Prompt Engineering 101
Advanced prompting techniques, privacy rules, and what never to share with AI
Section 1: Why Prompting Is a Skill
The difference between a mediocre AI response and an excellent one usually isn't the AI—it's how you ask. Here's the proof.
WEAK PROMPT
Result: Generic, formal, sounds like every other corporate email.
No context. No audience. No tone. AI guesses and gets it wrong.
ENGINEERED PROMPT
Result: Personal, appropriate tone, actually ready to send.
Everything specified: role, task, audience, tone, length, structure.
The pattern: Same model, same AI, completely different output. Better prompting = better results. It's a learnable skill.
Section 2: Zero-Shot Prompting
Zero-shot means: "I'm asking you something you weren't explicitly trained on, with no examples." AI tries to figure it out from general knowledge.
When to use Zero-Shot:
- Simple, straightforward requests
- Asking for something common (writing, explaining, analyzing)
- You're not sure what you want yet (brainstorming)
The limitation:
AI has to guess your style, format, tone, and depth. It might get it wrong.
Three Zero-Shot Examples:
Example 1: Simple Question
Good for: Quick facts, explanations, summaries
Example 2: Open-Ended
Good for: Brainstorming, generating options
Example 3: Creative Task
Good for: Creative work, when you don't care about the exact format
Pro tip: Zero-shot is fast and often good enough. Don't over-engineer every prompt. But when you need something specific, move to few-shot or chain-of-thought.
Section 3: Few-Shot Prompting
Few-shot means: "Here are examples of what I want. Now do more of that." It's more powerful than zero-shot because AI sees the pattern.
When to use Few-Shot:
- You have a specific style or format you want copied
- You've done something once and want consistency
- Zero-shot got close but wasn't quite right
Why it works:
AI is a pattern-matcher. Showing it 2-3 examples is like saying "here's the pattern I want"—much clearer than describing it.
Example: Product Descriptions
Here are 2 example product descriptions (your style):
Notice: short, benefit-focused, includes key details, uses pipes to separate specs, conversational but professional.
Now the prompt:
The magic: AI sees the pattern—brief, benefit-focused, feature-structured—and copies it for the new product. Without examples, it might write a paragraph. With examples, it nails your style.
Section 4: Chain-of-Thought Prompting
Chain-of-thought means: "Think step by step before you answer." It forces AI to work through complex problems methodically instead of guessing.
When to use Chain-of-Thought:
- Logic puzzles or reasoning problems
- Multi-step decisions
- Math or analysis questions
- Anything where the answer is "it depends"
Example: Without Chain-of-Thought
❌ Without CoT
AI might answer: "He spends $9,600 annually." (Wrong—it's $12,800 for 14 cows)
Example: With Chain-of-Thought
✓ With CoT
AI works through: 12 - 3 + 5 = 14 cows. 14 × $800 = $11,200. Correct!
Just add these phrases to any complex prompt:
- "Think step by step"
- "Work through this carefully"
- "Show your reasoning"
- "Break this into steps: 1) ... 2) ... 3)"
Section 5: Role Prompting
Give AI a role with deep expertise, and it'll answer from that perspective. Same question, different expert = different (and often better) answers.
The Magic Phrase:
"You are an expert [field] with 20 years of experience who explains complex things simply and practically."
Example: Same Task, Four Different Roles
Task: "Explain what makes a good website."
Financial Advisor's Answer
Focus: Trust, security, compliance, conversions
Skeptical Journalist's Answer
Focus: Credibility, sourcing, bias, manipulation tactics
UX Designer's Answer
Focus: Usability, accessibility, flow, user journey
Stand-Up Comedian's Answer
Focus: Absurd design trends, user frustrations, humor
Why it works: Each role has different priorities and knowledge. A financial advisor thinks about trust and security. A comedian thinks about what's ridiculous. Role shapes the answer.
Section 6: System Prompts & Custom Instructions
Instead of repeating context in every prompt, set up "standing instructions" that apply to all your conversations. It's like configuring AI once, then using it forever.
What They Are:
- System Prompt: A persistent instruction that shapes every response
- Custom Instructions: Available in ChatGPT and Gemini (called "Gems")
- Claude Projects: Project Instructions in Claude (same idea)
Where to Set Them:
- Claude: Create a "Project" → "Project Instructions"
- ChatGPT: Settings → "Custom Instructions" (free tier)
- Gemini: Create a "Gem" and set instructions there
Example Personal System Prompt:
Real Example for a Marketing Manager:
The payoff: After setting this up once, every conversation with AI is automatically tuned to your needs. You never have to repeat context. It's worth 10 minutes to set up.
Section 7: Five Universal Prompt Templates
Copy-paste these templates and fill in the blanks. They work across all professions.
1. The Explainer
Use for: Teaching, simplifying complexity, writing guides
2. The Writer
Use for: Drafting content, emails, proposals, articles
3. The Reviewer
Use for: Getting feedback, quality control, stress-testing ideas
4. The Planner
Use for: Project planning, organizing work, creating schedules
5. The Analyst
Use for: Data analysis, extracting insights, finding patterns
Section 8: Hands-On — Build Your Personal System Prompt
Set Up AI for Life
Create a system prompt that shapes how AI talks to you—from now on, in every conversation.
