The Attorney's Guide to AI Prompting

Why "summarize this" fails on case materials

AI output can be fluent and wrong in ways that take substantial time to catch. The model, source quality, product configuration, and instruction can all contribute. No partner hands an associate a 400-page deposition and says "summarize this." They say: "Two-page memo on every opinion Dr. Smith offered on causation, with page-line cites, flag anything inconsistent with his Rule 26 report, on my desk Thursday for the MSJ." The difference between those two instructions illustrates why a defined scope and deliverable matter, whether the reader is a human or a machine.

A general-purpose model has no idea which side you are on, what the document is for, or what “important” means in your matter. Asked only to summarize, it may produce a broad abstract that misses material issues. It may also fill gaps with plausible invention. Better instructions do not remove those risks, but they give the reviewer a clearer task, output, and source standard to test.

The CITE method: a prompt is an assignment memo

One useful control is to write the assignment memo you would give a team member, in four blocks. We call it CITE — Context, Instructions, Template, Evidence. Use all four when the task involves reviewing or drafting from a record.

• Context. Who you are, the matter, the document, and the posture. “You are assisting defense counsel in a trucking-liability case; the attached is the certified deposition of the plaintiff driver; the issues are speed at impact and phone use.” This states the perspective and issues the response should use.
• Instructions. The exact task, scoped and decomposed. Not "find the important parts" — that is a judgment call you have not given criteria for. Instead: "Extract every admission, every agreement with examining counsel, and every claimed lapse of memory on the key issues; capture the full Q&A for each, not just the answer."
• Template. The requested output shape. “A table per issue with columns: # · type · verbatim Q&A · cite · why it matters.” A consistent skeleton can make entries and omissions easier to inspect; it does not make the output filing-ready.
• Evidence. The requested source-location rule. “For every finding, cite page and line and quote the testimony verbatim; if support is not located, write ‘not found in the reviewed record’; do not infer.” The model can ignore or mishandle this instruction, so open every cited location and check the result.

That is the whole method. For a fuller walkthrough with before/after examples, see the CITE method page; to generate a finished prompt without writing one, use the builder.

The six record tasks

The CasePrompts library organizes record work into six recurring jobs. Naming the task is the first decision in any prompt, because it dictates the instructions and the output format. Every one of the 24 templates in the library is one of these, pre-briefed:

Extract key testimony

Request specific items from a long record—such as admissions or stated opinions—with the quoted language and a source location for each.

Build a chronology

Turn a pile of dated material into a single timeline, every entry attributed to its source and ordered by date.

Find contradictions

Identify passages that may conflict—such as testimony against a prior statement—so a reviewer can compare both sources.

Summarize the record

Compress a transcript or a file into a structured, cited summary built around the issues that matter, not a generic abstract.

Draft from the record

Request a first draft from identified record material—a fact section, a demand, or discovery requests—with source locations for factual assertions.

Interrogate the file

Set standing source and output rules, then ask recurring questions while continuing to verify each answer against the file.

Prompting by document type

The document type determines which source coordinate a prompt should request. A usable pointer is the purpose of the Evidence block, and the right form depends on what you are reviewing. These are the conventions CasePrompts uses, by document type:

For step-by-step source inventory and verification guidance, choose a record-type workflow.

The principle generalizes: never accept a finding the tool cannot point you to. If the model cannot produce the cite, treat the finding as unverified — which, until you check it, it is.

Prompting by practice area

Two lawyers can read the identical production and be hunting for opposite things. A PI plaintiff's lawyer reads the medical records for causation and permanency; insurance defense reads the same pages for gaps, prior claims, and pre-existing conditions. That difference belongs in the Context and Instructions blocks — and it is what the practice lenses encode. CasePrompts ships 12 of them:

• Civil / insurance defense
• Commercial litigation
• Criminal defense
• Employment litigation
• Family law
• Medical malpractice — defense
• Medical malpractice — plaintiff
• Personal injury — plaintiff
• Probate & estate litigation
• Prosecution
• Workers' compensation — claimant
• Workers' compensation — employer/carrier

Pick your practice on the by-practice pages and every template assembles with your practice priorities already composed in.

Choosing and using an AI tool

Prompt structure is only one part of tool choice. Products differ in source access, context limits, model behavior, administration, retention, training controls, and contract terms. Confirm the exact product and plan your organization approves before entering case material; product-family names alone do not establish the applicable data boundary.

The by-tool pages provide dated capability notes and official product sources. CasePrompts text is portable, but supported features and output quality are not interchangeable across products.

The verification duty

Treat AI output as material requiring review, not as validated work product. Check every material statement and citation against the complete record before relying on it, filing it, or sending it. The Evidence block requests source locations to support that process; it does not prove that the locations or quotations are correct.

ABA Formal Opinion 512 and Florida Bar Opinion 24-1 discuss issues including competence, confidentiality, supervision, candor, and fees. They are examples, not a complete or jurisdiction-specific answer. Check the rules, court requirements, client obligations, ethics opinions, and firm policy that apply to your work. The free builder does not accept document uploads; optional beta AI tools send the prompt text you enter only when you choose to run them. Nothing here is legal advice.

• ABA Formal Opinion 512: Generative Artificial Intelligence Tools (opens in a new tab)
• Florida Bar Ethics Opinion 24-1 (opens in a new tab)

Frequently asked questions