Prompt Engineering Fundamentals

Two marketers, one AI, completely different results

Maria needs a social media strategy for her company's new product launch. She opens ChatGPT and types:

"Give me a social media strategy."

She gets a generic list of platitudes: "Post consistently," "Engage with your audience," "Use hashtags." Useless. She closes the tab and spends three hours writing the plan herself.

Her colleague Devon opens the same tool and types:

"You are a senior social media strategist specializing in B2B SaaS product launches. I'm launching a project management tool aimed at remote teams of 10-50 people. Our budget is $5,000/month. Our audience is primarily on LinkedIn and Twitter. We have no existing following. Create a 30-day launch strategy with specific post types, posting frequency, and one viral-worthy content idea per week. Format as a table with columns: Week, Platform, Post Type, Topic, Goal."

Devon gets a detailed, actionable 30-day plan in 15 seconds. He tweaks two rows and ships it to his VP.

Same AI. Same task. The difference wasn't intelligence or experience -- it was prompt engineering.

By the end of this module, you'll master five core prompting techniques — zero-shot, few-shot, chain-of-thought, output format control, and prompt chaining — and you'll have built a real prompt chain for a task from your work.

Zero-shot vs. few-shot prompting

These are the two most fundamental techniques in prompt engineering, and the names are simpler than they sound.

Zero-shot means you give the AI a task with no examples. You're trusting it to figure out the format, tone, and structure on its own.

"Classify this customer review as positive, negative, or neutral: 'The delivery was late but the product quality was amazing.'"

Few-shot means you show the AI a few examples first, then give it the task. You're saying: "Here's the pattern -- now follow it."

"Classify these customer reviews:

Review: 'Absolutely love it, best purchase ever!' → Positive Review: 'Broke after two days, total waste of money.' → Negative Review: 'It's okay, nothing special.' → Neutral

Now classify: 'The delivery was late but the product quality was amazing.' →"

The rule of thumb: If you care about the exact format, tone, or classification criteria of the output, use few-shot. If you just need a quick answer, zero-shot is fine.

There Are No Dumb Questions

How many examples do I need for few-shot prompting?

Two to three is the sweet spot. One example shows the pattern. Two confirm it. Three lock it in. More than five usually wastes tokens without improving quality.

Does few-shot prompting work with every AI tool?

Yes. ChatGPT, Claude, Gemini, Copilot -- they all respond to examples. It's the most universal technique in prompt engineering because it works with how language models process patterns.

System prompts and role assignment

You already know the basics of role assignment from Module 1's Role + Context + Task formula. Here's why it's so powerful and how to take it further.

Every great prompt starts with telling the AI who to be. This is called role assignment, and it's the single highest-leverage technique in prompt engineering.

Why? Because the role shapes everything else. A financial analyst writes differently from a kindergarten teacher. A legal advisor focuses on different details than a marketing copywriter. When you assign a role, you're activating a specific subset of the AI's training data.

Think of it like a costume. The AI puts on whatever role you give it, and that costume changes how it talks, what it emphasizes, and what it skips.

Chain-of-thought prompting

This is the most powerful technique in this module — and the one that feels most like magic.

Here's a technique that dramatically improves accuracy on any task requiring reasoning: chain-of-thought prompting. Instead of asking for a final answer, you ask the AI to think through the problem step by step.

The difference is dramatic:

Without chain-of-thought:

"A store has 45 apples. They sell 60% on Monday and half of what's left on Tuesday. How many remain?" AI answer: "9 apples" (sometimes wrong, sometimes right -- it's guessing)

With chain-of-thought:

"A store has 45 apples. They sell 60% on Monday and half of what's left on Tuesday. How many remain? Think step by step." AI: "Step 1: 60% of 45 = 27 sold on Monday. Step 2: 45 - 27 = 18 remaining. Step 3: Half of 18 = 9 sold on Tuesday. Step 4: 18 - 9 = 9 remaining. Answer: 9 apples."

