Prompting AI for Legal Summaries: What Works and What to Avoid

You asked Claude to summarize a 40-page vendor agreement. It confidently stated a 90-day termination notice. The contract says 30. Your associate caught it in review — but only because they read the source. Legal summarization with AI works when prompts enforce citation, defined terms, and explicit uncertainty — and when a licensed attorney verifies output before anyone acts on it. Disclaimer: This article is educational prompt guidance, NOT legal advice. Always verify AI output with a licensed attorney. Never paste confidential client information into public AI tools. This guide covers Claude legal summarization patterns, source citation rules, the "Not specified" rule, defined terms handling, XML section tags, contract vs case law prompt templates, hallucination risks, human verification checklist, and FAQ.

What AI legal summaries are good for

First-pass extraction: payment terms, dates, parties, defined terms index, clause inventory. Comparison scaffolding: side-by-side term tables across two drafts — human confirms. Case research orientation: issue lists and procedural timeline — verified against primary sources. Not suitable for: final legal opinions, filing-ready briefs, privilege-sensitive strategy, or compliance sign-off without human review.

Mandatory security and ethics

Never input: client names + identifiable facts, privileged communications, sealed materials, or trade secrets into consumer ChatGPT/Claude without firm-approved enterprise agreements. Use firm-vetted tools with zero-retention and appropriate BAAs/DPAs where required. Label every AI summary: DRAFT — REQUIRES ATTORNEY REVIEW. Prompt footer: This output is for attorney review only, not legal advice to clients.

Claude legal summarization patterns

Claude Opus 4.8 handles long contracts and nested definitions with strong instruction-following. Patterns that work: 1. XML-tagged sections (document, task, rules, output_schema) 2. Quote-first extraction before paraphrase summary 3. Defined terms glossary pass as separate prompt step 4. "Not specified" instead of silent inference 5. Section ID on every bullet Avoid: "Summarize this contract" with no schema — produces fluent wrong numbers.

The "Not specified" rule

If a term, date, amount, or party obligation is not explicit in the source, output exactly: Not specified in document. Never: infer industry standard, assume mutual symmetry, or fill gaps from training data. Prompt line: Silence in source = Not specified. Do not use external knowledge of "typical" clauses. This single rule eliminates most dangerous hallucinations in contract summaries.

Cite sources — every claim

Format: [Claim] — Quote: "exact text" — Section: §4.2 / Page 12 / Heading "Termination" For tables: term | value | quote | section High-trust prompt: No paraphrase for dollar amounts, dates, or notice periods — quote only. If quote spans defined term, include definition section reference.

Defined terms discipline

Step 1 prompt — glossary extraction: DOCUMENT: [paste or attach] Extract all DEFINED TERMS (capitalized terms with explicit definitions in document). Output table: Term | Definition quote | Section If term used but not defined: flag UNDEFINED TERM. Step 2 uses glossary: Use defined term meanings exactly as in GLOSSARY — do not use colloquial meaning. Example failure: "Affiliate" colloquially vs defined narrowly in §1.1 — wrong summary without glossary pass.

XML section tags for legal prompts

Claude responds well to structured tags: <document>[text or reference]</document> <task>Extract termination provisions</task> <rules> - Quote exact language for all dates and amounts - Not specified if absent

• No external legal knowledge

</rules>

<output_schema>

Table: provision | quote | section | notes

</output_schema>

Tags separate instructions from source — reduces instruction bleed into quotes.

Contract summary prompt (full template)

<document>

[contract text]

</document>

<glossary>

[paste step 1 output or attach]

</glossary>

<task>

Produce attorney-review summary for associate onboarding.

</task>

<must_extract>

• Parties and roles
• Term and renewal
• Termination notice and cause
• Liability cap (exact amount)
• Indemnification scope
• Governing law and venue
• Payment terms

</must_extract>

<rules>

• Every MUST EXTRACT item: quote + section or Not specified in document
• Use GLOSSARY for defined terms
• No legal advice or recommendations
• Flag ambiguous language with AMBIGUITY: note

</rules>

<output>

1. Executive summary (max 200 words, no numbers without section ref)
2. MUST EXTRACT table
3. Open questions for attorney

</output>

Run on Claude Opus 4.8. GPT-5.5 acceptable for structured JSON table output with same rules.

