The CITE method
How attorneys should prompt AI: the CITE method
Most lawyers who try AI on case materials walk away unimpressed, and they are not wrong to. Ask a model only to “summarize this deposition” and the result may be a broad abstract that misses the questions your team cares about. It may also contain a confident citation to testimony that does not exist. The instruction is one part of that risk; model behavior and source quality matter too.
No partner hands an associate a 400-page transcript 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 report, by Thursday for the MSJ." That is a scoped assignment. The CITE method is that briefing pattern written down: Context, Instructions, Template, Evidence. It is a litigation AI prompt framework, not prompt engineering — four blocks, every time, in any tool.
C — Context
Context identifies the role, matter, source, posture, and issues in scope. Without that information, a general-purpose model has to infer what “important” means. Give it the assignment you would give a new team member: who the work is for, what document is being reviewed, and which questions matter. Context can focus the response, but it does not make the model understand the case or guarantee that it follows the stated perspective.
Weak prompt
Review this transcript.
CITE context
You are assisting defense counsel in a trucking-liability case. The attached document is the certified deposition transcript of the plaintiff driver. The issues that matter are vehicle speed at impact, cell-phone use before the collision, and seatbelt use.
I — Instructions
Instructions define the task and its boundaries. “Find the important parts” leaves the selection criteria unstated; a numbered list of concrete categories is easier to review. Break the work into checkable subtasks and say what is out of scope. The model can still omit or misclassify material, so treat its categories as instructions to test—not controls that enforce compliance.
Weak prompt
Pull out the key testimony.
CITE instructions
Extract every admission against interest, every agreement with examining counsel ("That’s fair," "Correct"), and every claimed lapse of memory ("I don’t recall") on the three key issues. Capture the full question-and-answer exchange for each item, not just the answer.
T — Template
Template describes the deliverable: a table with named columns, a memo with fixed headings, or a chronology with defined fields. A consistent output can make omissions and unsupported entries easier to spot. It does not make the content filing-ready, and it cannot ensure that every requested row, quotation, or citation is correct.
Weak prompt
Organize it clearly.
CITE template
Produce one table per issue with these columns: # | Type (admission / agreement / no-recall) | Verbatim Q&A | Cite | Why it matters (one sentence). After the tables, add a section "Top 5 most significant exchanges," ranked.
E — Evidence
Evidence tells the model which source locations to provide and what to say when support is not located. Requiring a page-line, Bates, paragraph, or timestamp citation can make a finding easier to check, but the model may still invent a citation, misquote the source, or overlook responsive material. “Not found in the reviewed source” is a requested response format, not proof of absence. Open every cited location and verify it before relying on the output.
Weak prompt
(most prompts simply omit this)
CITE evidence
For every finding, cite the page and line in the form page:line and quote the testimony verbatim. If you find no qualifying testimony on an issue, write "not found in the record." Never infer or extrapolate beyond what the words support.
Why generic prompts fail in legal work
A generic “summarize” instruction does not identify the issues, source boundaries, or deliverable. The model may produce fluent prose while missing a material exchange or supplying unsupported detail. Fluency is not evidence that the source was read correctly.
Legal work inverts ordinary AI priorities. You do not want the model's general knowledge, you want it confined to your record; you do not want a balanced overview, you want your practice priorities; and you never want a guess, you want "not found in the record." A generic prompt leaves those controls unstated. CITE makes the requested scope, format, and source locations explicit so a reviewer has a clearer basis for checking the result. It does not guarantee grounding, completeness, or accuracy.
Professional-responsibility references
These primary sources informed the general cautions on competence, confidentiality, candor, supervision, and verification. Requirements differ by jurisdiction; check the authority that applies to your work.
Common questions
Is the CITE method specific to one AI tool?
Do I need to learn prompt engineering?
Will the CITE method stop the AI from making things up?
What kinds of tasks does it handle?
Is this legal advice or a substitute for judgment?
Stop prompting like you're Googling.
Every template in the library is a CITE prompt, pre-briefed for a specific document and task. Pick one, fill in three or four blanks, and paste a citation-disciplined prompt into any AI tool.
Get new templates by email
New templates, new practice lenses, and the occasional tip on briefing an AI — delivered when they're ready, not before. No spam.
Prefer email? Contact hello@caseprompts.com.