Legal Tech12 min read

AI Contract Review in Singapore: A Claude Prompt Playbook

Production-tested Claude prompts for SPA, NDA, and shareholder-agreement review. PDPA-safe redaction. Aligned with MinLaw's GenAI guidance for the legal sector.

Haojun See

Founder & Director, On The Ground

Updated 1 May 2026

Where Claude actually helps

Contract review breaks into three distinct tasks: extraction (what does this document say), comparison (how does it differ from our standard, or from a benchmark), and judgement (is this acceptable given the deal, the client, the jurisdiction). Claude is excellent at extraction and comparison. It can summarise a 60-page SPA in 5 minutes. It can flag every deviation from your firm's standard NDA. It can produce a full clause-by-clause comparison table. Claude is partially useful for judgement. It can apply rules you give it ("flag any indemnity cap below 12 months' fees") but it cannot replace the partner-level judgement of whether *this particular* indemnity cap is acceptable for *this particular* client in *this particular* deal. The Singapore Ministry of Law's Guide for Using Generative AI in the Legal Sector is explicit on this: human accountability sits with the lawyer, not the model. AI assists; the lawyer decides.

Extraction prompts that work

Standard extraction: *"You are reviewing a [type of contract, e.g. SPA] under Singapore law. Extract the following into a structured table: parties, governing law, term, termination triggers, payment / consideration, indemnity scope and cap, liability cap and exclusions, IP assignment, confidentiality scope and tail, dispute resolution forum and seat, change-of-control trigger. For each, quote the operative clause. Cite the section / clause number. If a field is not present, say 'not present' — do not invent."* The "do not invent" instruction matters. Without it, Claude will sometimes hallucinate a default. With it, the failure mode shifts from false-positive to clearly-flagged-missing. Deviation extraction: *"Compare the attached SPA to the standard SPA in /standards/spa-template.md. List every clause-level deviation: what the standard says, what the document says, the practical effect of the difference, and a suggested negotiation position. Mark each as material / non-material. Do not opine on the merits — just flag."*

Comparison and benchmarking prompts

Cross-document comparison: When reviewing a stack of similar agreements (e.g. all NDAs from a counterparty's portfolio), comparison-style prompts produce strong output: *"Below are five NDAs we've signed with [counterparty type]. Build a comparison table: term length, definition of confidential information, exclusions, return obligations, jurisdiction. Identify the version most favourable to us and the version least favourable. Note which version we should propose as our standard going forward."* Benchmark against market standard: *"Review this employment agreement against typical Singapore employment market terms (notice periods 1–3 months, confidentiality 12–24 months tail, non-compete enforceability is narrow under Mano Vikrant Singh v Cargill). Flag any term that is materially outside market range, with a brief explanation."* For more prompt patterns, see Claude for Legal Research and The Singapore Prompt Library.

PDPA-safe workflows

Three patterns we use with SG clients: Pattern 1 — Pre-redact. Before any prompt, replace personal names with [PARTY_A], [PARTY_B]; replace specific monetary figures with [AMOUNT]; replace dates with [DATE]. The legal analysis is rarely sensitive to those identifiers. Pattern 2 — Use enterprise tier. Anthropic's enterprise plans guarantee no training on input. Standard practice for any firm processing PDPA-protected data routinely. Pattern 3 — On-device for the sensitive class. For matters where even enterprise-tier cloud LLMs are unacceptable (e.g. matrimonial, criminal defence, high-profile commercial), use on-device tooling. OTG Legal Box is built for this — ten legal tools running entirely on the firm's hardware. Always document the AI use in the matter file. The MinLaw guide does not require disclosure to the client in every case but does expect the firm to maintain its own audit trail. For the full eight-rule checklist, see PDPA-Safe Claude Prompts: A Lawyer's Checklist.

What Claude gets wrong

Three failure modes we see repeatedly. Quantitative reasoning. Claude can sometimes miscalculate or miscompare percentages and dates. Always verify the math. Singapore-specific case law. Claude knows broad common law principles but its training data is uneven on recent SG cases. For citations, verify against LawNet or SingaporeLaw.sg. Don't paste case-law summaries from Claude into a memo without verification. Strategic judgement. "Is this clause reasonable?" depends on the client, the deal, the leverage, and the lawyer's read of the counterparty. Claude can give the textbook answer; the textbook isn't always right for *this* situation.

Next steps

For a structured workflow on document review at scale, see AI E-Discovery in Singapore: 5-Step Claude Workflow. For drafting (not just reviewing), see Claude for Legal Research. OTG runs vibe coding workshops for Singapore law firms — partners and associates spend a day building tools they'll use the next week. Book a free 30-minute call to discuss training for your firm or a custom contract-review tool built as a Functional App Sprint.

Frequently asked questions

Can Singapore lawyers ethically use Claude for contract review?

Yes, with care. The Singapore Ministry of Law's [Guide for Using Generative AI in the Legal Sector](https://www.mlaw.gov.sg/files/Guide_for_Using_Generative_AI_in_the_Legal_Sector.pdf) sets out the obligations: human accountability, accuracy verification, confidentiality protection, and disclosure where appropriate. AI-assisted review is allowed; AI-only review without lawyer verification is not.

Is it PDPA-compliant to upload client contracts to Claude?

Use Claude Pro/Team/Enterprise tiers (no training on input). Redact personal identifiers and commercially sensitive figures before pasting. For consistently confidential workflows, consider on-device deployment via [OTG Legal Box](/apps/legal-box). See our [PDPA Prompting Checklist](/resources/pdpa-prompting-checklist).

Can Claude replace a junior associate for due diligence?

It can replace 60–80% of the first-pass extraction work. Senior verification of conclusions, novel issue spotting, and judgement on materiality remain a human responsibility under the MinLaw guide.

How does Claude compare to Harvey or Spellbook for Singapore lawyers?

Harvey and Spellbook are specialised legal-AI products with curated playbooks. Claude is general-purpose but more flexible and substantially cheaper. For high-volume standardised review, the specialised tools have edges. For smaller firms or bespoke matters, Claude with good prompts often wins on price-performance.

What clauses should I always have Claude flag?

Indemnity, liability cap, term and termination, change-of-control, IP assignment, governing law, dispute resolution, payment terms, and any clause referencing 'reasonable efforts' (these are notorious for ambiguity). Our [Singapore Prompt Library](/resources/singapore-claude-prompt-library) has a full checklist.

Previous7 Internal Tools Every Singapore SME Should Build with Claude Code Next PDPA-Safe Claude Prompts: A Singapore Lawyer's Checklist

Want to Apply This to Your Business?

We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.

Get in Touch View our services