Working prompts for legal tasks, written to produce output you can check. Each one names its material, asks for sources, and says what to do when the answer is not there. Open a category below, or pick a prompt from the list to jump straight to it.
Four habits behind every prompt here
Name the material. "Using only the attached document" beats "look at this".
Ask for the quote and the page. A source reference is what makes checking possible.
Give a rule for gaps. Specify how to report missing information. Then check whether the model followed that instruction.
Fix the shape. A table with named columns is easier to check than prose.
Confidential material goes only into tools your firm has approved, on the plan and settings that were approved. See Confidentiality and data flows.
Reviewing documents and evidence
4 items
Find every date, and say which are conditional
When: A letter, agreement, or email thread where timing matters.
Prompt
Using only the attached document, list every date or time reference.
For each one give: the exact words, the document ID and page, what the date refers to, and whether it is firm, estimated, conditional, or disputed.
If a date is conditional, quote the condition.
Do not infer dates that are not stated. If none are found, say "none found".
Output a table with columns: Quote | Document | Page | Refers to | Status.
What to check: Open each quoted passage. Look for conditions the table may have dropped ("subject to", "if", "unless").
When: First pass on a contract before you read it properly.
Prompt
From the attached agreement only, extract: parties, effective date, term and renewal, termination rights, payment terms, governing law, notice requirements, and any clause that limits liability.
For each item quote the clause and give its section number. Where the agreement is silent, write "silent".
Do not summarize or paraphrase inside quotation marks.
What to check: Confirm every section number against the document. Read the defined terms; extraction often misses a definition that changes a clause.
When: A counterparty returns a redraft, or you need to confirm what changed between drafts.
Prompt
Compare Version A and Version B of the attached document.
List every difference as a table: Section | Version A wording | Version B wording | Effect of the change (one sentence).
Include deletions and additions. Flag any change that shifts a date, an amount, a party, or a condition.
Do not describe a change that is not actually in the documents.
What to check: Run a proper redline as well; use the AI table to prioritize what to read first, not to replace the comparison.
Separate what the document says from what it implies
When: You want the model's reading of a document but need to keep evidence and inference apart.
Prompt
Read the attached document and answer the question below in two labelled parts.
Part 1, "Stated": only propositions the document says directly, each with a quote and page.
Part 2, "Inferred": conclusions you draw from Part 1, each labelled "inference" with the stated facts it relies on.
Do not put an inference in Part 1. Question: [your question]
What to check: Treat Part 2 as proposals. The useful work is checking Part 1; the inferences are only as good as the quotes they rest on.
When: A set of documents in a matter; you need the timeline before anything else.
Prompt
Build a chronology from the attached documents only.
One row per event. Columns: Date | Event (one sentence, factual) | Source document | Page or paragraph | Exact supporting quote | Certainty (stated / conditional / disputed / inferred).
Where two documents conflict, include both rows and mark them "conflict".
Do not fill gaps from general knowledge. If the date of an event is unknown, write "undated" and explain what is known.
What to check: Spot-check at least one row per document against its page. Then look for what is missing: events you know happened that do not appear.
When: You already have a chronology and want a second pair of eyes.
Prompt
Here is a draft chronology and the documents it was built from.
1. For each row, confirm the quoted passage exists at the cited location. Mark "verified", "not found", or "wording differs" (and show the difference).
2. List events in the documents that the chronology does not include.
3. List rows whose certainty label seems wrong, with the reason.
Do not add rows; report only.
What to check: A "verified" from the model means it found matching words, not that the row is right. Re-read every "wording differs".
When: Preparing for cross-examination or a motion from a discovery transcript.
Prompt
From the attached transcript, list every factual claim the witness made about [topic].
For each: the claim in one sentence, the page and line numbers, and the exact words.
Group claims by subject. Flag any two claims that appear inconsistent with each other and cite both locations.
Do not add claims the witness did not make. Do not characterize credibility.
What to check: Read the lines around each cited passage. Context frequently changes what a claim means.
When: Checking whether an answer to undertakings or a discovery response actually answers each request.
Prompt
Here is a list of numbered requests and a document responding to them.
For each request: quote the part of the response that addresses it, say whether the response answers it fully, partly, or not at all, and quote any refusal or qualification.
If a request has no corresponding response, say "no response".
Output a table with columns: Request # | Response quote | Answered? | Qualification.
What to check: Check every "fully"; the model is generous. The "partly" and "no response" rows are where the follow-up letter comes from.
