Short answer / recommendation
Use a retrieval-first workflow and a strict prompt template: (1) provide Claude only vetted source documents (statutes, contracts, briefs, internal emails) or a verified citation list, (2) set a firm system role that enforces citation/quotation rules and confidence flags, (3) require the model to stop and report “no on-point authority” when it can’t find support. Always require human attorney sign-off of any “final” memo.

Why this works
Hallucinations usually come from the model inventing supporting authority or over-generalizing. For legal memos you must (a) control the information Claude can cite, (b) force explicit provenance for every claim, and (c) bake review steps into the workflow.

Concrete prompt pattern (copyable)
System prompt (top-level):
You are a legal analyst producing memoranda for licensed in-house counsel. You must only cite sources provided in the attached corpus or explicitly named, and you must include for every legal claim a citation in Bluebook or official reporter form plus the exact quoted passage (≤ 200 characters) and a document identifier (source name, page/paragraph). If you cannot find authority supporting a claim, write: “No on‑point authority located in provided corpus.” Mark low-confidence conclusions with [LOW CONFIDENCE]. Do not invent statutes, cases, or citations.

User prompt (issue + constraints):
Facts: [concise facts].
Jurisdiction: [state/federal].
Task: Produce an IRAC memo limited to 800–1,200 words covering: (1) short answer, (2) analysis with citations and quoted excerpts, (3) risk assessment (likelihood of adverse outcome and exposure in $ ranges), (4) recommended mitigation steps prioritized. Sources allowed: [attach internal docs / list public sources]. Required: footnotes with full citations; inline bracketed provenance like {DocA ¶4}.

Verification follow-up (after draft):
List every citation you used with the exact URL or database identifier and the quoted text’s location (page/para). For any citation you did not quote, provide the sentence that links your conclusion to that authority.

Practical guardrails (must implement)
- Retrieval-first: run legal research separately (Westlaw/Lexis/official gazettes) and attach the results to Claude. Don’t rely on raw web search by the assistant.
- Source whitelist: tell Claude “only cite from attached/whitelisted sources.”
- Explicit provenance: require quote + source ID for every legal assertion.
- Low-confidence tags: require [LOW CONFIDENCE] for inferences or extrapolations.
- Failure mode: require exact string “No on‑point authority located” when none found.
- Version & logging: store the prompt, model name, and output for audit.

Human workflow checklist (quick)
1. Gather authoritative sources (cases, statutes, contracts) and attach them. 2. Run retrieval / search in legal database; create a source list. 3. Run Claude with the system + user prompt above. 4. Run the verification follow-up to get citation mapping. 5. Attorney reviews ALL quotes against primary sources and signs off. 6. Archive prompt/output + reviewer initials.

Best-for / Avoid-if
Best-for: internal memo drafts, long-context synthesis, contract risk summaries when you can attach primary authority. Avoid-if: you need certified legal opinions without human review, or you can’t run independent source verification.

Decision criteria
- Budget: tighter budgets may trade off the extra database access and manual verification. If you have subscription legal databases, enforce retrieval-first; otherwise restrict AI use to internal-doc summarization only.
- Skill level & team size: small teams must enforce stricter checklists; larger teams can add a dedicated reviewer role.
- Workflow stage: use Claude for first-draft analysis and structure; always perform human legal review before circulation.

If you want, I can write a one-click system+user prompt file you can paste into Claude and a short verification checklist formatted for your legal intake form.

Compare Claude and ChatGPT