These ten Claude prompts cut the mechanical slog out of discovery review — first-pass sorting, privilege flagging, chronology building — so attorneys can spend their time on the calls only they can make.
Discovery document review eats hours that small firms don’t have. A solo litigator or a two-attorney shop handling a mid-size commercial matter can spend days doing work that is, at its core, pattern recognition: Is this document responsive? Does this passage look privileged? What did this witness know and when? Claude (Anthropic’s large language model, currently available as Claude 3.5 Sonnet and Claude 3 Opus via claude.ai or the API) handles pattern recognition quickly and cheaply. These prompts were assembled for exactly that workflow. Each one produces a first-pass output. Responsiveness rulings, privilege determinations, and litigation strategy stay with the attorney — full stop. Treat every output as a draft memo from a very fast, very fallible paralegal who has never passed the bar.
How these prompts were chosen
The ten prompts below map to the most time-intensive mechanical tasks in a standard discovery workflow: intake sorting, privilege screening, timeline reconstruction, witness profiling, hot-doc identification, deposition prep, inconsistency spotting, opposing-party document analysis, production quality control, and ESI tracking. Each prompt is written for Claude’s extended context window (up to 200K tokens in Claude 3), which means you can paste long document batches directly. Adjust the bracketed variables for your matter. Where the prompt involves privilege or responsiveness, the instructions include explicit AI-limitation language — keep that language in. Removing it defeats the purpose.
1. Responsiveness first-pass on an email batch
Paste a batch of emails (stripped of any content that would itself create a conflict by passing to a third-party service — check your jurisdiction’s ethics rules on cloud AI tools first). Claude sorts each email into Likely Responsive, Likely Non-Responsive, or Unclear, with a one-line reason. The output is a triage list, not a ruling. Your attorney reviews every Unclear and spot-checks the Likely buckets.
You are a document review assistant. I will provide a batch of emails.
The lawsuit is: [ONE-SENTENCE CASE DESCRIPTION].
The discovery request at issue is: [PASTE RELEVANT REQUEST FOR PRODUCTION].
For each email, output a table row with four columns:
1. Document ID or subject line
2. Responsiveness: "Likely Responsive" / "Likely Non-Responsive" / "Unclear"
3. One-sentence reason
4. Any date, sender, or recipient worth flagging
Do NOT make final responsiveness rulings. Do NOT exclude any document
from attorney review. Flag every "Unclear" for mandatory attorney review.
Output format: plain text table.
[PASTE EMAIL BATCH HERE]Tweak: Replace the discovery request language for each RFP. Run separate passes per request number rather than combining — the output stays cleaner and audit trails are easier to reconstruct.
2. Privilege flag (attorney-client and work product)
This is the highest-stakes prompt in the set. Claude surfaces documents with privilege indicators — legal counsel on the CC line, subject lines referencing legal advice, language that sounds like legal strategy. It does not determine privilege. That call belongs to your attorney, every time, without exception. Do not remove the limiting instructions from this prompt.
You are a privilege review assistant. Your role is to flag documents
that MAY contain privileged material for attorney review.
You are NOT authorized to make privilege determinations.
Every document you flag requires attorney review before any withholding decision.
Review the following documents and flag any that show ONE OR MORE of these indicators:
- A licensed attorney is an author, recipient, or CC'd party
- The document discusses or requests legal advice
- The document references pending or anticipated litigation
- The document appears to be attorney work product (legal strategy, mental impressions, legal research)
For each flagged document output:
1. Document ID
2. Privilege indicator(s) present (list each one)
3. Attorneys identified (name/email if visible)
4. One-sentence description of the subject matter — NO legal conclusion
5. Recommended action: "HOLD FOR ATTORNEY PRIVILEGE REVIEW"
For documents with no privilege indicators: output document ID and "No privilege indicators detected — attorney confirmation still required."
IMPORTANT: This output is a first-pass screening tool only.
Final privilege determinations must be made by a licensed attorney.
[PASTE DOCUMENTS HERE]Tweak: Add a fifth indicator for common-interest doctrine situations if your matter involves co-defendants or affiliated entities. Name the attorneys in the matter so Claude can match against known names, not just titles.