The answer might be the same, but the chain-of-thought version is reliably correct because the model works through each step instead of jumping to a conclusion. For complex reasoning, this is the single biggest accuracy improvement you can make.

When to use it:

• Math and logic problems
• Multi-step analysis ("Compare these 3 options and recommend one")
• Debugging ("Here's my code -- find the bug")
• Complex writing tasks ("Outline, then draft, then refine")

Output format control

One of the most underused prompt engineering techniques: telling the AI exactly what shape the output should take.

Most people ask for information and let the AI decide the format. Pros specify it. Here's why this matters:

You can request virtually any format:

• JSON: "Return the data as a JSON object with keys: name, email, role"
• Markdown table: "Format as a markdown table"
• Bullet points: "Use bullet points, max 10 words per bullet"
• Numbered steps: "Give me numbered steps I can follow"
• Specific structure: "Use this structure: Problem, Cause, Solution, Timeline"

Pro tip for developers: When you need structured data, ask for JSON with a specific schema:

"Extract the following from this customer email and return as JSON:

json

{"sentiment": "positive|negative|neutral", "urgency": "low|medium|high", "category": "billing|technical|general", "summary": "one sentence"}
```"

This turns ChatGPT into a data extraction pipeline you can plug directly into code.

There Are No Dumb Questions

What about temperature and creativity settings?

Temperature controls how "creative" vs. "predictable" the AI's output is. Low temperature (0-0.3) gives you consistent, factual responses -- great for data extraction, classification, and code. High temperature (0.7-1.0) gives you more varied, creative responses -- great for brainstorming, creative writing, and generating diverse ideas. Most AI tools default to around 0.7. If you're using the API, you can set this directly. In ChatGPT's interface, you can't change temperature, but you can simulate it with instructions like "give me the most likely answer" (low temp) or "be creative and unconventional" (high temp).

Can I ask for multiple formats in one prompt?

Absolutely. "Give me a summary paragraph, then a bullet list of action items, then a table comparing options." The AI handles multi-format output well as long as you're explicit about where each section starts.

Prompt chaining: breaking big tasks into steps

Here's what separates casual users from prompt engineers: prompt chaining. Instead of cramming everything into one massive prompt, you break the task into a sequence of smaller prompts where each step feeds into the next.

Think of it like cooking a complex meal. You don't throw all the ingredients into one pot and hope. You prep ingredients, make the sauce, cook the protein, and assemble. Each step builds on the previous one.

Why chaining works better than one giant prompt:

1. Better quality -- the AI focuses on one thing at a time
2. More control -- you can course-correct between steps
3. Less confusion -- complex prompts with 10 instructions often get partially ignored
4. Reusable steps -- you can swap in different Step 1s and keep the rest of the chain

Now that you have the five core techniques, here's how professionals combine them into ready-to-use templates.

Common prompt patterns (your copy-paste toolkit)

Here are the seven patterns that cover 90% of real-world prompt engineering use cases. Bookmark these.

1. Summarize

"Summarize [this text] in [number] bullet points. Each bullet should be one sentence. Focus on [specific aspect]. The reader is [audience]."

2. Extract

"From the following text, extract: [field 1], [field 2], [field 3]. Return as [format]. If a field isn't mentioned, write 'N/A'."

3. Classify

"Classify the following [items] into these categories: [cat 1], [cat 2], [cat 3]. For each, provide the category and a confidence level (high/medium/low). Here are examples: [examples]."

4. Generate

"Generate [number] [type of content] for [audience]. Tone: [tone]. Constraints: [constraints]. Here's an example of what I'm looking for: [example]."

5. Transform

"Rewrite [this content] for [new audience/format/tone]. Keep the core message but change [specific aspects]. Length: [target length]."

6. Analyze

"Analyze [this data/text/situation]. Identify: [what to look for]. Format your analysis as [format]. Prioritize [criteria]."

7. Compare

"Compare [A] and [B] across these dimensions: [dim 1], [dim 2], [dim 3]. Format as a table. End with a recommendation for [use case]."

Anti-patterns: what NOT to do

