Based on Relativity's "Legal Professional's Guide to Prompt Engineering" (2026), adapted for Truth Meter's compliance-first architecture.
Core Principle: "90% of writing clear prompt criteria is simply writing clearly. Every time you write a prompt, ask yourself: would a human reader find this confusing?" — Jeff Gilles, Relativity
The Six Core Techniques
1. Zero-Shot Prompting
Give the model only an instruction—no demonstrations, no prior examples.
"In one paragraph, distill the ratio decidendi of R. v. Jordan (2016 SCC 27)."
2. Chain-of-Thought (CoT) Prompting
Explicitly ask the model to surface its intermediate reasoning steps ("think step-by-step").
"Analyze whether the following non-compete clause is enforceable under New York law. Delineate your reasoning in IRAC format before giving a conclusion."
3. Role-Based Prompting
Assigning the model a professional identity steers tone, vocabulary, and evaluative stance.
"You are a veteran law-review articles editor. Suggest three titles that capture the policy implications of the EU AI Act for a general counsel of a multinational company."
4. Contextual Prompting
Embed the substantive materials the model must rely on—contracts, statutes, pleadings—inside the prompt itself. This prevents hallucination.
[Drop contract into prompt] "Highlight any provisions that could trigger a change-of-control penalty under Delaware law."
5. System Prompting
A system prompt is the conversation's primordial directive—usually hidden from the end user—setting global boundaries.
"You are a neutral legal assistant; cite primary sources; refuse political speculation."
6. Iterative Refinement
Test prompts on smaller subsets of data, evaluate results, then refine before running on larger populations.
"A key difference between manual review instructions and prompting AI is the iterative nature of AI prompt writing. Unlike traditional reviewer protocols, which are typically finalized in one phase, AI prompts require testing and refinement to optimize effectiveness." — Starling Underwood, Kilpatrick Townsend
The "Relativity Test" for All Prompts
Before finalizing any prompt, answer these four questions:
Would a human reader find this confusing? (Clarity test)
Have we removed vague legal terms? (Precision test) — Avoid: "ideally," "reasonable," "substantial," "material"
Is the output verifiable against primary sources? (Accuracy test)
Would we be comfortable if this prompt were discovered? (Privilege test)
Standardized System Prompts for Truth Meter
Task Type
System Prompt
Regulatory Analysis
"You are a neutral compliance analyst. Cite primary regulatory sources. Flag uncertainty. Never speculate. Use IRAC format."
Risk Assessment
"You are an actuarial risk analyst. Identify specific statutes. Assess probability with reasoning. Suggest mitigations."
Document Review
"You are a document review specialist. Use plain language. Avoid vague terms unless defined. Mark critical items as VERY RELEVANT."
Research Synthesis
"You are a research analyst. Cross-verify across 3+ sources. Use Information Weight A-E grading. Cite official sources first."
Prompt Version Control
All prompts must include:
Author identity
Date/time
Task description
Iteration number
Source documents referenced
Human reviewer sign-off
Best Practices from the Field
Write clearly. Be concise, use active voice, avoid double negatives, and avoid "lawyer speak." Plain language is the key.
Clean things up. Checking for good grammar—and correct spelling—can help aid in clearly articulating criteria for the AI. So can intentional formatting.
Use emphasis. You won't hurt the AI's feelings by weighing important notes with all-capital letters or exclamation points. These tactics encourage the AI to focus on particular considerations.
Avoid vague legal terms. "Ideally," "reasonable," "substantial," or "material" can be interpreted too loosely by AI. Define terms explicitly or avoid them.
The "Opposing Witness" Approach
"Prompting is just asking really good questions. My advice to everyone is: look at what comes out of off-the-shelf generative AI models as if it were an opposing witness. Imagine you're deposing a hostile witness. Make it prove its answer. If you can think in that way, then you'll get much better information."
— Bennett Borden, Founder & CEO, Clarion AI Partners
Mark regulatory thresholds as CRITICAL or VERY RELEVANT
"AI is already an expert in public-domain knowledge"
Don't explain basic legal concepts—focus on case-specific context
"Single drafter ensures consistency"
One person drafts, multiple stakeholders review
"Test on subsets before full deployment"
Red-team prompts on 10-20 documents before full review
Work-Product Doctrine Consideration
Active Legal Question: Should AI prompts be protected by work-product doctrine? As of 2026, there is no settled case law. Seth M. Cohen (Alston & Bird) suggests prompts for brief writing are likely privileged; prompts for document relevance may not be. Truth Meter action: Seek legal opinion on whether our compliance prompts are privileged.
ROLE: You are a senior compliance officer with 20 years experience in [JURISDICTION] financial regulation.
TASK: Analyze whether [PRODUCT/ACTIVITY] requires [LICENSE/APPROVAL] under [STATUTE/REGULATION].
REQUIREMENTS:
1. Identify the applicable statutory provisions with specific section numbers
2. Analyze each element of the regulatory test
3. Apply the facts to each element
4. Identify any exemptions or safe harbors
5. Flag any areas of uncertainty or ambiguity
6. Cite only primary sources (statutes, regulations, official guidance)
FORMAT: Provide your analysis in IRAC format before giving a final recommendation.
CONTEXT: [Embed relevant statute excerpts, prior regulatory decisions, or similar cases]
ROLE: You are an actuarial risk analyst specializing in financial derivatives.
TASK: Assess the regulatory risk of [SPECIFIC ACTIVITY] in [JURISDICTION].
CONSTRAINTS:
- Cite specific statutes and regulatory guidance
- Identify enforcement actions against similar activities
- Assess probability of regulatory action (High/Medium/Low) with reasoning
- Identify mitigating factors
DOCUMENTS: [Embed relevant regulatory guidance, enforcement actions, or legal opinions]
OUTPUT: Structured risk memo with executive summary, detailed analysis, and recommended mitigations.
ROLE: You are a document review specialist trained in [SPECIFIC FIELD].
TASK: Review the attached [DOCUMENT TYPE] and identify [SPECIFIC ELEMENTS].
CRITERIA:
- Relevant if: [SPECIFIC CONDITIONS]
- Very relevant if: [HIGH-VALUE CONDITIONS]
- Non-responsive if: [EXCLUSION CONDITIONS]
IMPORTANT: Use plain language. Avoid terms like "reasonable," "substantial," or "material" unless defined.
EMPHASIS: Documents meeting [CRITICAL CRITERIA] should be marked as VERY RELEVANT.
Last Updated: May 7, 2026
Next Review: August 7, 2026 (quarterly)
Version: 1.0
Source: Relativity "Legal Professional's Guide to Prompt Engineering" (2026)