Document Review Classifier
During document review in litigation, as a first-pass classification before attorney review.
Document review is where litigation budgets go to die. In any case with meaningful production, someone has to read each document and decide whether it is responsive, how relevant it is, whether it is privileged, and whether it is a hot document that changes the case. A consistent first-pass classification, applied document by document against the actual discovery requests, dramatically speeds the workflow and helps ensure nothing important slips through unreviewed.
Claude is well suited to this first pass. Give it the case description, the specific discovery requests at issue, and the list of privilege holders, and it returns a structured classification for each document covering responsiveness, relevance tier, a privilege screen, confidentiality designation, the key facts established, and a hot-document flag. Used as a triage layer ahead of attorney eyes, it lets your team spend its expensive hours on the close calls rather than the obvious ones.
Use the prompt below as a starting point, and treat its output as a first cut, never a final call. Privilege determinations in particular demand attorney judgment; an AI flag is a prompt to look harder, not a decision. Claude's classification is a work-product draft for attorney review, not legal advice, and the responsible lawyer must confirm every responsiveness and privilege call before anything is produced or withheld.
The Prompt
Classify the documents below for discovery review. <requests> [PASTE THE DISCOVERY REQUESTS AT ISSUE, WITH NUMBERS] </requests> <document id="[DOC ID]"> [PASTE THE DOCUMENT TEXT; REPEAT THIS BLOCK FOR EACH DOCUMENT] </document> We are the: [PRODUCING / RECEIVING] party Case: [BRIEF CASE DESCRIPTION] Attorneys and privilege holders: [LIST IN-HOUSE AND OUTSIDE COUNSEL BY NAME, AND THE CLIENT ENTITIES] For each document, determine: 1. Responsiveness: does it fall within the scope of a request in <requests>? (Responsive / Not Responsive / Potentially Responsive, with the request number) 2. Relevance to the claims and defenses (Highly Relevant / Relevant / Marginally Relevant) 3. Privilege screen: - Attorney-client privilege (communication with counsel for legal advice?) - Work product (prepared in anticipation of litigation?) - Other privileges or protections 4. Confidentiality designation under any protective order 5. Key information: what facts the document establishes 6. Hot document flag: does it contain key evidence for either side? 7. Recommended coding tags For every responsiveness, relevance, privilege, and hot-document call, quote the sentence that supports it. Flag privilege calls for attorney review rather than resolving them. If something I need is missing or ambiguous, say so instead of assuming. Do not invent facts, citations, or quotations; mark anything you are unsure of [VERIFY]. This is a draft for attorney review.
Example Output
A classification card for each document with responsiveness, privilege, relevance, and key facts extracted.
Illustrative example — names, figures, and facts are fictional.
DOCUMENT REVIEW CLASSIFICATION Case: Halvorsen v. Pinecrest Medical Devices (product liability) Requests at issue: RFP Nos. 4-7 (design, testing, complaints) Privilege holders: In-house counsel J. Reyes; outside firm Calder LLP ----------------------------------------------------------------- DOC ID | RESPONSIVE | RELEVANCE | PRIVILEGE | HOT? ----------------------------------------------------------------- PCM-0012| Responsive | Highly Rel. | None | YES Key facts: Internal email; engineer warns the Gen-2 valve "fails the cycle test ~1 in 40" before launch. Smoking gun on notice. Suggested tags: NOTICE, DESIGN-DEFECT, KEY-WITNESS. ----------------------------------------------------------------- PCM-0034| Responsive | Relevant | None | No Key facts: QA testing log, Gen-2 valve, March 2024 cycle results. Tags: TESTING, RFP-5. ----------------------------------------------------------------- PCM-0051| Pot. Resp. | Marginal | A-C PRIVILEGE? | No Key facts: Email to J. Reyes seeking "legal read" on draft warning label. Likely attorney-client; ATTORNEY MUST REVIEW before any production. Tags: HOLD-PRIVILEGE, LOG. ----------------------------------------------------------------- PCM-0067| Not Resp. | -- | None | No Key facts: Cafeteria vendor invoice; outside RFP scope. Tags: NON-RESPONSIVE. ----------------------------------------------------------------- PCM-0078| Responsive | Highly Rel. | WORK PRODUCT? | No Key facts: Memo prepared after first claim letter, analyzing failure rate. May be work product (anticipation of litigation); ATTORNEY MUST REVIEW. Tags: HOLD-WP, LOG, CONFIDENTIAL. ----------------------------------------------------------------- NOTE: Privilege flags are preliminary screens only. All flagged documents require attorney confirmation before withholding or logging.
Tips
- •Never rely on AI alone for privilege calls; an attorney must review every privilege-flagged document.
- •Provide the specific discovery requests to improve responsiveness accuracy.
- •In federal court, seek a Federal Rule of Evidence 502(d) order before AI-assisted review.
- •Process documents in batches by custodian for better consistency.
- •Treat the output as a draft for attorney review; verify every citation, quotation, date and figure against the source before relying on it.
Frequently Asked Questions
When should I use Claude for document review classification?
Use it as a first-pass triage layer during the review phase, before attorney eyes-on review, ideally processing documents in batches by custodian for consistency. It is most valuable for sorting the obvious responsive and non-responsive material and surfacing likely privileged and hot documents, so your team concentrates its time on the close calls and the documents that actually move the case.
Can I produce or withhold documents based on Claude's classification alone?
No, never. Treat every classification, and especially every privilege flag, as a preliminary screen that requires attorney confirmation. An erroneous responsiveness call risks sanctions, and an AI privilege miss can waive privilege on a key document. The classification accelerates review; it does not replace the attorney judgment required before anything is produced, withheld, or logged.
How do I get the most reliable classifications?
Provide the exact text or numbers of the discovery requests at issue so Claude can assess responsiveness against the real scope, and list all privilege holders, including in-house and outside counsel, by name. Batch documents by custodian for consistency, and ask Claude to flag, rather than resolve, every potential privilege call so an attorney makes the final determination.
Is it safe to feed client documents into a consumer AI tool for review?
This is a high-stakes confidentiality question. Discovery documents are squarely protected by Model Rule 1.6, and many are privileged. ABA Formal Opinion 512 requires evaluating a tool's security, retention, and training practices and obtaining informed client consent where appropriate. Use an enterprise deployment with no-training and confidentiality terms; do not upload client documents to a consumer chatbot without proper data protections in place.
Get New Prompts Like This Every Week
Join the free Claude for Lawyers newsletter — weekly prompts, tutorials, and practice-specific guides.