Case law summary prompt

<document>

[opinion text]

</document>

<task>

Summarize for litigation associate — issue, holding, reasoning, disposition.

</task>

<rules>

• Distinguish facts vs holding vs dicta — label each
• Quote holding sentence verbatim with citation pin cite if present
• Not specified if relief or damages unclear
• Do not predict application to other facts

</rules>

<output_schema>

Caption | Court | Date

Facts (bullets, each with paragraph ref)

Issue (numbered)

Holding (quote + ref)

Reasoning (bullets with refs)

Disposition

Key quotes table

</output_schema>

Case summaries fail when model collapses dicta into holding — prompt explicitly separates them.

Contract vs case law — prompt differences

| Aspect | Contract | Case law |

| Primary risk | Wrong numbers, missed conditions | Wrong holding, merged dicta |

| Key artifact | Defined terms glossary | Pin cites and disposition |

| Output | Clause table + open questions | Issue/holding/reasoning structure |

| Never | Infer "standard" terms | Predict case outcome for new facts |

Do not reuse contract prompt for opinions — schema mismatch causes omissions.

Hallucination risks in legal AI

Invented cross-references to sections that do not exist.

Rounded dollar amounts or converted currencies.

Symmetric obligations ("each party may terminate") when only one side has the right.

Missing conditions precedent buried in definitions.

Citing statutes or cases not in the provided document (external knowledge bleed).

Mitigation: quote-first extraction, Not specified rule, no external knowledge line, human spot-check of every number.

Human verification checklist

Before relying on any AI legal summary:

• Every dollar amount and date traced to quote in source?
• Every MUST EXTRACT item present or marked Not specified?
• Defined terms used consistently with glossary?
• Any AMBIGUITY flags reviewed by attorney?
• No recommendations or "you should" language slipped in?
• Client confidential info removed from prompt logs?
• Output labeled DRAFT — REQUIRES ATTORNEY REVIEW?

Spot-check: randomly select 3 quotes — full-text search in source document.

Multi-document contract review

Base agreement + amendments:

<documents>

<doc id="A" label="Master Agreement">...</doc>

<doc id="B" label="Amendment 1">...</doc>

</documents>

<rules>

Later doc supersedes earlier on conflict. Cite both when amended.

Final value table: term | current value | source doc | quote

</rules>

Gemini 3.1 Pro handles multi-doc context; Claude Opus 4.8 for final merge accuracy.

JSON output for matter management tools

For pipeline integration with GPT-5.5 structured output:

Schema: { parties: [{name, role, section}], terms: [{topic, quote, section, value}], open_questions: [string], not_specified: [string] }

Same citation rules. Invalid if quote not substring of document.

Redlines and comparison prompts

<document_a label="Draft v3">...</document_a>

<document_b label="Draft v5">...</document_b>

<task>Material changes only — not punctuation</task>

<output>

Table: topic | v3 quote | v5 quote | materiality (high/medium/low)

</output>

Attorney confirms materiality ratings — AI proposes, human disposes.

Privilege and work product

Do not use AI summaries to create or replace privilege logs without partner review.

Prompts about strategy ("weakest argument") are work product — treat input as sensitive.

Enterprise Claude/OpenAI with zero retention ≠ privilege protection — firm policy governs.

Model routing

Claude Opus 4.8: primary for contract and opinion fidelity

GPT-5.5: JSON extraction pipelines, structured tables

Gemini 3.1 Pro: very long appendices + main agreement in one context

Gemini 3.5 Flash: defined term candidate list only — verify on Opus

Never ship Flash-only summary on binding terms without Opus second pass.

Common failures

"Summarize" without MUST EXTRACT list → omits liability cap

No glossary pass → wrong meaning of defined terms

Allowing paraphrase on numbers → 30 days becomes "about three months"

External knowledge enabled → cites wrong statute

Skipping verification because summary "reads well"

Pasting client email threads with PII into public Claude

PromptMake workflow

PromptMake /text: paste MUST EXTRACT list + document type → generate XML-tagged prompt scaffold → run in firm-approved Claude → attorney verification.

Guests: 3 generations/day. Registered: 5/day. Do not paste confidential matter text into PromptMake public tier.

FAQ

Is AI legal summarization safe for client work?

Only with firm policy-compliant tools, no confidential data in public tiers, and mandatory attorney review. This article is not legal advice.

