AI tools can draft a case brief, a statute summary, or a client email in seconds. What they do not do by default is produce the same structure twice. Ask the same question on Monday and Thursday and you may get a tidy numbered brief one day and a wall of text the next.
This guide is about that single problem: getting predictable formatting and staying inside the material you provided. It gives you five prompt patterns, three copy-ready templates, and a simple method for testing any prompt on your own tool and documents before you rely on it.
Please read first: This article is about formatting and workflow. It is not legal advice, and I am not a lawyer. AI output can contain invented citations, misquoted statutes, and wrong jurisdictions even when it looks perfectly formatted. Anything that goes into a filing, a client communication, or a legal decision must be checked against the primary source by a qualified person.
Why formatting and scope matter in legal work
Imagine three people at the same small firm, each with their own saved prompt for “summarize this case.” One gets a numbered list, one gets a single block of text, and one gets something the length of a law review note when a two-paragraph email was all that was needed. (This is a hypothetical scenario, used to illustrate the problem.)
None of those outputs is necessarily wrong. The trouble is that nobody can compare them, review them quickly, or hand them to a colleague without reformatting. In legal work, where numbering, citations, and exact wording matter, inconsistent structure also makes errors harder to spot.
Two ideas explain most of the fix:
Format control means the output follows the same structure each time: the same headings, in the same order, with the same citation style and level of detail.
Context control means the model works only from the material you gave it. It does not add cases you did not provide, invent statutory text, or mix in another jurisdiction unless you asked for a comparison.
A prompt can fail at either one. A clean-looking brief that quietly cites a case you never supplied has good format control and poor context control, which is the more dangerous combination.
Five common legal research prompt types
Each type below has a typical way of going wrong. These are design observations based on how language models generally respond to vague instructions, not measured results.
1. Case briefs
Goal: turn an opinion or case summary into a structured brief.
Typical weakness: a request like “summarize this case as a brief” leaves the model to decide what counts as a fact, what counts as reasoning, and how many sections to use.
Fix: list the exact section names, in order, and say that no sections may be added, merged, or renamed.
2. Statute and regulation summaries
Goal: explain a statute in plain language while keeping the subsection numbering exact.
Typical weakness: with a long, multi-part statute, a single-step request makes it easy for subsections to be skipped, merged, or paraphrased without their numbers.
Fix: split the task in two. First list every subsection by its exact designation with no commentary. Then summarize each one separately.
3. Contract clause analysis
Goal: identify risks, ambiguities, and possible revisions in a clause.
Typical weakness: without a boundary, the model may comment on terms that were not in the text you pasted or add general contract-law commentary.
Fix: add one clear scope line: analyze only the clause provided, and do not refer to other terms or outside standards unless asked.
4. Memos and client letters
Goal: turn research notes into something a partner or client can read.
Typical weakness: the tone may be fine while the structure varies, so the recommendation can end up buried at the bottom.
Fix: give labeled sections in a fixed order, with the recommendation near the top.
5. Multi-jurisdiction comparisons
Goal: compare how two or three jurisdictions treat the same issue.
Typical weakness: a narrative comparison invites blending, where a rule from one jurisdiction is attributed to another, especially when the standards are similar but not identical.
Fix: require a table with one column per jurisdiction and identical row categories, and supply the legal text for each jurisdiction yourself rather than relying on the model’s memory.
Three techniques that carry across all five
Name the sections instead of describing them. “Write this in a clear, structured format” leaves everything open. “Use exactly these five headings in this order” does not.
Use two steps for complex material. When the source has many parts, such as a statute with many subsections or a multi-party contract, have the model first confirm the structure, then fill it in. The first step lets you catch a missed subsection before it becomes a wrong summary.
State what the model must not do. Lines such as “do not cite any case that is not in the text provided” or “do not compare to other states unless asked” give the model a boundary, not just a goal.
One caution: these techniques reduce variation, but they do not eliminate it. A model can still ignore an instruction, and a prompt that behaves well on one tool or version may behave differently on another. That is why the testing method later in this article matters.
Template 1: Case brief with a fixed structure
This template uses five sections. If your firm or course uses IRAC (Issue, Rule, Application, Conclusion), swap the headings, but keep the same rule: fixed names, fixed order, nothing added.
