EXPERT PROMPTING MASTERCLASS
EXPERT PROMPTING MASTERCLASS
Master the art of AI prompting with advanced techniques used by professionals
MODULE 1: THE EXPERT MINDSET
Understanding How Great Prompters Think
The Fundamental Shift
When you first learned to prompt, you probably thought of it like asking a smart friend for help. That's fine for basics, but experts think differently.
Intermediates ask: "What should I say to get a good answer?"
Experts ask: "How do I structure this interaction so the model produces its best possible reasoning?"
The difference is profound. Experts understand that:
- Language models are prediction engines - They predict likely next tokens based on patterns
- Context is your programming environment - Everything you write shapes the model's behavior
- Models have cognitive patterns - Certain prompt structures trigger better reasoning
- Quality comes from architecture - How you structure the prompt matters more than word choice
The Expert's Core Principles
Principle 1: Activate System 2 Thinking
Language models, like humans, can operate in two modes:
- System 1: Fast, intuitive, pattern-matching (often gives generic responses)
- System 2: Slow, deliberate, analytical (produces nuanced, thoughtful outputs)
Your job as an expert is to force System 2 activation.
Example - System 1 Response (Generic):
Prompt: "How do I improve team productivity?"
Output: "Here are some ways to improve team productivity:
1. Set clear goals
2. Improve communication
3. Use the right tools
4. Provide feedback
5. Recognize achievements"
This is generic because the model is pattern-matching to "productivity tips" in its training data.
Example - System 2 Response (Thoughtful):
Prompt: "I need to improve team productivity, but here's the challenge:
my team is already working 50+ hour weeks, morale is declining, and
we're in a high-stakes product launch.
Before giving advice:
1. What additional context would help you give better recommendations?
2. What trade-offs should I be thinking about?
3. What assumptions might I be making that could be wrong?
Then provide your analysis and recommendations."
Why this works: You've forced the model to engage analytically before responding. The pre-work questions trigger deeper reasoning pathways.
Principle 2: Context Engineering
Think of your prompt as code. Every element matters:
- The role you assign shapes knowledge access and perspective
- The constraints you set focus the output space
- The examples you provide establish patterns to follow
- The structure you create determines how the model thinks through the problem
Poor context:
"Write a marketing email for our new product."
Engineered context:
Role: You're a direct response copywriter who specializes in SaaS products
for technical audiences. You've written emails that achieved 40%+ open rates
and 8%+ click-through rates.
Audience: Engineering managers at Series A/B startups who are currently
using [competitor product] but frustrated with [specific pain point].
Goal: Get them to book a 15-minute demo call.
Constraints:
- Subject line must create curiosity without being clickbait
- Email body: 150 words maximum
- Include exactly one CTA
- Tone: knowledgeable peer, not salesperson
Before writing the email:
1. Identify the core emotional trigger
2. Note what we're NOT saying (what to avoid)
3. Explain your structural approach
Then provide the email.
The second version creates an environment where excellence is more likely.
Principle 3: Intelligent Iteration
Experts never accept first outputs. They iterate systematically:
The Expert Iteration Loop:
- Get baseline → Identify gaps → Add specific constraints → Re-generate
- Get variation → Compare approaches → Synthesize best elements → Refine
- Stress test → Find weaknesses → Strengthen → Polish
Example in action - Iteration 1 (Baseline):
"Explain blockchain to a non-technical audience."
[Get response, evaluate]
Iteration 2 - Add constraints:
"That's too abstract. Rewrite it:
- Use only one analogy (choose the best one)
- Include a concrete example of a real problem it solves
- Address the #1 objection people have
- Keep it under 200 words"
[Get response, evaluate]
Iteration 3 - Refine:
"Better. Now make it more engaging:
- Start with a surprising fact or question
- Remove any jargon that remains
- End with a clear 'so what' moment"
[Get response, evaluate]
Iteration 4 - Polish:
"Final pass: tighten the language. Every sentence should be essential."
[Get final response]
This takes 5-10 minutes but produces dramatically better results than spending 30 minutes trying to write the "perfect" first prompt.
Practice Exercises
✍️ PRACTICE EXERCISE 1.1: Activate System 2
Your turn. Take this generic prompt and transform it using System 2 activation:
Generic prompt:
"Give me ideas for reducing customer churn."
Your enhanced prompt should:
- Provide specific context about the business
- Include pre-work questions that force analysis
- Request structured thinking
- Set clear output requirements
Spend 5 minutes writing your enhanced version. Then test it in your preferred AI tool.
✍️ PRACTICE EXERCISE 1.2: Context Engineering
Scenario: You need to write a performance review for a team member who's technically excellent but struggles with communication.