Knowing what to avoid is just as important as knowing what to do. Here are the most common prompt engineering mistakes.

Anti-pattern 1: The kitchen sink prompt

"Write a blog post about AI in healthcare that's SEO-optimized and includes statistics and is written for executives but also accessible to general readers and should be 1000 words but could be longer if needed and include a call to action and make it engaging but professional and use examples from real companies but don't name any specific companies..."

This prompt tries to do everything and achieves nothing well. The AI will ignore half of your constraints. Fix: Break it into a chain or prioritize your top 3 constraints.

Anti-pattern 2: Being polite instead of precise

"Could you maybe try to possibly write something along the lines of a marketing email, if it's not too much trouble? Something kind of professional but also friendly?"

Hedging language ("maybe," "kind of," "along the lines of") makes the AI hedge its output too. Fix: Be direct. "Write a marketing email. Tone: professional but warm."

Anti-pattern 3: Contradictory instructions

"Write a short, detailed, comprehensive summary."

Short and comprehensive are opposites. The AI picks one and ignores the other. Fix: Decide what matters most. "Write a 100-word summary covering only the three main findings."

Anti-pattern 4: No constraints at all

"Tell me about machine learning."

This produces a generic overview that could be an intro to a textbook. Without knowing WHO you are, WHY you're asking, and HOW MUCH detail you need, the AI defaults to "explain it like a Wikipedia article." Fix: Add audience, purpose, and length.

Real-world prompt templates

Here are five battle-tested templates you can copy and customize right now.

Email writer:

"You are a [role]. Write an email to [recipient + their role] about [topic]. Goal: [what you want them to do after reading]. Tone: [specify]. Length: under [X] words. Include: [specific elements]. Avoid: [things to skip]."

Meeting prep:

"I have a meeting with [who] about [topic] in [timeframe]. My goal is to [outcome]. Prepare: (1) an agenda with 3-4 discussion points, (2) two questions I should ask them, (3) two objections they might raise and how to respond. Keep each section under 50 words."

Content creator:

"Write a [content type] about [topic] for [audience]. The reader's main problem is [pain point]. Start with a hook that addresses this pain point. Include [number] specific examples. End with [CTA type]. Tone: [specify]. Length: [specify]."

Data analyzer:

"Here is [data type]: [paste data]. Analyze it and tell me: (1) the top 3 insights, (2) any anomalies or red flags, (3) what I should do next based on this data. Present insights as bullet points with the metric, the finding, and why it matters."

Decision maker:

"I need to decide between [option A] and [option B]. Context: [situation]. My priorities are: [list 3-4 priorities in order]. Compare both options across these priorities in a table. Then give me your recommendation with reasoning. Think step by step."

Back to Maria and Devon

One week after Maria learned the fundamentals, her outputs started matching Devon's — not because she worked harder, but because her instructions got sharper. She stopped opening ChatGPT with a vague hope and started opening it with a role, a context, a task, and a format in mind. The 30-day social media strategy she once wrote by hand over three hours she now gets in a first draft in 15 seconds, then refines through two rounds of iteration. Devon hasn't gained any advantage — the gap closed the moment the prompts did. The tool didn't change. The instructions did.

Key takeaways

• Zero-shot is quick and works for simple tasks. Few-shot (adding examples) is more reliable for anything requiring a specific format or classification.
• Role assignment is the highest-leverage single technique. "You are a [specific expert]" changes everything about the output.
• Chain-of-thought ("think step by step") dramatically improves reasoning accuracy by forcing the model to work through each step before reaching a conclusion.
• Output format control (JSON, tables, bullet points) turns AI from a chatbot into a productivity tool.
• Prompt chaining beats giant prompts. Break big tasks into 3-5 linked steps.
• Anti-patterns to avoid: kitchen sink prompts, hedging language, contradictory instructions, and accepting first-draft output.

You now have the prompting toolkit. In Module 3, you'll put these techniques to work on writing — drafting emails, blog posts, and professional documents with prompt chaining and tone control.