Socratic method that gave you anxiety attacks in Contracts. The IRAC structure you thought was just a straitjacket for your legal writing assignments. The issue-spotting instinct that made 1L exams feel like a fever dream. All of it turns out to be training for the skill the rest of the world is now calling "prompt engineering." You just did not know it yet.

But you do not have to take my word for it. Recent research from Wharton, Anthropic, and NeurIPS has quantified which prompting techniques actually work. Every single one maps to something you learned before you passed the bar.

Here it is.

1. Say precisely what you mean. (Contracts drafting)
2. Tell the reader what you want before you tell them what you know. (IRAC)
3. Curate ruthlessly. Include only what matters. (Brief writing)
4. Show, don't just tell. (Precedent and analogical reasoning)
5. Never accept the first answer. (The Socratic method)
6. Ask the specific question, not the general one. (Issue spotting)
7. Change one fact and see what breaks. (Hypo-shifting)
8. Argue the other side. (Moot court)
9. "You are a legal expert" is worth exactly nothing. (The anti-lesson)

Say Precisely What You Mean

Every contracts professor who ever circled the word "reasonable" in red ink and wrote "reasonable to whom, under what standard, measured when?" in the margin was training this skill. Legal drafting is the art of eliminating ambiguity. You learned to replace "promptly" with "within five business days," to specify governing law instead of leaving it implied, to define every capitalized term because you understood that undefined words invite disputes.

That instinct is the single most valuable skill in AI prompting.

Wharton's Prompting Science research and Anthropic's own documentation on Claude both converge on the same finding: specificity is the highest-impact prompting technique. The quality of the task specification itself matters more than any other variable. Not the model. Not the prompt template. Not the persona you assign. The precision of your instruction.

Consider the difference between these two prompts:

Vague: "Review this NDA."

Precise: "Identify all non-compete, non-solicitation, and non-disclosure obligations in this NDA. For each, specify the restricted activity, geographic scope, duration, and any carve-outs. Flag any provision that would survive a change of control."

The first prompt is the contractual equivalent of "best efforts." It sounds like something, but it means almost nothing. The second prompt does what every good contract does: it defines the scope of the obligation, specifies the deliverable format, and identifies the trigger condition.

LLMs narrow the probability space of their output based on input constraints. A more precise prompt constrains the output in productive ways, the same way a well-drafted definition section constrains how terms are interpreted throughout an agreement. Vague prompts get vague answers for the same reason vague contract terms invite disputes: they leave too much room for interpretation.

Your contracts professor was not pedantic. Your contracts professor was teaching you prompt engineering twenty years before prompt engineering existed.

Tell the Reader What You Want Before You Tell Them What You Know

IRAC was the first legal framework most of us ever learned, and most of us resented it. Issue, Rule, Application, Conclusion felt like a straitjacket. Why can't I just write the analysis the way it flows in my head?

Because structure is not a constraint on thinking. It is a tool for communication. And it turns out that what researchers call "structured framework prompting" is exactly what IRAC teaches: decompose the problem, provide the governing standard, point to the facts, and specify the output you need.

Here is IRAC as a prompt framework, mapped one to one:

• Issue = State the task. "Analyze the indemnification provisions in the Stock Purchase Agreement between Buyer Corp and Target Inc."
• Rule = Provide the standard or criteria. "Identify cap amounts, basket mechanisms, survival periods, and carve-outs for fundamental representations and fraud."
• Application = Point to specific facts or documents. "Review Sections 7.1 through 7.5 of the attached SPA."
• Conclusion = Define the output format. "Present findings in a table with columns for provision type, specific terms, and page citation."

That four-part structure gives an AI the same context you would give a junior associate when delegating a research assignment. It answers the questions that every associate asks (and every AI needs answered): What am I looking for? What standard am I applying? Where should I look? How should I present what I find?

For complex tasks, break IRAC into sequential prompts. First: "Identify all indemnification provisions and their section numbers." Then: "For each provision you identified, extract the cap, basket, survival period, and carve-outs." Then: "Flag any terms that deviate from market standard for a middle-market acquisition."

This mirrors how you would actually delegate to a junior associate. One step at a time, checking each deliverable before moving on. You would never hand a first-year a 200-page SPA and say "tell me everything." You would break the assignment into discrete tasks. AI works the same way.

The straitjacket, it turns out, was training wheels for structured thinking. You can take them off now. But the structure stays.

Curate Ruthlessly. Include Only What Matters.

Brief writing taught you one of the hardest skills in professional communication: the discipline of exclusion. Page limits forced you to decide what mattered and what did not. Every sentence had to earn its place. You learned that including a weak argument does not add to your case. It dilutes your strongest points.

This discipline maps directly to how you should provide context to AI.

Research on context quality shows that LLM performance degrades as input length grows. A Stanford study published in TACL documented what researchers call the "lost in the middle" effect: models attend more carefully to the beginning and end of a prompt, with information in the middle receiving significantly less weight. Chroma's "context rot" research found that performance drops noticeably as input tokens increase, becoming increasingly unreliable in longer contexts.

