Prompt Engineering Fundamentals

What Is Prompt Engineering?

Prompt engineering is the practice of crafting effective instructions for AI language models to get accurate, useful, and relevant outputs. Think of it as learning to communicate clearly with a highly capable but literal assistant — the quality of what you get out is directly tied to the quality of what you put in.

As AI models like Claude, GPT-4, and Gemini have become more powerful, the skill of writing good prompts has become one of the most valuable abilities in the modern workplace. Whether you're drafting emails, analyzing data, generating reports, or brainstorming ideas, better prompts lead to dramatically better results.

Why This Matters

Studies from organizations like McKinsey and Boston Consulting Group have found that workers who effectively use AI tools can be 20–80% more productive depending on the task. The difference between average and exceptional AI usage almost always comes down to how well you prompt.

The CRAFT Framework

CRAFT is a structured approach to writing prompts that consistently produce high-quality outputs. Each letter stands for a key element you should consider when writing a prompt. You don't need to use every element in every prompt, but thinking through each one helps you craft more effective instructions.

C — Context

Context is the background information the AI needs to understand your situation. Without context, the AI has to guess about your circumstances, audience, and constraints — and those guesses are often wrong.

Weak prompt:

"Write an email about the project delay."

Strong prompt with context:

"I'm a project manager at a software company. Our mobile app launch has been delayed by 2 weeks due to unexpected QA issues. I need to inform our client, Acme Corp, who has been a partner for 3 years. They're expecting the launch next Friday."

R — Role

Assigning a role tells the AI what perspective or expertise to bring to the task. This shapes the vocabulary, depth, and approach of the response.

Example roles:

"Act as an experienced marketing strategist."
"You are a patient math tutor for middle school students."
"Respond as a senior financial analyst reviewing quarterly earnings."
"You are a technical writer creating documentation for non-technical users."

Role Stacking

You can combine roles for nuanced responses. For example: "Act as a marketing strategist who also has deep expertise in behavioral psychology." This gives the AI permission to draw from multiple domains.

A — Action

The action is the specific task you want the AI to perform. Be explicit and precise about what you want. Vague actions lead to vague outputs.

Vague action:

"Help me with my presentation."

Specific action:

"Create an outline for a 15-minute presentation with 8–10 slides. Include speaker notes for each slide and suggest one data visualization per section."

F — Format

Format tells the AI how to structure the output. This is one of the most impactful elements because it determines whether the response is immediately usable or requires heavy editing.

Common format instructions:

"Format as a numbered list with bold headers."
"Present this as a comparison table with pros and cons."
"Write this as a 3-paragraph executive summary, no more than 200 words."
"Structure as: Problem → Analysis → Recommendation → Next Steps."
"Output as JSON with keys for name, category, and priority."

T — Tone

Tone sets the voice and style of the response. The same information can be delivered in wildly different ways depending on tone, so matching tone to your audience is critical.

Tone examples:

"Use a professional but approachable tone, as if writing to a colleague."
"Keep the language simple and jargon-free — this is for a general audience."
"Write in a formal, academic style with citations."
"Be concise and direct. No filler. Every sentence should carry weight."

Putting CRAFT Together

A complete CRAFT prompt might look like: "You are a senior UX researcher [Role]. I'm redesigning the checkout flow for our e-commerce platform that serves 50,000 monthly users [Context]. Analyze these 3 user complaints and identify the top usability issues [Action]. Present findings as a prioritized list with severity ratings (high, medium, low) and a one-sentence recommendation for each [Format]. Use a clear, data-driven tone suitable for a product team meeting [Tone]."

Prompting Techniques

Zero-Shot Prompting

Zero-shot prompting means asking the AI to perform a task without providing any examples. You rely entirely on the model's training to understand what you want. This works well for straightforward tasks where the expectation is clear.

Zero-shot example:

"Classify the following customer review as positive, negative, or neutral: 'The product arrived on time but the packaging was damaged. The item itself works fine though.'"

Zero-shot works best when the task is common (summarization, classification, translation), the desired output format is obvious, and you don't need the AI to follow a specific pattern.

Few-Shot Prompting

Few-shot prompting means providing one or more examples before asking the AI to perform the task. The examples teach the model the pattern, format, and style you expect. This is one of the most powerful techniques available because it lets you "show, not tell."

Few-shot example:

"Convert these product descriptions into one-line taglines:

Product: A lightweight laptop with 20-hour battery life and a carbon fiber body.
Tagline: Power that goes the distance. Impossibly light.

Product: Noise-canceling headphones with spatial audio and 40-hour battery.
Tagline: Silence the world. Hear everything.

Product: A smart water bottle that tracks hydration and syncs with fitness apps.
Tagline:"

How Many Examples?

For most tasks, 2–3 examples are enough. Use more examples (5+) when the task has subtle nuances or when you need very consistent formatting. Make sure your examples are diverse enough to cover edge cases.

Chain-of-Thought Prompting

Chain-of-thought (CoT) prompting asks the AI to show its reasoning step by step before arriving at an answer. This dramatically improves accuracy for tasks that require logic, math, analysis, or multi-step reasoning.

Without chain-of-thought:

"A store has 45 apples. They sell 60% on Monday and half the remaining on Tuesday. How many are left?"

With chain-of-thought:

"A store has 45 apples. They sell 60% on Monday and half the remaining on Tuesday. How many are left? Think through this step by step, showing your work at each stage before giving the final answer."

