Prompt Patterns: Roles, Constraints & Rubrics — A Complete Guide

The 7 Essential Prompt Patterns for Business

Prompt engineering is not about memorising magic phrases. It is about understanding a set of patterns that consistently produce better results. These 7 patterns work across all AI tools — ChatGPT, Claude, Copilot, Gemini — and all business contexts.

Pattern 1: Role Prompting

What it is: Assign the AI a specific expert persona before giving your request.

Why it works: AI produces more relevant, detailed outputs when it has a clear perspective to adopt. A "senior HR consultant" gives different advice than a generic AI.

Template:

You are a [specific role] with [years] of experience in [domain]. You specialise in [specialisation]. Your audience is [who you're writing for]. [Your actual request]

Business examples:

• "You are a CFO with 20 years of experience in Singapore financial services..."
• "You are a compliance officer specialising in PDPA and data protection..."
• "You are a management consultant who advises mid-size companies on AI adoption..."

When to use: Always. Role prompting should be your default starting point.

Pattern 2: Constraint-Based Prompting

What it is: Set explicit boundaries on format, length, tone, scope, and content.

Why it works: Without constraints, AI tends to produce generic, meandering outputs. Constraints force precision and relevance.

Template:

[Request]. Constraints:

• Format: [table/bullets/paragraphs/numbered list]
• Length: [word count or page count]
• Tone: [formal/conversational/technical/executive]
• Audience: [who will read this]
• Must include: [required elements]
• Must exclude: [elements to avoid]

Business examples:

• "Maximum 200 words, written at a Grade 8 reading level"
• "Output as a table with 5 columns: [specify columns]"
• "Use British English spelling, formal tone, no jargon"

When to use: When you need control over the output format.

Pattern 3: Chain-of-Thought

What it is: Instruct the AI to reason through a problem step by step before giving an answer.

Why it works: Forces the AI to show its reasoning, which produces more accurate and defensible conclusions. Also makes it easier to spot where the logic goes wrong.

Template:

[Question/Task]. Think through this step by step:

1. First, [initial analysis]
2. Then, [secondary consideration]
3. Next, [evaluation]
4. Finally, [conclusion/recommendation]

When to use: Complex analysis, financial calculations, strategic decisions, and any situation where the reasoning matters as much as the conclusion.

Pattern 4: Few-Shot Prompting

What it is: Provide 1-3 examples of the output quality and format you expect before asking for new content.

Why it works: Examples communicate quality standards more effectively than descriptions. The AI matches the pattern, tone, and detail level of your examples.

Template:

Here are examples of what I want:

Example 1: [input] → [desired output] Example 2: [input] → [desired output]

Now do the same for: [new input]

When to use: When consistency across multiple outputs matters (job descriptions, product descriptions, email templates, social media posts).

Pattern 5: Rubric-Based Prompting

What it is: Provide explicit evaluation criteria and ask the AI to assess against them.

Why it works: Produces structured, fair evaluations. Reduces subjectivity and ensures completeness.

Template:

Evaluate [subject] against this rubric:

Criterion	Weight	1 (Poor)	3 (Good)	5 (Excellent)
[criterion 1]	[%]	[description]	[description]	[description]
[criterion 2]	[%]	[description]	[description]	[description]

For each criterion, provide: score, evidence, and recommendation.

When to use: Vendor evaluations, performance reviews, proposal assessments, quality audits.

Pattern 6: Comparative Analysis

What it is: Ask the AI to compare options side-by-side against defined criteria.

Template:

Compare [Option A] vs [Option B] vs [Option C] on these dimensions:

1. [Dimension 1]
2. [Dimension 2]
3. [Dimension 3] For each dimension: explain the key differences, rate each option (1-5), and identify the clear winner. Conclude with an overall recommendation.

When to use: Technology selection, strategic alternatives, vendor comparison, policy options.

Pattern 7: Iterative Refinement

What it is: Start with a broad request, then refine through follow-up prompts.

Template — Round 1: Generate the first draft. Round 2: "Make it more concise — cut to 300 words" Round 3: "Add specific data points and examples for the Malaysia market" Round 4: "Change the tone from academic to conversational" Round 5: "Format as a table with action items"

When to use: Always. Rarely is the first output perfect. Plan for 2-4 rounds of refinement.

Combining Patterns

The most effective prompts combine multiple patterns:

Role + Constraint + Chain-of-Thought: You are a senior management consultant. Analyse whether Company X should enter the Indonesian market. Think step by step. Maximum 500 words. Output as: Executive Summary, Analysis, Recommendation.

Role + Few-Shot + Rubric: You are an HR director. Evaluate these 3 candidates using the scorecard below. Here is an example of how I want each evaluation structured: [example]. Now evaluate: [candidates].