You already know this intuitively. You do not dump every case you found into a brief. You select the most relevant authorities, organize them strategically (strongest arguments first and last, weaker ones in the middle), and present a curated body of evidence. Brief page limits taught you this discipline. Federal rules did not give you 50 pages because they thought you needed all 50. They gave you a ceiling, and the best brief writers never came close to it.

Apply the same discipline to AI prompts. When analyzing a 50-page agreement, do not paste the whole thing. Paste the specific sections relevant to your question. If you need to analyze the full document, break it into sections and analyze each separately. Put the most critical context first and last. Treat every word in your prompt the way you treat every word in a reply brief: if it does not advance the analysis, cut it.

Show, Don't Just Tell

Precedent is not just legal authority. It is a communication tool. When you cite a case, you are not merely invoking a rule. You are showing the reader an example of how a court applied that rule to facts, which teaches the reader how to apply the same rule to your facts. The entire common law system is built on learning by example.

In AI prompting, this technique is called "few-shot prompting," and it is one of the most effective approaches available. Anthropic recommends providing a few diverse, high-quality examples. But recent research reveals a nuance that any lawyer would recognize: quality matters more than quantity. Two or three well-chosen examples are highly effective. Beyond that, you hit diminishing returns. And poorly chosen examples actively degrade performance.

This is precedent selection. You cite cases not because they exist but because they are on point. You choose the case from the same jurisdiction, with analogous facts, decided under the same legal standard. A dozen peripheral cases do not help your argument. Two strong, on-point precedents do.

Apply the same instinct to AI. If you want the model to extract provisions in a specific format, show it one good example of the format you want before asking it to do 50 more. If you want risk assessments structured a certain way, provide a sample output from a previous deal. Just like in a brief, a few strong examples beat a dozen weak ones.

The common law trained you to teach by showing. AI learns the same way.

Never Accept the First Answer

This is the heart of it.

Picture the moment. First semester, Contracts or Torts or Civil Procedure. The professor scans the room. The silence is excruciating. Then: your name.

You give your answer. You think it is pretty good. The professor does not nod. Does not smile. Instead: "But what about...?" "Is that always true?" "What happens when...?"

Your heart rate spikes. You scramble. The professor is not being cruel (probably). The professor is doing something very specific: asking targeted follow-up questions that expose the gaps in your reasoning. Not "try again." Not "be more precise." But: "What about the exception for fraud?" "Does that analysis change if the jurisdiction is Delaware?" "You said the statute requires notice, but notice to whom?"

That visceral, anxiety-producing process is exactly what makes multi-turn AI prompting effective.

A NeurIPS 2025 study found that targeted, specific follow-up reliably improves AI output quality. But here is the critical finding: vague feedback ("make it better," "try again," "be more thorough") causes quality to plateau or actually reverse. The AI does not know what "better" means any more than you knew what to do when a professor just stared at you in disappointed silence.

The difference between effective and ineffective iteration is precision. Professors did not say "try again." They said "What about the exception for fraud?" That specificity is what makes the Socratic method work, and it is what makes multi-turn prompting work.

Walk through a three-turn example:

Turn 1: "Identify all change-of-control provisions in this customer agreement."

The AI produces a competent but surface-level answer: the agreement contains an anti-assignment clause in Section 9.1 that restricts assignment without consent.

Turn 2: "What about the permitted transfer carve-out in Section 3.2(b)? Does that create a gap in the buyer's protection?"

Now the AI engages more deeply. It identifies that Section 3.2(b) allows transfers to affiliates without consent, which could permit a post-closing restructuring that effectively circumvents the anti-assignment protection.

Turn 3: "How does that interact with the anti-assignment clause in Section 9.1? Are those provisions consistent, or does the affiliate transfer carve-out create an internal conflict?"

The AI's analysis transforms. It identifies the tension between the two provisions, notes that the definition of "affiliate" in Section 1.1 is broad enough to create a genuine gap, and flags this as a point for negotiation.

The output improved across each turn. Not because you said "be better," but because you asked the precise follow-up question that exposed the gap. Just like the professor did. Just like the professor always did.

Your heart rate might still spike a little when someone says your name in a quiet room. But the skill that caused the spike is the skill that makes you good at this.

Ask the Specific Question, Not the General One

The real skill gap with AI is not about how you ask. It is about what you ask. This is where legal training creates the widest advantage.

A non-lawyer looking at a stock purchase agreement asks: "Summarize this agreement." A first-year associate asks about the representations and warranties. A senior M&A attorney asks about the change-of-control triggers in the customer agreements, the consent requirements that could delay closing, the anti-assignment provisions that might impair the value of the acquired book of business, the indemnification caps relative to enterprise value, and the survival periods on fundamental reps versus general reps.

The difference is not sophistication of language. It is issue spotting. The senior attorney knows what to look for because they have seen what goes wrong. They know that the most consequential provision in a data room is often the one nobody thought to ask about.