When: You need a first list of questions to ask a witness about a specific document.
Prompt
Read the attached document. Draft questions for its author on examination for discovery, organized by topic.
Each question should be answerable yes or no or with a specific fact, and should refer to the passage it concerns by page.
Include questions about who else saw the document, when, and what was done in response.
Do not assume facts not in the document.
What to check: This is a starting list, not a sequence. Re-order to your own method and remove questions that invite narrative answers if that is not what you want.
When: Starting research on a legal issue with a tool that retrieves authorities.
Prompt
State the legal propositions relevant to [issue] in [jurisdiction], one per line.
For each proposition, cite the authority you rely on with a full citation and a pinpoint, and quote the passage that supports it.
Mark each citation "verify" until I confirm it.
If you are not confident an authority exists or supports the proposition, say so instead of citing it.
What to check: Open every authority in a recognized source. Confirm it exists, confirm the pinpoint, read the passage. Nothing goes into a factum until that is done; check the certification required for the court and document.
When: You have a decision and want to know quickly whether it helps.
Prompt
Read the attached decision. Answer these questions, each with paragraph references:
1. What was the issue?
2. What did the court decide and on what reasoning?
3. What facts did the outcome turn on?
4. Does it address [my question]? Quote the relevant paragraphs.
5. Was it followed, distinguished, or appealed, if the document says?
If the decision does not address something, say so.
What to check: Paragraph references are easy to check and often wrong by one or two; confirm each. Check subsequent history yourself.
When: Any first draft: a letter, a memo section, an affidavit paragraph.
Prompt
Draft [document] for [audience] to achieve [purpose], in [tone].
Use only [identified source files and versions].
Target length: [n] words. Format: [format].
Before drafting, check whether a missing instruction leaves materially
different possible assignments, such as accepting an offer versus merely
acknowledging it. If so, ask the smallest question needed and wait.
Do not ask about details already supplied or minor presentational choices.
A missing fact is different: mark [FACT NEEDED: describe] where a useful
draft can proceed. If that fact determines the proposed position, explain
the dependency and stop. Do not turn my preferred answer into evidence.
Draft from the supplied record. After each factual statement, add a
bracketed source and page reference. Do not invent facts or quotations.
Preserve conditions and uncertainty. List unresolved items after the draft.
What to check: Answer material instruction questions; verify factual references separately. A preference does not prove an event happened. For an already clear assignment, simpler wording is sufficient.
When: Tightening a draft without losing facts, conditions, or hedges.
Prompt
Revise the draft below for clarity and length. Keep every fact, date, amount, condition, and qualification exactly as stated.
After the revision, list every substantive change you made and every sentence you removed, so I can confirm nothing material was lost.
Do not change quoted material.
What to check: Compare the facts, not the prose. Look especially for a "may" that became "will" or a condition that disappeared.
Review [draft] against [identified sources], addressing [issue].
Return up to three material objections. For each, quote the draft passage,
cite the relevant source and location, explain the objection, and identify
any unresolved evidence question. Separate an actual contradiction from a
hypothetical alternative. Do not invent facts or authorities.
If no material supported objection is found, say so. If the sources are
missing, identify what you need and limit the review to internal logic.
Report only; do not rewrite the draft.
What to check: Test each objection against its source. Agreement is not proof of correctness; disagreement is not proof of useful review. Accept, reject with reasons, or investigate each material objection.
When: End of day; you have rough notes and need clean dockets.
Prompt
Convert these rough notes into time entries.
Each entry: date, matter, duration in tenths of an hour, and a description in the firm's style: [example of your style].
Keep every task I listed; do not merge or drop any. Do not invent durations: where I gave none, write "[time?]".
Do not include anything that reveals privileged content beyond what is necessary to describe the task.
What to check: Your notes should not contain client confidences unless the tool is approved for them. Confirm the total against your day.
When: A thread with many messages that needs a short record.
Prompt
Summarize the attached email thread as a file note.
Give: participants, the date range, each decision or commitment made (with the sender, date, and a quote), and open items.
Keep the order chronological. Do not characterize anyone's motives.
End with a list of attachments mentioned and whether they were included.
What to check: Confirm every quoted commitment against the message it came from, including who actually sent it.
When: Any task you delegate to an agent that can read or change files.
Prompt
Task: [one sentence].
You may: read files in folder [X]; write a new file named [Y]. You may not: modify or delete existing files, send anything, or access the internet.