3. Chronology extraction from a document set
Building a master timeline from hundreds of documents manually takes a full day or more. This prompt pulls every datable event, communication, or fact-reference from a document batch and returns a sorted chronology. You’ll still need to verify dates against originals — Claude misreads date formats and occasionally hallucinates plausible-sounding dates from ambiguous text.
You are a litigation timeline assistant. Extract every datable event
from the documents below and compile a chronology.
For each event output:
- Date (use format YYYY-MM-DD; if approximate write "circa [date]")
- Document ID or source
- Parties involved
- Event description (one sentence, factual, no legal characterization)
- Direct quote from document (max 30 words) supporting the date
Sort output chronologically, earliest to latest.
If a document contains no datable events, note it as "No date extracted — [Document ID]."
Flag any date that appears internally inconsistent
(e.g., a reply email dated before the email it replies to).
This chronology is a first-pass drafting tool.
Dates must be verified against original documents by the attorney.
[PASTE DOCUMENTS HERE]Tweak: For large batches, run in chunks of 20-30 documents and then run a consolidation prompt that merges the partial chronologies and removes duplicates.
4. Summary memo per witness
Before depositions, you need to know what the documents say about each witness: what they knew, when they knew it, who they communicated with, and what positions they took. This prompt builds that memo from the document set. Feed it documents that reference or were authored by the target witness.
You are a litigation research assistant preparing a witness document summary.
Target witness: [FULL NAME, TITLE, ORGANIZATION]
Matter: [ONE-SENTENCE DESCRIPTION]
Review the documents below. Summarize everything the documents show about this witness:
1. Role and responsibilities (as reflected in documents — do not infer beyond what documents state)
2. Key communications authored by or sent to this witness (date, counterparty, subject)
3. Documents the witness is referenced in but did not author
4. Any positions, statements, or representations attributed to this witness in the documents
5. Apparent knowledge of [KEY ISSUE IN THE MATTER] based on documents — cite Document ID and direct quote for each item
6. Gaps: dates or topics where you would expect documents but none appear
Format as a memo. Each factual claim must cite a Document ID.
Do NOT draw legal conclusions. Do NOT characterize credibility.
This memo is a first-pass research tool for attorney review and refinement.
[PASTE WITNESS-RELATED DOCUMENTS HERE]Tweak: Run a separate pass for each witness rather than batching. Separate memos are faster to review and cleaner to share with co-counsel.
5. Hot-doc detection (high-impact passages)
Hot documents are the ones that end up in opening statements. This prompt flags passages that are factually significant, emotionally resonant, or directly contradict a party’s stated position. It tends to over-flag — you’ll review more documents than you’d expect — but missing a hot doc is worse than reviewing one extra.
You are a litigation document analyst. Identify high-impact passages
in the documents below — passages likely to be significant at trial
or in dispositive motion practice.
Flag a passage if it meets ONE OR MORE of these criteria:
- Directly contradicts a known party position (describe the contradiction)
- Contains an admission, concession, or acknowledgment of a problem
- Shows knowledge of [KEY DISPUTED FACT] prior to [KEY DATE]
- Contains unusually candid internal language about risk, liability, or fault
- Is from a senior decision-maker on a core issue in the matter
For each flagged passage output:
- Document ID
- Page/paragraph location if available
- Direct quote (verbatim, max 75 words)
- Why it may be significant (one sentence — factual, no legal conclusion)
- Significance tier: "High" / "Medium" / "Review"
Do NOT make legal or strategic conclusions.
An attorney must evaluate each flagged passage for actual evidentiary value.
This is a first-pass triage, not a trial exhibit list.
[PASTE DOCUMENTS HERE]Tweak: Populate the bracketed fields with your actual disputed facts and key dates. The more specific those fields are, the fewer false positives you get.
6. Deposition prep outline from a document set
This prompt takes the document record for a witness and produces a topic outline with document anchors — the scaffolding for a deposition outline, not the finished product. The attorney writes the actual questions, decides the order, and handles strategy. Claude builds the document map the attorney works from.
You are a litigation support assistant. Build a deposition preparation outline
for the following witness based solely on the documents provided.