Quick Reference Card

Related Reading

• Prompting Structured Outputs — Get consistent, formatted results from AI tools
• Prompting Evaluation and Testing — Systematic approaches to testing prompt effectiveness
• ChatGPT Output Evaluation — How to evaluate and improve the quality of AI outputs

Seven Reusable Pattern Templates for Enterprise Workflows

Prompt patterns function as standardized blueprints that teams can adapt across departments without rebuilding instructions from scratch. The following templates emerged from Pertama Partners workshop facilitation across organizations in Singapore, Malaysia, Indonesia, Thailand, and Vietnam throughout 2025.

Pattern 1 — Persona-Context-Task (PCT). Structure: "You are a [professional role] with expertise in [domain]. Given [contextual background], please [specific deliverable]." This pattern consistently produces more relevant outputs from ChatGPT Enterprise, Claude Teams, and Google Gemini than unstructured requests. Effectiveness research published by Microsoft Research in August 2025 demonstrated thirty-seven percent improvement in output relevance scores when persona framing preceded task specification.

Pattern 2 — Chain-of-Verification (CoV). After generating initial output, append: "Now verify each factual claim in your response. List any statements that cannot be confirmed from your training data and flag them explicitly." This meta-cognitive pattern reduces hallucination rates by approximately forty-two percent according to benchmarking conducted by Stanford HAI Laboratory in November 2025.

Pattern 3 — Comparative Analysis Matrix. Template: "Create a comparison table evaluating [items] across these dimensions: [criteria list]. Include a recommendation column explaining tradeoffs for [specific audience]." Particularly effective for procurement evaluations, vendor assessments, and strategic planning scenarios.

Pattern 4 — Iterative Refinement Loop. Initial prompt followed by: "Rate your response on a scale from one to ten for [quality criteria]. Identify the weakest section and rewrite it with improved [specific attribute]." This self-evaluation mechanism leverages model introspection capabilities available in Claude Sonnet, GPT-4o, and Gemini Advanced.

Pattern 5 — Constraint Specification Framework. Template: "Generate [deliverable] following these constraints: maximum [word count] words, reading level appropriate for [audience], tone matching [reference example], excluding [prohibited elements], formatted as [output structure]." Explicit boundary definition prevents scope creep in generated outputs.

Pattern 6 — Few-Shot Exemplar Anchoring. Provide three carefully curated examples before the target request: "Here are three examples of [deliverable type] that match our quality standards: [Example 1] [Example 2] [Example 3]. Now create a new one addressing [specific topic] following the same structure and quality characteristics." DeepLearning.AI coursework published by Andrew Ng's team validated that three exemplars outperform both single-shot and five-shot configurations for most business document generation tasks.

Pattern 7 — Audience Adaptation Cascade. Template: "Rewrite the following content for three distinct audiences: (1) C-suite executives requiring strategic implications in under two hundred words, (2) department managers needing implementation details and timeline estimates, (3) frontline employees wanting practical step-by-step guidance." This single prompt generates three calibrated versions from one source document, reducing content production time by approximately sixty percent across corporate communications workflows.

Advanced prompt pattern taxonomies formalized through Vanderbilt University's software engineering research group identify twenty-three distinct compositional strategies organized across output customization, error identification, prompt improvement, interaction, and context control categories. The Persona Pattern assigns domain-expert identities (chartered financial analyst, maritime arbitration barrister, avian epidemiologist) shaping vocabulary register and inferential reasoning pathways. The Flipped Interaction Pattern inverts conventional query-response dynamics, enabling the model to elicit requirements through Socratic questioning sequences before generating deliverables. Template Patterns employ placeholder-delimited skeletal structures using Mustache, Handlebars, or Jinja2 templating syntax conventions ensuring output structural consistency across iterative generation cycles. Cognitive Verifier Patterns decompose ambiguous queries into constituent sub-questions referencing Bloom's Revised Taxonomy cognitive process dimensions before synthesizing comprehensive responses. Chain-of-Thought variations including zero-shot-CoT, few-shot-CoT, and Tree-of-Thoughts branching exploration architectures represent distinct computational reasoning strategies benchmarked against GSM8K mathematical reasoning, StrategyQA multi-hop retrieval, and HumanEval code generation evaluation suites. Meta-prompting patterns incorporating self-consistency voting, universal self-adaptive prompting, and skeleton-of-thought parallel decomposition leverage ensemble diversity principles from Dietterich's machine learning methodology. Practitioners across pharmaceutical regulatory affairs, actuarial reserving, and petroleum reservoir engineering verticals construct domain-specific pattern libraries curated through Obsidian knowledge management vaults, Notion relational databases, and Roam Research bidirectional linking architectures. ReAct (Reasoning and Acting) patterns interleave deliberative reasoning traces with tool-invocation action steps, enabling grounded information retrieval through Serper, Tavily, and Exa search API integrations preventing confabulatory drift during complex analytical workflows.