Home → ChatGPT & Prompt Engineering → Module 3

Module 3: Advanced Prompting Techniques

Power user techniques: few-shot learning, chain-of-thought, and parameter control

📅 Week 1 📊 Intermediate

You: "Calculate 47 x 23 and explain your reasoning step-by-step."

ChatGPT: "Let me think through this... 47 x 20 = 940, then 47 x 3 = 141, so 940 + 141 = 1,081."

That's chain-of-thought prompting — one of several advanced techniques that unlock ChatGPT's full reasoning capabilities. These techniques separate casual users from power users earning $150K+ as prompt engineers.

💡 By the end of this module, you'll master:

  • Zero-shot vs few-shot prompting (when to use each)
  • Chain-of-thought reasoning for complex problems
  • Role-based prompting for expert-level outputs
  • System messages and conversation control
  • Temperature, top-p, and other parameters

🎯 Technique 1: Zero-Shot vs Few-Shot Prompting

Zero-Shot Prompting

What it is: Giving ChatGPT a task without examples, relying on its pre-training to understand.

Example:

Classify this product review sentiment: "The laptop is okay but battery life is terrible."
Output: Negative (mentions "terrible" battery life, dampens "okay")

When to use: Simple, well-defined tasks ChatGPT already understands (sentiment analysis, translation, basic Q&A).

Few-Shot Prompting

What it is: Providing 2-5 examples before asking ChatGPT to perform the task. Shows the pattern you want.

Example:

Classify product review sentiment. Examples: Review: "Love this phone! Fast and sleek." Sentiment: Positive Review: "Terrible customer service, never buying again." Sentiment: Negative Review: "It's decent for the price." Sentiment: Neutral Now classify: "The laptop is okay but battery life is terrible."
Output: Negative (recognizes mixed sentiment leans negative due to "terrible")

When to use: Custom formats, nuanced tasks, or when you want specific style/structure. Examples train ChatGPT on your patterns.

✅ Zero-Shot Wins
  • Standard tasks (translate, summarize)
  • Quick one-off requests
  • When examples are hard to provide
  • General knowledge questions
✅ Few-Shot Wins
  • Custom output formats
  • Domain-specific jargon
  • Complex classification tasks
  • Matching your brand voice

🧠 Technique 2: Chain-of-Thought (CoT) Prompting

What is Chain-of-Thought?

Concept: Asking ChatGPT to "think step-by-step" or "explain your reasoning" before answering. Forces sequential logic instead of jumping to conclusions.

Why it works: Large language models perform better on complex reasoning when they break problems into steps (just like humans).

📊 Research finding: Adding "Let's think step-by-step" improved GPT-3 accuracy on math problems from 17% to 78% (Google Research, 2022).

CoT Example 1: Math Problem

❌ Without CoT
A store has 23 apples. They sell 7, then receive 15 more. How many do they have?
Output: 31 apples. (Correct, but no reasoning shown)
✅ With CoT
A store has 23 apples. They sell 7, then receive 15 more. How many do they have? Let's think step-by-step.
Output:
Step 1: Start with 23 apples
Step 2: Sell 7 → 23 - 7 = 16 apples
Step 3: Receive 15 → 16 + 15 = 31 apples
Final answer: 31 apples

CoT Example 2: Logic Problem

Question: If all roses are flowers, and some flowers fade quickly, can we conclude that some roses fade quickly? Think through this step-by-step using logical reasoning.
ChatGPT's reasoning:

Step 1: Analyze premise 1 → "All roses are flowers" (roses ⊂ flowers)
Step 2: Analyze premise 2 → "Some flowers fade quickly" (doesn't specify which flowers)
Step 3: Check logical connection → Premise 2 doesn't guarantee roses are in the "some flowers" subset
Step 4: Conclusion → No, we cannot conclude this. The roses might not be the flowers that fade quickly.

🎯 CoT Trigger Phrases

  • "Let's think step-by-step."
  • "Explain your reasoning before answering."
  • "Break this down into steps."
  • "Walk me through your thought process."
  • "First, analyze... Then, conclude..."

🎭 Technique 3: Role-Based Prompting

What is Role-Based Prompting?

