Prompts Library — Guide

Prompt Engineering Guide 2026

Master the six core techniques that separate mediocre AI outputs from exceptional ones. Each technique includes an explanation, when to use it, and a ready-to-use example you can adapt.

1. Role-Playing (Persona Assignment)

Assigning a specific expert persona to the AI dramatically changes the depth, vocabulary, and perspective of its response. The more specific the role, the better the output.

When to use: Whenever you need domain-specific expertise, a particular tone, or a specialized perspective. Works for everything from legal analysis to creative writing.

Key principle: Specify the expert’s experience level, their communication style, and what they prioritize. “You are a senior product manager at a Series B SaaS company” is far better than “You are a product manager.”

You are a senior data scientist with 15 years of experience at a Fortune 500 company. You specialize in explaining complex statistical concepts to non-technical stakeholders. You always use real-world business analogies and avoid jargon. When you must use a technical term, you define it in parentheses. Your communication style is clear, patient, and structured. Now explain [CONCEPT] to a marketing team who needs to understand it for a board presentation.

2. Chain-of-Thought Reasoning

Asking the AI to “think step by step” before answering improves accuracy on complex problems by 40-70%. This technique forces the model to show its reasoning rather than jumping to conclusions.

When to use: Math problems, logical reasoning, multi-step analysis, strategic decisions, debugging, and any task where the reasoning matters as much as the answer.

Key principle: Explicitly instruct the AI to break down its thinking. Phrases like “think step by step,” “show your reasoning,” and “before answering, consider…” activate this pattern.

I need to decide whether to [DECISION]. Before giving your recommendation, think through this step by step: Step 1: List all the relevant factors to consider. Step 2: Evaluate each factor as favoring Option A or Option B. Step 3: Identify which factors carry the most weight and why. Step 4: Consider what could go wrong with each option. Step 5: Give your final recommendation with a confidence level. Context: [PROVIDE DETAILS]

3. Few-Shot Learning (Examples)

Providing 2-3 examples of the desired input/output format teaches the AI your exact expectations far more effectively than lengthy descriptions. Show, don’t just tell.

When to use: When you need a specific output format, consistent tone, or a pattern the AI should replicate. Essential for classification, formatting, and style matching tasks.

Key principle: Your examples should demonstrate the pattern, not just the content. Include examples that cover different scenarios so the AI understands the range of expected outputs.

Convert customer feedback into structured insights. Follow this exact format: Input: “The app crashes every time I try to export a PDF. This is really frustrating and I’m considering switching to a competitor.” Output: Category: Bug | Severity: High | Feature: PDF Export | Sentiment: Negative | Churn Risk: Yes | Action: Escalate to engineering Input: “Love the new dashboard! The charts are much clearer now. Would be great if I could customize the date range though.” Output: Category: Feature Request | Severity: Low | Feature: Dashboard | Sentiment: Positive | Churn Risk: No | Action: Add to backlog Now process these customer messages: [PASTE MESSAGES]

4. System Prompts (Persistent Instructions)

System prompts set the AI’s behavior for an entire conversation. They define who the AI is, what rules it follows, and how it formats responses — before any user message arrives.

When to use: Building chatbots, Custom GPTs, API integrations, or any scenario where you want consistent behavior across multiple interactions.

Key principle: Structure system prompts with clear sections: identity, rules (always do / never do), output format, and boundaries. Bullet points work better than paragraphs for rules.

[SYSTEM PROMPT] You are a customer onboarding assistant for [PRODUCT NAME]. IDENTITY: – Name: [ASSISTANT NAME] – Tone: Helpful, concise, encouraging RULES: – Always greet new users by name if provided – Break instructions into numbered steps (max 5 per response) – If a user asks about pricing, direct them to [PRICING URL] – Never make up features that do not exist – If you are unsure about an answer, say “Let me connect you with our support team” and provide [SUPPORT EMAIL] OUTPUT FORMAT: – Use short paragraphs (2-3 sentences max) – Bold key actions the user needs to take – End every response with a clear next step or question

5. Temperature and Sampling Tips

Temperature controls randomness in AI outputs. Understanding when to adjust it is the difference between reliable factual answers and creative brainstorming sessions.

Temperature scale: 0.0 = most deterministic and focused. 0.7 = balanced (default for most models). 1.0+ = highly creative and unpredictable.

Key principle: Match temperature to the task. Use low temperature (0.0-0.3) for factual queries, code generation, data extraction, and classification. Use medium (0.5-0.7) for general writing and summarization. Use high (0.8-1.0) for brainstorming, creative writing, and generating diverse options.

WHEN USING THE API, set temperature based on your task: For data extraction (temp: 0.0): “Extract all dates, dollar amounts, and company names from this contract. Return as a JSON array. Be precise — do not infer data that is not explicitly stated.” For balanced writing (temp: 0.5): “Write a professional product announcement for [PRODUCT]. Follow our brand voice: confident but not boastful, clear but not simplistic.” For creative brainstorming (temp: 0.9): “Generate 20 wildly different tagline ideas for [BRAND]. Push boundaries. Include serious, playful, provocative, and unexpected options. Quantity over quality at this stage.”

6. Output Formatting and Constraints

Specifying the exact output format eliminates the biggest source of AI frustration: getting the right information in the wrong structure. Tell the AI exactly what shape you want the answer in.

When to use: Always. Every prompt should include some formatting guidance. This is the most universally applicable technique on this list.

Key principle: Be explicit about format (JSON, markdown, table, bullet points), length constraints (word count, number of items), and structure (sections, headers, ordering). If you want a table, describe the columns. If you want JSON, show the schema.

Analyze the competitive landscape for [PRODUCT] and return your analysis in this exact format: ## Market Overview [2-3 sentences on market size and trajectory] ## Competitor Matrix | Competitor | Pricing | Key Strength | Key Weakness | Threat Level | |————|———|————–|————–|————–| [Fill 5 rows] ## Our Positioning – **Differentiator 1:** [One sentence] – **Differentiator 2:** [One sentence] – **Differentiator 3:** [One sentence] ## Recommended Actions 1. [Immediate action – this week] 2. [Short-term action – this month] 3. [Strategic action – this quarter] Keep the total response under 500 words.

Frequently Asked Questions

Which technique should I learn first?

Start with Output Formatting — it applies to every single prompt and gives you immediate improvements. Then add Role-Playing for domain expertise. Chain-of-Thought is essential for complex tasks. Few-Shot Learning comes naturally once you start recognizing patterns you want to replicate.

Can I combine multiple techniques in one prompt?

Absolutely, and you should. The best prompts typically combine 2-3 techniques. For example: assign a role (technique 1), ask for step-by-step reasoning (technique 2), and specify the output format (technique 6). Start simple and layer techniques as needed.

Do these techniques work across all AI models?

Yes, these are universal prompting principles that work with ChatGPT, Claude, Gemini, Llama, and other major models. The specific syntax may vary slightly, but the core concepts — context, examples, constraints, and reasoning instructions — improve output quality across all language models.