The magic phrase is often as simple as adding "Think step by step" or "Let's work through this reasoning carefully." Research from Google and others has shown that chain-of-thought prompting can significantly boost accuracy on reasoning tasks.

System Prompts vs User Prompts

When using AI tools through their APIs or advanced interfaces, you'll encounter two types of prompts:

System Prompt

Sets the overall behavior, personality, and constraints for the AI. Think of it as the "job description" given before the conversation starts.

Defines the AI's role and persona
Sets rules and guardrails
Persists across the entire conversation
Usually hidden from the end user

User Prompt

The actual message or question sent by the user during the conversation. This is the turn-by-turn input.

Contains the specific request or question
Changes with every message
Provides task-specific context
Visible to the user

Where You See System Prompts

In consumer products, system prompts are used behind the scenes. For instance, Claude Projects let you set custom instructions that act as a system prompt. ChatGPT custom GPTs use a system prompt to define behavior. When using the API, you set the system prompt directly.

Prompt Template Library

Here are reusable templates you can adapt for common tasks. Copy them, fill in the brackets, and customize as needed.

Template 1: Content Creation

You are a [ROLE, e.g., content strategist]. I need to create [CONTENT TYPE, e.g., a blog post] about [TOPIC]. Audience: [DESCRIBE TARGET AUDIENCE] Goal: [WHAT SHOULD THE READER DO/FEEL/KNOW AFTER?] Length: [WORD COUNT OR SECTION COUNT] Tone: [PROFESSIONAL / CASUAL / ACADEMIC / ETC.] Please include: - A compelling headline and subheadline - An engaging opening hook - [NUMBER] main sections with headers - A clear conclusion with a call to action Avoid: [ANY TOPICS, WORDS, OR APPROACHES TO SKIP]

Template 2: Analysis and Decision Making

I need to make a decision about [DESCRIBE DECISION]. Here is the context: - Current situation: [DESCRIBE] - Constraints: [BUDGET, TIME, RESOURCES] - Stakeholders: [WHO IS AFFECTED] - Goals: [WHAT SUCCESS LOOKS LIKE] Please analyze this by: 1. Identifying the top 3–5 options 2. Listing pros and cons for each 3. Rating each option on feasibility, impact, and effort (1–5 scale) 4. Recommending the best option with reasoning 5. Outlining 3 immediate next steps Present as a structured comparison table followed by your recommendation.

Template 3: Learning and Research

I want to understand [TOPIC] at a [BEGINNER / INTERMEDIATE / ADVANCED] level. My background: [RELEVANT EXPERIENCE OR KNOWLEDGE] My goal: [WHY I'M LEARNING THIS] Time available: [HOW DEEP TO GO] Please: 1. Start with a one-paragraph overview 2. Explain the 3–5 most important concepts 3. Use analogies to explain complex ideas 4. Provide a real-world example for each concept 5. End with 3 questions I should explore next 6. Recommend 2–3 resources for further learning If I have misconceptions based on what I've said, please correct them.

Template 4: Email and Communication

Write a [TYPE: email / Slack message / memo] to [RECIPIENT AND RELATIONSHIP]. Purpose: [WHAT I NEED TO COMMUNICATE] Key points to include: - [POINT 1] - [POINT 2] - [POINT 3] Tone: [FORMAL / FRIENDLY / URGENT / DIPLOMATIC] Length: [BRIEF (2–3 sentences) / MEDIUM (1–2 paragraphs) / DETAILED] Call to action: [WHAT I WANT THEM TO DO NEXT] Additional context: [ANY SENSITIVE ASPECTS OR RELATIONSHIP DYNAMICS]

Template 5: Meeting Preparation

Help me prepare for a meeting about [TOPIC]. Meeting details: - Attendees: [WHO WILL BE THERE AND THEIR ROLES] - Duration: [LENGTH] - My role: [PRESENTER / PARTICIPANT / FACILITATOR] - Objective: [WHAT SHOULD THE MEETING ACCOMPLISH] Please create: 1. A structured agenda with time allocations 2. 3 key talking points I should raise 3. Potential questions I might be asked, with suggested answers 4. A one-page pre-read summary I can send to attendees 5. Follow-up action item template

Common Prompting Mistakes

Mistake	Why It's a Problem	Fix
Being too vague	The AI fills in gaps with assumptions that may not match your needs	Be specific about what you want, for whom, and in what format
Overloading a single prompt	Asking for too many things at once leads to shallow treatment of each	Break complex tasks into sequential prompts (prompt chaining)
Not iterating	Accepting the first response even when it's not quite right	Refine your prompt based on what was missing or wrong in the output
Ignoring the audience	The AI defaults to a generic tone that may not suit your readers	Always specify who the output is for
Skipping format instructions	You get a wall of text instead of a structured, usable output	Tell the AI exactly how to structure the response
Not providing enough context	The AI produces generic content that doesn't fit your specific situation	Include relevant background, constraints, and goals

Prompting Is Not "One and Done"

The best prompt engineers iterate. Your first prompt is a draft. Read the output critically, identify what's missing or off, and refine. Over time, you'll develop intuition for what works, but even experts revise their prompts.

Resources

Article

Prompt Engineering Guide

Anthropic

Anthropic's official guide to prompt engineering with Claude, covering techniques from basic to advanced with practical examples.