Concept: Telling ChatGPT to act as a specific expert (lawyer, doctor, data scientist, comedian, etc.). This shapes vocabulary, depth, and perspective.

Why it works: Pre-training data includes countless examples of how different experts write/speak. Role assignment activates relevant patterns.

Example: Same Question, Different Roles

Question: "How do I reduce stress?"

🧑‍⚕️ As a Clinical Psychologist:

Act as a clinical psychologist with 15 years of experience. How do I reduce stress?
Response: "From a clinical perspective, chronic stress activates your HPA axis, releasing cortisol. Evidence-based interventions include: 1) Cognitive Behavioral Therapy (CBT) to reframe negative thought patterns, 2) Progressive Muscle Relaxation (PMR) to reduce physiological arousal, 3) Mindfulness-Based Stress Reduction (MBSR) — 8-week programs show 30% cortisol reduction..."

🏋️ As a Fitness Coach:

Act as a certified fitness coach. How do I reduce stress?
Response: "Exercise is your secret weapon! Here's what works: 1) 30 minutes cardio 3x/week releases endorphins (natural mood boosters), 2) Strength training builds resilience, 3) Yoga combines movement + breathwork, 4) Even a 10-minute walk outdoors reduces cortisol 15%. Start small: commit to one 20-minute workout tomorrow..."

🧘 As a Buddhist Monk:

Act as a Buddhist monk who teaches mindfulness. How do I reduce stress?
Response: "Stress arises from attachment and resistance to what is. Practice: 1) Observe thoughts without judgment — they are clouds passing through sky of awareness, 2) Return to breath — the anchor to present moment, 3) Accept impermanence — this too shall pass. Daily sitting meditation: start with 5 minutes. Not to eliminate stress, but to change relationship with it..."

💡 Pro tip: Combine role + expertise level + context for maximum precision:

"Act as a senior data scientist at Netflix with 10 years of experience in recommendation systems. Explain collaborative filtering to a product manager..."

⚙️ Technique 4: System Messages & Conversation Control

What are System Messages?

Concept: In API/advanced interfaces, system messages set persistent behavior for entire conversation. Think of it as ChatGPT's "operating instructions."

In ChatGPT web: Use "Custom Instructions" (Settings → Personalization) to set default behavior.

Example Custom Instructions:

What would you like ChatGPT to know about you? I'm a product manager at a B2B SaaS startup. I work with engineers, designers, and executives. I value data-driven decisions. How would you like ChatGPT to respond? - Be concise (max 3 paragraphs unless I ask for detail) - Use bullet points for lists - Include relevant metrics/data when discussing strategy - Avoid jargon without explanation - End responses with 1-2 follow-up questions

Result: Every response is now tailored to your role and preferences automatically.

Powerful System Message Templates

1. The Socratic Teacher

You are a Socratic tutor. Never give direct answers. Instead, ask questions that guide me to discover the answer myself. If I'm stuck, provide hints.

2. The Ruthless Editor

You are a ruthless editor. Prioritize clarity and concision. Cut unnecessary words. Point out weak arguments. Be direct but constructive.

3. The Expert Consultant

You are a senior consultant at McKinsey. Structure all responses using: 1) Key insight upfront, 2) Supporting evidence, 3) Actionable recommendation. Use frameworks (SWOT, Porter's 5 Forces) when relevant.

🎛️ Technique 5: Controlling Parameters (Temperature, Top-P)

Advanced Controls (API/Pro Users)

Temperature (0.0 - 2.0)

What it controls: Randomness/creativity of outputs.

0.0 (Deterministic)
2.0 (Chaotic)
  • Low (0.0-0.3): Factual, consistent, predictable. Use for: data extraction, code, technical writing
  • Medium (0.7-1.0): Balanced. Default for most tasks
  • High (1.5-2.0): Creative, surprising, risky. Use for: brainstorming, creative writing, unconventional ideas

Example: "Write a tagline for eco-friendly shoes"

Temperature 0.1: "Sustainable footwear for a better planet."

Temperature 1.5: "Walk softly, leave only footprints of change."

Top-P / Nucleus Sampling (0.0 - 1.0)

What it controls: Diversity of word choices. Limits AI to top X% probability words.