Steps: 1. [step] 2. [step] 3. [step].
Stop and ask me if: a file cannot be read, a step would change something outside [Y], or you are unsure which of two interpretations is right.
When finished, list every file you read and every action you took.
What to check: The permissions have to be enforced by the tool, not just stated here. Read the list of actions it reports and compare it with what you expected.
When: A folder of documents that needs a manifest before review.
Prompt
Create an inventory of the documents in folder [X].
For each file: file name, type, page count if available, date created and modified, and the first line of text if the file contains text.
Flag files that could not be opened or that contain no readable text (likely scans needing OCR).
Write the inventory to [X]/inventory.csv. Do not open any file outside [X] and do not change any file.
Report what you did when finished.
What to check: Open three or four files at random and compare them with their rows. Make sure the files flagged as unreadable really are.
A review-table workflow (Harvey, Legora, or similar)
When: Many documents, the same questions for each.
Prompt
Columns (one question per column), applied to every document in the set:
1. Document date (quote the date line).
2. Parties named (quote).
3. Does the document mention [topic]? Quote the passage, or write "no mention".
4. Any deadline or notice requirement? Quote it with the clause number.
For every cell, give the quote and page. Leave a cell blank rather than guess.
What to check: Review-table tools make checking faster, not unnecessary. Sample cells from every column, and check all "no mention" cells for a sample of documents by searching the text yourself.
When: You want an automated first check of a chronology or table before a person reviews it.
Prompt
You are checking the attached table against the attached source documents.
For each row: find the quoted passage at the cited location. Report "found", "not found", or "wording differs" with the actual wording.
Do not fix the table. Do not add rows. Output a table: Row # | Result | Note.
What to check: This finds citation errors, not reasoning errors. A person still reads the rows that matter.
Optional: check the substance of a claim
For a few material claims, gather evidence in a fresh session without showing the proposed answer. Reading the source directly is often simpler.
Turn the claim into a neutral, open evidence question. For “Alder approved the revised schedule”, ask what the record establishes about Alder’s response to that schedule.
Start a fresh session with the complete bounded record but no draft, proposed answer, or prior checker answers. Confirm project context or memory does not reintroduce the draft. If separation cannot be established, describe this as ordinary source checking.
Use the separate evidence prompt below. Then compare its answer with the claim: supported, contradicted, or unresolved. Open the cited passages before revising.
Check source coverage, question neutrality, and qualifications. Fresh context does not make the model an independent witness. Review source-to-draft completeness separately.
Evidence prompt — fresh session
Question: [neutral evidence question].
Permitted record: [source IDs and versions].
Use only that record. Give the relevant passages and locations, including
qualifications and conflicting passages. Then state what those passages
establish and what remains unresolved. Distinguish no evidence found
from evidence that an event did not happen.
If a required source is missing or unreadable, report the coverage gap.
Do not fill it from memory. Stop after the evidence answer.
When: Before anything generated goes into real work.
Prompt
Before I rely on this output, list every date, amount, name, quotation, citation, and conditional statement it contains, each with where it came from in the source material.
For any item with no source, mark it "unsourced".
Then list anything in the source material that seems relevant but did not make it into the output.
What to check: Unsourced items get removed or sourced. The last list is where omissions hide, which are the errors that look least like errors.
When: Before reusing an extraction, classification, or drafting prompt across many documents.
Prompt
Task and prompt to test: [paste].
Fictional example input: [paste].
Required result and prohibited errors: [specify].
Propose six short tests: two straightforward cases, two changes that
should preserve the relevant result, and two changes that should alter it.
For each altered case, change one relevant feature at a time and show
the exact before-and-after text. State which output field should remain
stable or change, and why. Do not require identical prose.
Return two separate sections:
A. Test inputs, with IDs and no answers.
B. Proposed answer key: expected fields, source support, prohibited
errors, and any uncertainty about the expected result.
If my success criteria are too vague to determine a result, ask for the
missing rule before creating the tests. Do not infer substantive legal
rules. Stop after proposing the tests; do not claim they were run.
How to run it
Review and correct the proposed answer key yourself.
Run the prompt under test on inputs alone in separate sessions.
Retain a fresh example that did not inform revisions.
What to check: Confirm each edit changes only the intended feature. A name change is not irrelevant if identity is the issue. Small synthetic tests diagnose failures; they do not estimate real-world error rates.
Prompts are starting points, not guarantees. Products interpret instructions differently; test a prompt on fictional material before using it on a matter.