- Decide your profession and main AI use cases (teaching, coding, writing, research, etc.)
- Write your system prompt using the template from Section 6. Spend 5 minutes thinking about your actual preferences. (Most people skip this and regret it.)
- Go to Claude Projects (or ChatGPT Custom Instructions, or Gemini Gems) and paste your system prompt.
- Test it with 3 different requests. Does it respond in your preferred style? Does it avoid things you said never?
- Refine once if needed. Now you're done. Every conversation from now on will be tuned to you.
Why bother? This system prompt saves you from explaining your context hundreds of times. It compounds. In a year, you'll have saved dozens of hours.
Section 9: What NOT to Share with AI — Protecting Your Privacy
AI tools are powerful — but they are not a safe place for sensitive personal or financial information. Before you paste something into an AI chat, it is worth understanding what happens to that data and what you should never share.
Treat every AI chat window like a public forum. If you would not post it on a notice board, do not paste it into a chat. This applies to all AI tools — Claude, ChatGPT, Gemini, Copilot, and others — unless you are on a verified enterprise plan with explicit data privacy guarantees.
Never Paste These Into an AI Chat
- Social Security Number (SSN) or National ID
- Passport number or Driver's license number
- Date of birth combined with name and address
- Immigration or visa document numbers
- Bank account numbers or routing numbers
- Full credit card or debit card numbers
- Tax returns, W-2s, or income statements
- Brokerage or investment account details
- Passwords or PINs of any kind
- API keys or secret tokens
- Recovery phrases or two-factor backup codes
- Corporate VPN or system credentials
- Full medical records or diagnostic reports
- Insurance policy numbers and claim details
- Mental health history or therapy notes
- Genetic test results
- Customer lists with names, emails, or phone numbers
- Employee salary or HR records
- Student records or grades (FERPA)
- Anyone's private messages without their consent
- Unreleased product roadmaps or financials
- M&A details, contracts, or legal strategy
- Proprietary source code (in consumer tools)
- Client or patient data covered by NDA/HIPAA
Why Does This Matter? How AI Data Works
When you type something into most AI chat interfaces, that text is sent to a cloud server and processed there. Depending on the provider and plan:
Safe Ways to Use AI with Sensitive Context
You can still get value from AI for sensitive topics — just anonymize or generalize the data first:
"Here is my W-2 from 2024. My SSN is 123-45-6789. Can you help me file my taxes?"
"I earned $75,000 in salary and $3,000 in freelance income last year. What deductions should I look into?"
"Here is our customer database export. Can you find patterns in the purchase history?" [pastes real names + emails]
"I have purchase data: 3,200 customers, avg order $47, 22% repeat buyers. What patterns should I analyze?"
When It IS Safe: Enterprise & Privacy Plans
If your organization uses an enterprise plan from Anthropic (Claude for Work/Enterprise), OpenAI (ChatGPT Enterprise), or Microsoft (Copilot for M365), your data typically has stronger protections:
- Data is not used to train models
- No human review of your conversations
- Data stays within your organization's contracted region
- Business associate agreements (BAA) available for HIPAA contexts
Still check: Even on enterprise plans, verify the specific data handling terms before processing data covered by regulations like HIPAA, GDPR, or FERPA.
The 3-Second Privacy Check
Before pasting anything into AI, ask yourself:
If the answer to any of these is "yes" — anonymize before sharing, or don't share at all.
- Zero-Shot: No examples. Ask directly. Fast, good for simple tasks.
- Few-Shot: Show 2-3 examples. AI copies the pattern. Best for consistency.
- Chain-of-Thought: "Think step by step." Forces logical reasoning. Great for complex problems.
- Role Prompting: "You are a [role]." Different experts give different (better) answers.
- System Prompts: Persistent instructions that apply to all conversations.
- The Explainer: "You are [role]. Explain [topic] to [audience]..."
- The Writer: "Write a [type] about [topic] for [audience]. Tone: [tone]..."
- The Reviewer: "Review this [type] and identify: 1) errors 2) improvements 3) rating..."
- The Planner: "Create a [timeframe] plan to achieve [goal]. Constraints: [list]..."
- The Analyst: "Analyze [content]. Tell me: [questions]..."
- "Think step by step"
- "Show your reasoning"
- "You are an expert [field]"
- "Use an analogy"
- "Be critical"
- "Give me 3 alternatives"
- "Avoid [jargon/clichés]"
- "Keep it under [length]"
- Simple question? → Zero-shot (fastest)
- Want consistent style? → Few-shot (show examples)
- Complex reasoning? → Chain-of-thought (think step-by-step)
- Want different perspective? → Role prompting (be a [role])
- Same preferences every time? → System prompt (set once, use forever)
- Identity: SSN, passport number, driver's license number
- Financial: bank accounts, card numbers, tax returns
- Passwords, API keys, or 2FA recovery codes
- Medical records, insurance numbers, health history
- Other people's personal data without their consent
- Confidential business data: M&A, unreleased financials, client lists
- Replace names with "[Customer A]" or "[Employee 1]"
- Use ranges instead of exact figures: "$70K–80K" not "$76,234"
- Describe what the data looks like rather than pasting it
- Ask your question with hypothetical numbers, not real ones
- 3-second check: "Would I be OK if this appeared in a news story?"