  • Low (0.1-0.5): Focused, safe word choices
  • High (0.9-1.0): Allows rare, unexpected words

Pro tip: Use high temperature OR high top-p, not both. Combining amplifies chaos.

Max Tokens

What it controls: Maximum response length (1 token ≈ 0.75 words).

Use case: Force concise responses (max 100 tokens) or allow long-form content (max 4000 tokens).

⚠️ Note: ChatGPT web interface doesn't expose these controls (except via Custom Instructions). API users and ChatGPT Plus with Advanced Data Analysis can adjust parameters.

📊 Technique 6: Structured Outputs & JSON Mode

The Problem with Unstructured Responses

You ask ChatGPT for product data, and it responds: "Sure! Here's the information you requested: The product is called..." — mixing explanatory text with data, making parsing impossible.

The solution: Request structured outputs (especially JSON) with clear schemas. ChatGPT will return clean, parseable data you can use in code, databases, or APIs.

Why JSON Mode Matters

  • ✅ Parse data reliably: No extra text to strip away
  • ✅ Integrate with code: Direct API/database insertion
  • ✅ Validate structure: Ensure required fields are present
  • ✅ Scale workflows: Batch process hundreds of responses
  • ✅ Professional AI apps: Table stakes for production systems

Basic JSON Request

Extract product information and return as valid JSON with this schema: { "product_name": string, "price": number, "features": array of strings, "rating": number (1-5), "in_stock": boolean } Source text: "The UltraBook Pro laptop costs $1,299. Key features include 16GB RAM, 1TB SSD, and 14-hour battery life. Customers rate it 4.7 out of 5 stars. Currently available for immediate shipping." Return ONLY valid JSON, no explanatory text before or after.
{ "product_name": "UltraBook Pro", "price": 1299, "features": ["16GB RAM", "1TB SSD", "14-hour battery life"], "rating": 4.7, "in_stock": true }

Advanced: Multi-Object JSON

Extract all products from this page and return as a JSON array. Schema for each product: { "name": string, "brand": string, "price": number, "discount_percent": number or null, "colors": array of strings } [Paste product listing HTML or text here] Return format: {"products": [...]} No markdown formatting. Pure JSON only.

Schema Design Best Practices

Tips for bulletproof JSON schemas:

  1. Be explicit about types: string, number, boolean, array, object
  2. Specify required vs optional fields: Use "or null" for optional
  3. Give examples: Show 1-2 sample outputs
  4. Request validation: "Ensure all fields are present before returning"
  5. Handle missing data: Specify defaults (null, empty array, etc.)

Error Handling & Validation

Extract customer feedback and return JSON: { "sentiment": "positive" | "negative" | "neutral", "score": number (1-10), "key_themes": array of strings (max 3), "urgency": "low" | "medium" | "high" } Before returning: 1. Validate all fields are present 2. Ensure sentiment is one of the 3 allowed values 3. Ensure score is between 1-10 4. If any field can't be determined, use null Source: [customer review text]

Real-World Use Cases

📊 Data Extraction

Extract structured data from unstructured text, PDFs, or images (via vision)

🔌 API Integration

Generate JSON that directly feeds into your application's API endpoints

📋 Form Filling

Convert conversational inputs into structured form data

📈 Report Generation

Create consistent, parseable reports for dashboards and analytics

⚠️ Common Mistakes:

  • Forgetting "no extra text": Always emphasize "return ONLY JSON"
  • Vague schemas: Be explicit about every field type and structure
  • No validation: Ask ChatGPT to validate before returning
  • Not handling nulls: Specify what to do when data is missing

🎯 Practice Exercise

Challenge: Design a JSON schema and prompt for extracting job postings

Requirements:

  • Job title, company, location, salary range
  • Required skills (array), experience years (number)
  • Remote/hybrid/onsite (enum)
  • Handle missing data gracefully

Test with: 2-3 real job posting paragraphs

🧠 Technique 7: Advanced Reasoning Patterns

Beyond basic chain-of-thought, these patterns unlock deeper reasoning for complex problems.

ReAct: Reasoning + Acting

What it is: Interleave thinking (reasoning) with actions (tool use, lookups, calculations). Think → Act → Observe → Repeat.