Witness: [NAME, TITLE]
Matter summary: [TWO SENTENCES MAX]
Primary areas of inquiry: [LIST 3-5 TOPICS FROM YOUR CASE THEORY]
For each topic area:
1. List the documents that relate to it (Document ID + one-line description)
2. Identify factual points the documents establish about this witness and this topic
3. Identify any apparent inconsistencies between documents on this topic
4. Flag any documents that would benefit from the witness's explanation
Output format: numbered outline, one section per topic.
Each factual point must cite a Document ID.
Do NOT draft deposition questions.
Do NOT suggest litigation strategy.
Do NOT make credibility assessments.
This is a document map for the attorney's review —
final deposition strategy is determined by the attorney.
[PASTE WITNESS-RELATED DOCUMENTS HERE]Tweak: After running this, feed the outline into Prompt 7 (inconsistency follow-ups) to automatically surface questions worth exploring.
7. Follow-up question generator from inconsistencies
This prompt takes two or more documents that appear to contradict each other and generates neutral factual questions designed to surface the inconsistency. “Neutral factual” is deliberate — Claude writes questions that elicit information, not questions built on argumentative premises. The attorney decides which questions to use and how to weaponize them.
You are a deposition preparation assistant. I will provide two or more documents
that appear to contain inconsistent factual statements.
Your task:
1. Identify each apparent inconsistency between the documents (be specific — quote both sides)
2. For each inconsistency, draft 2-3 neutral, open-ended factual questions
designed to explore the inconsistency without telegraphing the conflict
3. Note which Document IDs underlie each inconsistency
Question format requirements:
- Open-ended (cannot be answered yes/no)
- Factual, not argumentative
- Do not embed the answer or characterize the witness
- No compound questions
Output: numbered list of inconsistencies, each followed by its draft questions.
IMPORTANT: These are draft questions for attorney review only.
The attorney determines which questions to use, in what order,
and how to handle the witness's responses.
This output is not deposition strategy — it is a drafting starting point.
Document A: [PASTE OR LABEL]
Document B: [PASTE OR LABEL]
[Add Document C, D as needed]Tweak: Works well after running Prompt 6. Pull the inconsistencies Claude flagged in the deposition outline and feed them directly into this prompt as the source documents.
8. Opposing-side document summary
When opposing counsel produces a large document set, the first job is figuring out what they sent you and what they’re trying to show. This prompt builds a production summary — what categories of documents appear, what time periods they cover, what narrative they seem to support. Useful before your attorney digs in, not instead of that dig.
You are a litigation support analyst. Summarize the opposing party's document production below.
Provide:
1. Document types present (emails, memos, contracts, invoices, reports, etc.)
with approximate counts
2. Date range of documents (earliest to latest)
3. Key individuals who appear as authors or recipients
4. Topics and subject matters covered across the production
5. Any apparent gaps: topics you would expect given the case description
but that do not appear in these documents
6. Documents that appear particularly significant — quote directly, cite Document ID
Case description: [TWO SENTENCES]
Known disputed issues: [LIST]
Do NOT draw legal conclusions. Do NOT assess litigation strategy.
Do NOT characterize document authenticity.
This is a first-pass inventory for attorney review.
The attorney determines the significance of this production.
[PASTE OPPOSING PRODUCTION HERE]Tweak: The gaps section is often the most useful output. Populate “known disputed issues” carefully — vague inputs produce vague gap analysis.
9. Production set quality check (gaps and missing dates)
Before you certify a production or send a deficiency letter, you want a quality check on what’s in the set versus what should be there. This prompt reviews a production index or document batch for date gaps, missing custodians, and metadata anomalies worth investigating. It does not tell you whether the production is legally complete — that’s your call.
You are a document production quality control assistant.
Review the production index or document set below and identify potential gaps or anomalies.
Check for:
1. Date gaps: periods within the relevant date range where document volume drops
unexpectedly to zero or near-zero (flag each gap with start/end date and duration)
2. Missing custodians: individuals expected in the production based on
the documents present who do not appear as authors or recipients in certain periods
3. Thread breaks: email chains where replies exist but originating emails do not
4. Metadata anomalies: documents with no date, no author, or implausible dates
(e.g., created date after modified date)
5. Format inconsistencies: document types present in some periods but absent in others
without apparent explanation
Relevant date range: [START DATE] to [END DATE]
Expected custodians: [LIST KEY NAMES]
Matter summary: [ONE SENTENCE]
Output: numbered list of findings. Each finding: description, date range or Document IDs affected, and a suggested follow-up question for the producing party.