Why it is built this way: the section list is explicit, the source is fenced in, and the “Possible gaps” line gives the model a place to admit missing information instead of filling it in.
Template 2: Statute summary in two steps
Run Step 1 on its own. Check the list against the statute yourself. Only then run Step 2.
Step 1: Structure check
Why it is built this way: the confirmation step turns a hidden failure (a quietly skipped subsection) into something you can see and check.
Template 3: Contract clause review with a scope line
Template 4: Memo skeleton
Template 5: Multi-jurisdiction comparison table
A caution: asking a model to compare jurisdictions from memory is one of the riskier uses of these tools. Supplying the text for each jurisdiction yourself, as above, is safer, and the result still needs to be checked against the official sources.
How to test any of these prompts yourself
A prompt that looks good on its first run proves very little. Here is a simple way to find out whether a prompt actually holds.
- Pick a source you already know well, such as a case you have read or a statute you can check line by line.
- Run the same prompt three times in fresh sessions, with the same source text.
- Score formatting consistency. Full pass: same headings, same order, similar level of detail in all three runs. Partial: two of three match. Fail: none match.
- Score context control. Pass only if the output uses nothing beyond the pasted material. Any outside case, invented quotation, or unrequested commentary is a fail.
- Check the substance against the source, line by line: numbering, quotations, party names, dates, and holdings.
- Change one thing at a time (add a boundary line, rename a section, split into two steps) and run the three-session test again.
- Retest after a model update, since the same prompt can behave differently on a new version.
A small test like this is not a scientific study, and results from one tool, one document type, and one day do not transfer automatically to another. Record what you tried and what you saw, so your notes are based on your own results.
Verification checklist before anything leaves your desk
- Every case, statute, and quotation checked against an official or authoritative database.
- Citations confirmed to exist and to say what the output claims.
- Subsection numbers, dates, and party names compared with the source.
- Jurisdiction confirmed, and no rule carried over from another one.
- Current law confirmed, since a pasted text may be outdated or amended.
- A qualified person has reviewed anything intended for a client, court, or decision.
Courts have sanctioned lawyers for submitting AI-generated citations to cases that did not exist. Good formatting is not evidence of accuracy.
Common mistakes when writing legal research prompts
Describing the format instead of specifying it. “Make it look like a memo” invites interpretation. A named list of sections does not.
Trusting a single good run. One clean output says nothing about the next one.
Writing longer prompts to fix structure problems. Extra paragraphs about tone and style can bury the structural instructions. Often the better fix is a shorter prompt that is exact about sections and boundaries.
Adding background “for context” in comparison tasks. More material gives the model more to cross-reference. Supply only what the comparison needs.
Relying on the prompt to do the checking. Telling a model to “verify” its own output is not a substitute for you verifying it.
Final thoughts
The most useful change you can make is to write legal research prompts like forms rather than questions: name the sections, fence in the source, say what the model must not do, and split complex tasks into a confirmation step and a content step. None of that replaces legal judgment or verification. What it does is make the output easier to review, compare, and correct.
Take one of the templates above, run it three times on a document you know well, and see what holds. Your own results will tell you more than any general claim, including the ones in this article.
FAQ’s
Does prompt length matter more than clarity?
Clarity about the output format is generally more useful than length. A short prompt with named sections and a clear boundary gives the model less room to improvise than a long prompt with vague style instructions. Test both on your own material to see which holds up for your tool.
Can these approaches work for other document types?
The underlying ideas (named sections, two steps for complex material, explicit boundaries) are general and can be adapted to other kinds of structured writing. For medical, financial, or other high-stakes documents, the content needs review by a qualified professional in that field, and a template alone is not enough.
Should AI-generated legal research ever be used without human review?
No. Formatting consistency is not the same as legal accuracy. Anything used in a filing, client communication, or legal decision should be verified against the primary sources by a qualified person.
How often should I retest my prompts?
Retest after any significant model or tool update, and whenever you change the type of document you feed in. A prompt that worked on one version may behave differently on another.
Rehan Malik is a blogger and the creator of PromptByJob. He publishes articles and prompt examples about using AI tools for different tasks. He uses AI-assisted drafting in his writing process. He is not a healthcare professional, and his nursing-related content should not be treated as medical advice or evidence of clinical testing.