Best for: Multi-step problems requiring external information or calculations

Answer this question using ReAct (Reasoning + Acting) pattern: "What's the current market cap of Apple compared to 5 years ago?" Format: Thought: [Your reasoning] Action: [What information you need] Observation: [What you found/calculated] Thought: [Next reasoning step] ... (repeat until final answer) Answer: [Final conclusion]

Self-Consistency: Multiple Paths to Truth

What it is: Generate multiple independent reasoning paths, then identify the consensus answer. Reduces errors on complex problems.

Best for: Math, logic puzzles, critical decisions

Solve this problem using 3 different approaches, then identify the consensus answer: Problem: "A store offers 20% off, then an additional 10% off the sale price. If an item originally costs $100, what's the final price?" Approach 1: [Calculate step-by-step] Approach 2: [Use a different method] Approach 3: [Verify with alternative logic] Consensus answer: [If all 3 agree, that's the answer. If not, identify the error.]

Reflection: Self-Critique & Improve

What it is: Ask ChatGPT to critique its own response and improve it. Creates iterative refinement loop.

Best for: Writing, strategy, analysis—anything benefiting from revision

Write a cold outreach email for B2B sales. Then: 1. Critique your own email: What's weak? What could be better? 2. Rewrite it addressing those critiques 3. Final version only (no explanations)

Advanced version:

[Your initial task/question] After completing, do a self-review: - Score your answer 1-10 on: accuracy, completeness, clarity - Identify 2-3 specific weaknesses - Provide an improved version addressing those weaknesses - Explain what makes the improved version better

Tree-of-Thought: Explore Multiple Reasoning Branches

What it is: Instead of one linear reasoning chain, explore multiple paths simultaneously, evaluate each, then choose the best.

Best for: Strategic decisions, creative problems, complex tradeoffs

Problem: Should our startup expand to Europe or double down on North America? Use tree-of-thought reasoning: Branch 1: Optimistic expansion scenario - Key assumptions: [...] - Pros: [...] - Cons: [...] - Likely outcome: [...] Branch 2: Conservative consolidation scenario - Key assumptions: [...] - Pros: [...] - Cons: [...] - Likely outcome: [...] Branch 3: Data-driven hybrid approach - Key assumptions: [...] - Pros: [...] - Cons: [...] - Likely outcome: [...] Evaluation: Which branch has the strongest reasoning? Why? Recommendation: [Choose best path with rationale]

When to Use Each Pattern

Use ReAct when...
  • Need external information
  • Multi-step research required
  • Problem needs tool/calculation use
Use Self-Consistency when...
  • High stakes decision
  • Math or logic problems
  • Need error checking
Use Reflection when...
  • Quality matters most
  • Creative/strategic work
  • Time for iteration
Use Tree-of-Thought when...
  • Complex tradeoffs
  • Multiple viable paths
  • Strategic planning

🎯 Practice Challenge

Pick one pattern and apply it:

Problem: "Should I quit my job to start a business?"

  • ReAct: Research → Think → Research more
  • Self-Consistency: 3 different risk analyses
  • Reflection: Initial advice → Critique → Better advice
  • Tree-of-Thought: Quit now vs wait 6mo vs side hustle first

Which pattern gives you the most insight?

🚀 Power Move: Combining Techniques

Ultimate Prompt Template (All Techniques)

[ROLE] Act as a senior financial analyst at Goldman Sachs with 15 years of experience. [CONTEXT] I'm evaluating whether to invest $50K in cryptocurrency. [FEW-SHOT EXAMPLES] When analyzing investments, use this format: - Risk level: [High/Medium/Low] - Timeframe: [Short/Medium/Long-term] - Key factors: [3-5 bullets] - Recommendation: [Clear yes/no with reasoning] [CHAIN-OF-THOUGHT] Think through this step-by-step: 1. Assess market conditions 2. Evaluate risk vs potential return 3. Consider my situation 4. Provide recommendation [SPECIFIC ASK] Should I invest in cryptocurrency now? Use the format above.

Result: Highly structured, expert-level analysis that matches your needs exactly.