This is a QC drafting tool.
Whether any gap constitutes a deficiency requiring a formal letter
is a legal determination made by the attorney.
[PASTE PRODUCTION INDEX OR DOCUMENT BATCH HERE]Tweak: Pair this with a document review platform’s index export (Relativity, Everlaw, or even a simple Excel export) rather than pasting raw documents — the structured data produces cleaner gap analysis.
10. ESI status check
ESI issues surface early and cause problems late. This prompt reviews a set of case-related communications — meet-and-confer notes, ESI protocol drafts, correspondence — and produces a status summary: what’s been agreed, what’s still open, and what topics haven’t been addressed yet. Think of it as a checklist assistant, not an ESI counsel.
You are a litigation support assistant tracking ESI discovery status.
Review the documents below (meet-and-confer correspondence, ESI protocol drafts,
court orders, and related communications) and produce a status report.
For each ESI topic listed, report current status based solely on the documents:
Topics to track:
1. Custodian list — agreed / disputed / not yet addressed
2. Date range — agreed / disputed / not yet addressed
3. Search terms — agreed / disputed / not yet addressed
4. Document sources (email, cloud storage, mobile, etc.) — agreed / disputed / not yet addressed
5. Production format (native, TIFF, load file specs) — agreed / disputed / not yet addressed
6. Privilege log format and timing — agreed / disputed / not yet addressed
7. Litigation hold notices — confirmed issued / not confirmed / disputed
8. Any outstanding objections or unresolved disputes
For each item: cite the document and date that supports the status you assign.
Flag any item with no supporting document as "No record in provided documents — confirm with counsel."
Output as a table: Topic | Status | Supporting Document/Date | Open Issues
This is a status tracking tool only.
ESI compliance obligations and dispute resolution strategy
are determined by the attorney.
[PASTE ESI-RELATED DOCUMENTS AND CORRESPONDENCE HERE]Tweak: Add jurisdiction-specific ESI protocol requirements (FRCP 26(f), local rules, standing orders) as a note at the top of the prompt so Claude can flag topics your jurisdiction treats as mandatory. Remove any items not relevant to your matter to keep the output tight.
Notes on using these prompts
Model choice matters here. Claude 3.5 Sonnet is faster and cheaper; Claude 3 Opus handles longer, more complex document batches more reliably. For anything over 50,000 words of pasted documents, Opus is worth the extra cost. Both are available at claude.ai (Pro plan, $20/month) or via Anthropic’s API if you’re building a more automated intake.
Every prompt above includes explicit limiting language. That language is not legal boilerplate decoration — it shapes the output. Claude is more likely to hedge appropriately and less likely to over-conclude when the prompt specifies its role. If you strip out the “first-pass only” instructions to save tokens, expect outputs that sound more confident than they should be.
Before you paste client documents into any AI tool, including Claude, check your state bar’s ethics opinions on cloud AI tools and confidentiality. Several states have issued guidance; others have not. The answer is not “don’t use AI” — it’s “know what your bar says and act accordingly.” That’s your call, not ours.
These prompts break most often on: handwritten or poorly OCR’d documents (Claude reads text, not images), documents with heavy redactions that alter meaning, and very large batches where early documents in the paste get less attention than later ones. For high-stakes matters, run batches in chunks of 20-30 documents and spot-check early-batch outputs against a manual review of the same documents.
The ten prompts above cover the mechanical layer of discovery. The judgment layer — what matters, what to do about it, how to use it — stays with the attorney running the matter.
Related reading
- Briefpoint Tested: AI Discovery Drafting for Solo and Small-Firm Litigators
- 10 ChatGPT Prompts Every Solo Lawyer Should Save (Tested on Real Matters)
- The 5-Prompt Sequence for First-Pass Contract Review with Claude or GPT
- 7 Demand Letter Drafting Prompts That Actually Save Time
- Lexis+ AI vs Westlaw Precision vs CoCounsel: The 2026 Legal Research AI Showdown for Small Firms