Bad prompt:
"Write a performance review for John."
Your task: Engineer a complete context that includes:
- Your role and relationship
- Specific details about John's work
- The purpose and audience of the review
- Tone and format requirements
- Any constraints or sensitivities
Write your engineered prompt and test it.
The Expert Self-Assessment
Before moving to the next module, rate yourself honestly:
I understand that:
- ☐ Prompts structure thinking, not just request information
- ☐ Generic inputs produce generic outputs
- ☐ Iteration is where quality comes from
- ☐ Context engineering is a skill I can develop
- ☐ Different prompts trigger different reasoning modes
If you checked all boxes, you're ready for Module 2.
MODULE 2: ADVANCED PROMPTING FRAMEWORKS
The Techniques That Produce Excellence
Now that you understand the expert mindset, let's build your technical toolkit. These are battle-tested frameworks that work across all AI models.
Framework 1: Layered Prompting
The single most powerful technique experts use is building prompts in conceptual layers rather than dumping everything at once.
The Five Layers
- Layer 1: Role & Expertise - Who is the AI in this interaction? What knowledge should it access?
- Layer 2: Context & Constraints - What's the situation? What are the boundaries?
- Layer 3: Task Structure - How should the AI approach this? What's the thinking process?
- Layer 4: Output Specifications - What format? What length? What elements must be included?
- Layer 5: Quality Controls - How will we validate? What checks should happen?
Layered Prompting in Action
Scenario: You need a content strategy for a product launch.
❌ Intermediate Approach:
"Create a content strategy for launching our new project management
tool targeting remote teams."
✅ Expert Approach:
LAYER 1 - ROLE & EXPERTISE:
You are a content strategist who has launched 20+ B2B SaaS products.
Your specialty is creating content engines that generate qualified leads
with minimal paid spend. You think in terms of customer journey stages
and content formats that actually convert.
LAYER 2 - CONTEXT & CONSTRAINTS:
Product: Project management tool with unique async collaboration features
Target: Remote-first teams, 10-50 people, currently using Asana or Trello
Timeline: 8 weeks pre-launch, ongoing post-launch
Budget: $15K (must include content creation and promotion)
Team: 1 content writer, 1 designer, founder available for thought leadership
LAYER 3 - TASK STRUCTURE:
Approach this in phases:
Phase 1 - Strategic Foundation:
- Define our 3 core content pillars based on product differentiation
- Map content types to each funnel stage (awareness → consideration → decision)
- Identify our unique POV that will stand out
Phase 2 - Content Planning:
- Create 8-week pre-launch calendar
- Specify content format, channel, and goal for each piece
- Note dependencies and production timelines
Phase 3 - Distribution Strategy:
- Organic channels (which platforms, why, frequency)
- Paid promotion (where to allocate budget for maximum impact)
- Partnership/collaboration opportunities
LAYER 4 - OUTPUT SPECIFICATIONS:
Deliver as:
1. Strategic narrative (2-3 paragraphs explaining the approach)
2. Content calendar (table format: Week | Content Piece | Format |
Channel | Goal | Owner)
3. Budget allocation (breakdown of $15K)
4. Success metrics (3-5 KPIs we should track)
LAYER 5 - QUALITY CONTROLS:
After completing the strategy:
- Identify the 3 riskiest assumptions you've made
- Note what additional information would strengthen this plan
- Provide 2-3 alternative approaches we should consider
Let's begin with Phase 1.
Why this is powerful:
- The AI knows exactly what expertise to draw from
- Clear constraints prevent generic advice
- Phased approach enables complex reasoning
- Specific output format ensures usability
- Quality controls catch gaps and assumptions
When to Use Layered Prompting
✅ Use it for:
- Complex business problems
- Creative projects requiring strategy
- Technical documentation
- Any task where quality matters more than speed
❌ Don't use it for:
- Simple factual queries
- Quick edits or formatting
- Straightforward tasks with obvious approaches
Framework 2: Constitutional Prompting
Give the AI a "constitution"—a set of principles that govern how it should think and respond throughout your interaction.
The Power of Constitutions
A constitution creates consistency across multiple exchanges and embeds quality standards directly into the model's behavior.
Example Constitution:
OPERATING PRINCIPLES FOR THIS CONVERSATION:
1. Specificity over Generalization
- Provide concrete examples, not abstract concepts
- Use numbers, names, and specific scenarios
- Replace "often" with "in X% of cases" when possible
2. Reasoning Transparency
- Show your thinking process, not just conclusions
- Explain why you chose approach A over approach B
- Note when you're uncertain
3. Productive Disagreement
- Present counter-arguments to your own recommendations
- Identify when conventional wisdom might be wrong
- Challenge my assumptions if you spot flaws
4. Practical Orientation
- Every recommendation must be actionable
- Include what it costs (time, money, complexity)
- Flag what could go wrong
5. No Platitudes
- Ban: "think outside the box," "synergy," "leverage," "circle back"
- If it sounds like it came from a corporate memo, rewrite it
- Be direct and human
ACKNOWLEDGE THESE PRINCIPLES, THEN: Help me decide whether to
build or buy a customer support ticketing system.