🎯 Hands-On Exercise: Technique Testing Lab

📝 Test Each Technique

Task: Try the same question with different techniques, compare results.

Question: "How can I improve my website's conversion rate?"

  1. Zero-shot: Ask the question directly.
    Observe: Generic advice?
  2. Few-shot: Provide 2 examples of analysis you like, then ask.
    Observe: Does it match your style?
  3. Chain-of-thought: "Let's think step-by-step. First analyze current state, then identify bottlenecks, then suggest improvements."
    Observe: More structured reasoning?
  4. Role-based: "Act as a conversion rate optimization expert at Amazon..."
    Observe: More specific tactics?
  5. Combined: Use role + CoT + examples.
    Observe: Best quality output?

Reflection: Which technique gave you the most valuable answer? Why?

📚 Summary: Advanced Techniques Arsenal

  • ✅ Zero-shot vs few-shot: No examples vs 2-5 examples to train patterns
  • ✅ Chain-of-thought: "Think step-by-step" → 78% accuracy boost on reasoning tasks
  • ✅ Role-based prompting: "Act as [expert]" → shapes vocabulary, depth, perspective
  • ✅ System messages: Custom Instructions for persistent behavior across conversations
  • ✅ Parameters: Temperature (creativity), top-p (diversity), max tokens (length)
  • ✅ Combining techniques: Role + CoT + few-shot = ultimate precision

🎯 Key Takeaway: These techniques aren't just "tricks" — they're how professional prompt engineers earn $150K+. Master them, and you'll extract 10x more value from ChatGPT than 95% of users. Next, we'll package these into reusable frameworks.

📝 Test Your Understanding

Question 1: What's the main difference between zero-shot and few-shot prompting?

Zero-shot is always better
Few-shot provides 2-5 examples to show the pattern you want
They're the same thing
Few-shot only works for coding

Question 2: What does chain-of-thought prompting do?

Forces ChatGPT to show step-by-step reasoning before answering
Makes responses longer
Only works for math problems
Slows down ChatGPT unnecessarily

Question 3: What does low temperature (0.0-0.3) produce?

More creative, random outputs
Slower responses
Deterministic, factual, consistent outputs
Shorter responses

Question 4: Why does role-based prompting work?

It's just a placebo effect
Training data includes examples of how different experts write, roles activate relevant patterns
ChatGPT actually becomes an expert
Only works with medical/legal roles

Question 5: What's the best approach for complex reasoning tasks?

Just ask the question directly
Use high temperature for creativity
Combine role-based + chain-of-thought + specific examples
Keep prompts as short as possible

🎓 Advanced AI Learning Paths

You've mastered advanced techniques. Take your skills to professional level with specialized courses and certifications:

🏆

ChatGPT Prompt Engineering for Developers

Coursera (DeepLearning.AI) | Free Audit / $49 Certificate

Taught by Andrew Ng: Learn directly from the pioneer of modern AI. This course covers prompt engineering best practices, API usage, and building AI applications. Highly recognized certification for your resume.

💡 Perfect for: Developers and technical professionals who want university-quality education from AI pioneers. Includes hands-on coding exercises and API integration projects.

  • Expert Instructors: Andrew Ng + Isa Fulford (OpenAI)
  • Hands-On Projects: Build real AI applications
  • Shareable Certificate: Add to LinkedIn
  • Flexible Schedule: Complete at your pace
Enroll on Coursera →
🚀

Advanced ChatGPT & API Mastery

Udemy | ₹499-899

Practical implementation: 15-25 hour courses covering advanced techniques, API integration, custom GPT building, and real-world business applications. Great for hands-on learners.

  • API Integration: Python scripts with OpenAI API
  • Custom GPTs: Build specialized AI assistants
  • Business Cases: Real-world implementation examples
  • Q&A Support: Direct instructor access
Browse Advanced Courses →

🚀 Next Step: Master Proven Frameworks

You've learned individual techniques. Now let's package them into battle-tested frameworks — RISEN, RTF, CRAFT — that prompt engineers use for consistent, high-quality results every time.

Coming up in Module 4: Learn three proven prompt frameworks (RISEN, RTF, CRAFT) that structure your thinking and deliver professional-grade outputs. Plus, build your personal prompt template library.