The Curiosity Prompt for Lawyers: 'Ask Me What Else You Need to Know' (and Why It Works)

“The lawyers who are most frustrated with AI treat it like a search engine.” That is Brooke Loesby, a former BigLaw lawyer, writing in the ABA Journal in January 2026, and it explains most of the disappointment. A search engine returns what exists. A language model produces what is plausible, including the facts you did not give it. Type “draft a demand letter for a breach of contract” and you get a letter about a contract the model invented.

The curiosity prompt for lawyers is the fix, and it is not a longer prompt but a shorter one with a different verb: one sentence that turns the model from a search engine into a junior associate being briefed, so it asks before it drafts. You would not let a new associate guess the facts. You would let them ask.

The one line: “Ask me what else you need to know to give me the most accurate response”

That is the whole technique, verbatim from Loesby’s article. Her description of what happens next: “It functions very much like a supervising attorney questioning a junior associate before research or drafting begins.”

I need [a first draft of a settlement proposal / a summary of the attached lease / a plain-English explanation of the discovery timeline] for [audience] under [jurisdiction] law.
Before you produce anything, ask me what else you need to know to give me the most accurate response. Put all your questions in one message, most important first. I will answer, and then you start.

Loesby’s stress-test variant goes one step further: “Here is how I am thinking about positioning this argument. Ask me what else you need to know and then challenge my assumptions.” The second half turns the interview into a moot.

Why it works: the attorney stays the source of all material facts

Loesby’s own explanation is the professional-responsibility one: “Because the model interviews the user, the attorney remains the source of all material facts.” That, she argues, is why the prompt supports the “core duties of competence, diligence and supervision”. The lawyer supplies the facts; the model organises them.

The technical explanation is about gaps. A model given three lines of facts fills the other twenty from patterns in its training data, confidently. Stanford’s “Large Legal Fictions” study found that models “often uncritically accept users’ incorrect legal assumptions” (Dahl et al.); Damien Charlotin, who runs the hallucination database, is sharper: “The harder your legal argument is to make, the more the model will tend to hallucinate, because they will try to please you.” An interview surfaces the assumption, the missing party, the wrong jurisdiction before the model builds on it. It does nothing for invented authority (see why AI agrees with you); every citation still goes through a database.

Reverse Intake: “Do not write the document yet. Ask me up to 10 targeted questions”

Justia’s guide names the drafting version “Reverse Intake” and gives the template: “[State your high-level drafting goal, e.g., a dispute letter or contract clause]. Do not write the document yet. Ask me up to [Number, e.g., 5 to 10] targeted clarifying questions to gather the necessary factual context and jurisdictional nuances” (Justia Onward).

The words that do the work are “Do not write yet”. Without them, most models draft first and ask afterwards, and you are left editing a plausible letter built on invented facts.

I need a first draft of a letter before claim for a [commercial debt / breach of warranty] matter under [English law and the relevant pre-action protocol]. Audience: the counterparty's in-house counsel.
Do not write the document yet. Ask me up to eight targeted questions about the facts, the contract terms, the amount and how it is calculated, the counterparty's likely response, and the outcome my client actually wants. Group them by importance in one message, and tell me which documents you would want to see. After I answer, produce the draft.

When to cap the questions (and how OpenAI’s own guide conditions it)

An interview that never ends is its own failure mode. OpenAI’s GPT-5.1 prompting guide conditions the behaviour explicitly: when key information is missing, “pause and ask 1–3 brief clarifying questions”, but “For users who sound rushed or decisive, minimize questions and instead move ahead with defaults” (OpenAI). Anthropic’s scope-control prompt for Claude Opus 5 draws the same line from the other side: “Make routine judgment calls yourself, and check in only when different readings of the request would lead to materially different work” (Anthropic).

Two rules of thumb: ask for the questions in a single message, grouped by importance, so you answer once; and on a reasoning model skip “think step by step”, which OpenAI says is unnecessary for its o-series.

Five worked examples: demand letter, NDA, memo, client email, LinkedIn post

Suppose you run the curiosity prompt on five common tasks. The questions below are the kind a well-briefed model should come back with; they are illustrative, not transcripts. Each is a question you would want a junior to ask.

I am about to ask you to review an NDA we have received against our standard positions. Before reviewing, ask me what you need to know: which side we are on and whether the NDA is mutual; our accepted term; our position on residuals, non-solicitation and governing law; who must approve a deviation. One message. Once I answer, read the whole NDA and flag deviations, quoting each clause.

Meta-prompting: let the model write the prompt

The interview also works one level up: ask the model to interview you and then write the prompt. Nico Kuhlmann of Hogan Lovells gives the German version in LTO: “Du bist ein Experte für Prompt Engineering. Erstelle den optimalen Prompt, um einen Unternehmenskaufvertrag im Hinblick auf rechtliche Risiken zu analysieren.”

You are an expert in prompt engineering for legal work. I want a reusable prompt that produces [a supplier-side SaaS redline against our playbook / a client-ready case update email]. Interview me: ask the ten questions you need about my inputs, my audience, my quality bar, my house style and the failure modes I have seen. Then write the prompt with a role line, a context block with placeholders, an explicit output format, and constraints including a [VERIFY] rule for citations and a "do not invent facts" rule. Explain in five bullets why you structured it that way.

Test the result on three past matters where you know the right answer. OpenAI’s guide uses the same technique to debug prompts, finding contradictory rules such as “be concise” alongside “err on the side of completeness”. The wider patterns are in prompt engineering for lawyers; the ways this goes wrong are in legal prompting mistakes.

How the vendors productised it: cold-start interviews and playbooks

The idea is now built into products. Anthropic’s open-source legal plugins each begin with a “cold-start interview” that “takes 10–20 minutes per plugin” and writes a practice profile every later skill reads from; the README warns that “Skipping setup is the single most common reason a skill produces generic output” (GitHub). Immigration lawyer Greg Siskind makes clarifying questions his third tip: “Think of it like working with a junior associate or paralegal: give feedback, ask for revisions, and experiment with different approaches to improve the output” (Visalaw.ai).

You do not need a plugin. Ernie Svenson built his with one instruction: “I had Claude interview me about my background, my preferences and how I work. I told it to ask me 40 questions. It did” (Attorney at Work). The result became the master instruction that shapes everything Claude does for him; the step-by-step is in Claude Projects and custom GPTs for law firms, and the firms running whole practices this way are in AI-native law firms explained.

Where to go next: the 40-prompt library shows the curiosity line inside full research, review and drafting prompts; the prompting hub and the prompt library hold the rest. This is the first habit we install in AI Lab for Lawyers: by the second session, participants have stopped typing one-line prompts.