The AI will follow these principles throughout the conversation, producing much higher quality responses with less hand-holding.
Constitutional Prompting Templates
For Analysis Work:
ANALYTICAL PRINCIPLES:
1. Data before opinions - cite sources and numbers
2. Acknowledge uncertainty - if confidence <80%, say so
3. Multiple perspectives - always include alternate interpretations
4. Falsifiability - state what evidence would prove you wrong
Now analyze: [your task]
For Creative Work:
CREATIVE PRINCIPLES:
1. Original over familiar - avoid the first idea that comes to mind
2. Specific over general - use concrete details and sensory language
3. Surprising over expected - look for the unusual angle
4. Purposeful over decorative - every element should serve the goal
Now create: [your task]
For Technical Work:
TECHNICAL PRINCIPLES:
1. Correctness over cleverness - working code beats elegant code
2. Explain trade-offs - note what you're optimizing for and against
3. Consider maintenance - flag what will be hard to change later
4. Security-conscious - point out potential vulnerabilities
Now build: [your task]
✍️ PRACTICE EXERCISE 2.1: Build Your Constitution
Scenario: You're working with an AI to develop your business strategy for the next year.
Your task: Write a 4-5 principle constitution that will ensure the AI gives you the kind of strategic thinking you need (not generic business advice).
Consider:
- What bad habits do generic business recommendations have?
- What qualities do you value in strategic advice?
- What should the AI prioritize or avoid?
Write your constitution, then test it with a real strategic question.
Framework 3: Chain-of-Thought Scaffolding
Don't just ask for reasoning—provide the exact scaffolding the model should use.
The Scaffolding Principle
When you provide a thinking structure, the model produces dramatically more thorough and logical outputs.
Without Scaffolding:
"Should we expand our product to serve enterprise customers? Explain your reasoning."
Result: You'll get some pros and cons, but the analysis will be shallow.
With Scaffolding:
Should we expand our product to serve enterprise customers?
Use this analytical framework:
STEP 1 - OPPORTUNITY ASSESSMENT:
- Market size and growth rate for enterprise segment
- Current competitors and their positioning
- Our unique advantages in this segment
STEP 2 - CAPABILITY ANALYSIS:
- What product capabilities do we have?
- What gaps exist for enterprise needs?
- Estimated development cost and time for each gap
STEP 3 - GO-TO-MARKET CHALLENGE:
- What does enterprise sales require (team, process, timeline)?
- Expected CAC and sales cycle length
- First-year revenue realistic forecast
STEP 4 - RISK ASSESSMENT:
- What could go wrong?
- What could distract from our core business?
- What's the opportunity cost?
STEP 5 - DECISION FRAMEWORK:
- Under what conditions is this a YES?
- Under what conditions is this a NO?
- What information would change your recommendation?
STEP 6 - RECOMMENDATION:
Based on the above analysis, provide a clear recommendation with:
- Your confidence level (1-10)
- The 3 most important factors in your decision
- Next steps if we proceed
- Alternative approaches we should consider
Why this works: Each step builds on the previous one, forcing comprehensive analysis rather than surface-level thinking.
Advanced Scaffolding Patterns
The Comparison Scaffold:
Compare [Option A] vs [Option B] using this structure:
FOR EACH OPTION:
1. Core strengths (top 3)
2. Critical weaknesses (top 3)
3. Best-case scenario
4. Worst-case scenario
5. Hidden costs or complications
THEN COMPARE:
6. Which is better for [specific criterion]?
7. Which has lower risk?
8. Which has higher upside?
9. What would make you choose one over the other?
10. Is there a hybrid approach that takes the best of both?
The Problem-Solving Scaffold:
Problem: [describe problem]
STEP 1 - PROBLEM DISSECTION:
Break this into sub-problems. What are the 3-4 core issues?
STEP 2 - ROOT CAUSE ANALYSIS:
For each sub-problem, what's causing it? Go at least 2 levels deep.
STEP 3 - SOLUTION GENERATION:
For each root cause, generate 2-3 potential solutions.
Rate each: Impact (1-10), Difficulty (1-10), Time to implement.
STEP 4 - DEPENDENCIES & SEQUENCING:
What needs to happen first? What depends on what?