Think back to the exam room. Three hours. A fact pattern dense with issues. Your grade depended not on how well you analyzed the issues you spotted, but on how many issues you spotted in the first place. That timed-exam instinct, the ability to scan a fact pattern and identify what matters, is precisely what AI lacks and what you provide.

AI does not know what questions to ask itself. It will answer whatever you ask, competently, confidently, and sometimes incorrectly. But it will not tell you that you asked the wrong question. That is your job. And it is a job that three years of law school and years of practice have trained you to do better than almost anyone.

Change One Fact and See What Breaks

Every law professor has a version of this move. You give your answer. It seems solid. Then: "Now assume the buyer is a competitor." "What if the closing condition fails?" "Same facts, but the governing law is California instead of Delaware."

One fact changes. Your entire analysis might collapse. Or it might hold. Either way, you learn something.

This technique, hypo-shifting, is one of the most powerful ways to test whether an AI is actually reasoning about your problem or just pattern-matching against its training data.

Here is how to apply it. Ask the AI to analyze a non-compete provision under Delaware law. Get the analysis. Then change one variable: "Now assume the governing law is California." If the AI's answer does not change meaningfully (California is famously hostile to non-competes), the AI was not analyzing. It was generating plausible-sounding text based on patterns. If the answer does change, and changes in the right ways, you have evidence that the analysis is substantive.

Change one variable at a time. Jurisdiction. Dollar threshold. Time period. Party identity. Each shift should produce a different analysis if the model is actually engaging with the substance. An indemnification cap analysis should change when you shift the deal size from $50 million to $500 million. A non-solicitation analysis should change when you shift the restricted party from employees to customers. A termination analysis should change when you shift from a convenience right to a cause-only right.

If the analysis stays the same when the facts change, you are not getting analysis. You are getting a template. And you know the difference, because your professors spent three years training you to spot it.

Argue the Other Side

Moot court was exhausting. You prepare your argument. It is airtight. Then they tell you to argue the other side. Suddenly, all the gaps you did not see become obvious.

This adversarial instinct translates directly to AI. After getting an analysis you are satisfied with, try this: "What would opposing counsel say about this analysis?" or "What are the three strongest counterarguments to the conclusion you just reached?"

Research on self-critique in AI confirms that this technique produces measurable improvements, but only when the critique is directed against specific criteria. "What is wrong with this?" is too vague. "What would a seller's counsel argue about the enforceability of this non-compete under California law?" gives the model a specific adversarial lens.

In M&A diligence, this is especially powerful for risk assessment. After identifying an issue, prompt the AI to argue why it might not be as significant as it appears. "This customer contract has a termination for convenience clause with 30 days notice. Argue that this is not a material risk to the buyer." The resulting analysis often surfaces mitigating factors (notice period requirements, cure provisions, termination payments) that a single-perspective analysis misses.

You learned in moot court that you do not truly understand your own position until you can argue against it. The same principle applies to AI output. The first answer is your opening brief. The adversarial follow-up is the reply.

"You Are a Legal Expert" Is Worth Exactly Nothing

Here is the counterintuitive kicker. The single most popular prompting technique on the internet is persona setting. "You are a senior M&A attorney at a top-10 law firm with 20 years of experience." Every prompting guide recommends it. Every AI tutorial starts with it.

It does not work.

Wharton tested persona prompting rigorously in 2025 across six leading language models. The finding was unambiguous: expert personas produced performance "statistically indistinguishable from the baseline." Telling an AI it is an expert does not make it smarter, more accurate, or more thorough. It is the prompting equivalent of telling a first-year associate to "think like a partner." It sounds motivating. It changes nothing about the quality of their work product.

Why does this matter? Because if the most popular prompting technique does not work, then the techniques that do work become even more important. And those techniques are: specificity. Structure. Strategic context. Targeted follow-up. Adversarial thinking.

Look at that list. Specificity is contracts drafting. Structure is IRAC. Strategic context is brief writing. Targeted follow-up is the Socratic method. Adversarial thinking is moot court.

The research does not just suggest that lawyers can prompt AI effectively. It says that the techniques that actually improve AI output ARE the techniques lawyers already practice, and that the one technique everyone else relies on is the one that does not work.

One caveat: persona setting can affect tone and register. "Respond in a formal legal memorandum style" will change how the output reads. It just will not change whether the analysis is correct. Tone is not substance. You learned that distinction in legal writing, too.

You Were Trained for This

So there it is. The nine things. Everything you need to know about prompting AI, you learned in law school.

The most effective AI users are not technologists. They are not prompt engineers with computer science degrees. They are rigorous thinkers who know how to specify precisely, structure analytically, curate context, press on weak answers, and argue both sides. They are people trained to never accept the first draft, to always ask "what about the exception," and to treat every word as if it carries legal consequence.

The rest of the world is paying for prompt engineering courses and buying prompting playbooks and attending webinars on "how to talk to AI." You already took that course. It was called law school. It just came with a heavier reading load, worse coffee, and a Socratic method that still makes your heart rate spike when someone says your name in a quiet room.

The syntax is new. The thinking is not.

---