Why Claude for legal summaries?

Claude Opus 4.8 follows long citation rules and XML structure reliably. Always verify — no model is infallible.

What must never appear in a public AI prompt?

Client names, identifiable facts, privileged content, sealed records, and trade secrets — unless using approved confidential AI environment.

How do I handle missing clauses?

Output Not specified in document. Do not ask model what "should" be there.

Can AI compare two contract versions?

Yes for first-pass diff tables. Attorney marks what is actually material.

Should I use JSON or prose output?

Prose for human review memos. JSON for matter management integration — same quote rules either way.

What if the model adds legal recommendations?

Ban in rules: No recommendations, risk ratings, or strategic advice. Strip and re-prompt if violated.

How long can the document be?

Gemini 3.1 Pro and Claude Opus 4.8 handle long agreements; split by exhibit if over context with cross-reference index prompt first.

Does Not specified apply to implied terms?

Yes — if not explicit in provided text, Not specified. Implied terms require attorney analysis outside AI.

Can associates skip reading the source?

No. AI summary is index and draft only. Verification checklist is mandatory.

NDA summary prompt

<document>[NDA text]</document>

<must_extract>Parties, confidential information definition, term, exclusions, return/destruction, injunctive relief, governing law</must_extract>

<rules>Quote definition of Confidential Information verbatim. Not specified if mutual vs one-way unclear from text alone.</rules>

NDAs fail when model paraphrases narrow definition — quote-first is critical.

Employment agreement prompt

<must_extract>Compensation, at-will language quote, non-compete (duration/geography), IP assignment, arbitration, termination benefits</must_extract>

Flag: Non-compete enforceability varies by jurisdiction — prompt: do not assess enforceability; quote clause only.

SaaS MSA prompt

<must_extract>SLA uptime, credit cap, data processing, subprocessor list, audit rights, auto-renewal, price increase notice</must_extract>

SaaS tables often split terms across Order Form + MSA — multi-doc prompt required.

Deposition transcript summary

<task>Witness outline for attorney prep — not trial advocacy</task>

<rules>Attribute statements to witness. Label QUESTION vs ANSWER. Not specified if witness unclear. No credibility assessments.</rules>

<output>Topic headings | page:line refs | key quotes</output>

Regulatory excerpt summary

For compliance teams reviewing SEC/FDA excerpt paste:

<rules>No external regulatory knowledge. Quote only from provided excerpt. Not specified if obligation conditional on facts not in text.</rules>

Attorney confirms applicability to entity — AI extracts text obligations only.

Worked example: indemnification clause

Weak output: "Vendor indemnifies customer for third-party claims."

Strong output: Quote full indemnity scope, carve-outs, cap cross-reference §8.2, procedure quote — or Not specified per element.

Attorney spot-check: compare cap cross-ref exists in §8.2.

Splitting long agreements

Step 1 — exhibit and section index:

<task>List all sections and exhibits with heading + first sentence</task>

Step 2 — per-section MUST EXTRACT passes merged by reduce prompt with conflict rules.

Gemini 3.1 Pro for index; Claude Opus 4.8 for clause extraction on critical sections.

Cross-border contracts

Note governing law and currency in constraints header.

Prompt: Do not convert currencies. Quote amounts in original currency.

Defined terms in translation: if bilingual doc, specify which language controls.

Associate workflow integration

1. Associate runs glossary + MUST EXTRACT prompt
2. Checklist verification on numbers
3. Partner reviews open questions only
4. Summary stored as DRAFT with prompt version logged

Saves partner time on extraction — not on judgment.

Related articles

Summarization prompts without losing details — citation and MUST KEEP patterns

Structured output prompting — JSON schemas for clause tables

Positive framing — Not specified vs don't guess

Litigation teams should store prompt version, model ID, and timestamp alongside every DRAFT summary in the matter file. Reproducibility matters when a partner questions whether a number in a memo came from the model or the contract.

For cross-border deals, state governing law, currency, and controlling language in the prompt header before any MUST EXTRACT pass — bilingual documents fail when the model assumes English controls undefined terms.

Bottom line

Legal AI summarization demands quote-first extraction, defined terms discipline, Not specified for gaps, XML structure, and attorney verification.

Not legal advice. No confidential data in public tools. Claude Opus 4.8 for fidelity. Human checklist before any summary influences a decision.