STEP 5 - RECOMMENDATION:
Given the above, what's the optimal sequence of actions?
What resources are needed? What could block success?
✍️ PRACTICE EXERCISE 2.2: Create Your Scaffold
Choose one of these scenarios:
- A) You need to decide whether to hire a senior developer or two junior developers
- B) You're evaluating whether to rebrand your company
- C) You're deciding which of three marketing channels to focus on
Your task: Create a complete analytical scaffold (6-8 steps) that would force thorough, logical analysis.
Then test it with an AI and see the difference in output quality.
Framework 4: Few-Shot Mastery
The most underrated expert technique: showing the AI exactly what good looks like through examples.
The Power of Examples
Few-shot learning means providing 2-4 examples of exactly what you want before asking the AI to generate something new.
Why it's powerful:
- Examples are more precise than descriptions
- The AI can pattern-match to your specific quality bar
- You control style, tone, format, and structure
- Works across all models
Few-Shot Structure
I need [type of output]. Here are examples of exactly what I want:
EXAMPLE 1:
[paste example]
EXAMPLE 2:
[paste example]
EXAMPLE 3:
[paste example]
KEY PATTERNS TO NOTICE:
- [element 1 that's important]
- [element 2 that's important]
- [element 3 that's important]
Now create one for: [your new case]
Few-Shot in Practice
Scenario: Product Descriptions
I need a product description for wireless earbuds.
Here are 3 examples of our brand voice:
EXAMPLE 1 (Phone Case):
"Your phone survives your life. Military-grade protection meets
slim design—because bulk is not a security strategy. $24."
EXAMPLE 2 (Laptop Stand):
"Your neck deserves better. Aluminum stand puts your screen
at eye level. Folds flat, works anywhere. $49."
EXAMPLE 3 (USB-C Cable):
"The cable that stays plugged in. Reinforced connector survives
12,000+ bends. Charges fast, transfers faster. $15."
PATTERNS TO MATCH:
- Opens with customer pain point
- Technical benefits in plain language
- Crisp, confident tone
- Ends with price
Now write for: Wireless earbuds, noise cancellation, 24hr battery, $129.
✍️ PRACTICE EXERCISE 2.3: Few-Shot Training
Your task: Find 3 examples of something you create regularly (emails, posts, reports, code).
Create a few-shot prompt with:
- Your 3 examples
- Explicit pattern notes
- A new case to generate
Test it and compare quality to outputs without examples.
Framework 5: Constraint-Based Creativity
Counter-intuitive truth: More constraints = better creativity.
The Constraint Paradox
Total freedom produces generic outputs. Smart constraints force originality.
Why constraints work:
- They eliminate obvious/generic options
- They force deeper search in possibility space
- They create novel combinations
- They give you control over output
The Constraint Formula
[Task] + [What to avoid] + [Format] + [Unusual element] = Original output
Example:
Write a job posting for senior engineer.
Constraints:
- Avoid: "rockstar," "ninja," "passionate," "fast-paced"
- Format: 4 paragraphs max, no bullets
- Include: One surprising perk, one honest challenge
- Tone: How you'd describe it to a friend
Write it.
✍️ PRACTICE EXERCISE 2.4: Constrain for Quality
Add 4-5 constraints to force better output:
- A) "Write blog post about time management"
- B) "Create social post about new product"
- C) "Explain blockchain technology"
Test constrained vs generic version. Notice the difference.
🎯 MODULE 2 CHECKPOINT
You've learned five advanced frameworks:
- Layered Prompting - Build in conceptual layers
- Constitutional Prompting - Set operating principles
- Chain-of-Thought Scaffolding - Provide thinking structures
- Few-Shot Mastery - Show exactly what you want
- Constraint-Based Creativity - Use limits for originality
Integration exercise: Create a competitive analysis using at least 3 frameworks combined.
MODULE 3: MODEL-SPECIFIC MASTERY
Getting the Best from Each AI Tool
Not all AI models are created equal. Each has distinct strengths, weaknesses, and quirks. Experts know how to optimize for each one.
ChatGPT (GPT-4) Mastery
Model Characteristics
- Strengths: Creative tasks, conversational flow, broad knowledge, coding, accessibility
- Weaknesses: Can be verbose, sometimes overconfident, may require more constraint for precision
- Best for: Brainstorming, content creation, explanations, code generation, conversational interactions
GPT-4 Optimization Techniques
Technique 1: Custom Instructions as Environment Variables
ChatGPT Plus allows custom instructions. Use these like programming environment variables.
Example Custom Instructions:
WHAT WOULD YOU LIKE CHATGPT TO KNOW:
- I'm a B2B SaaS founder, technical background, 8 years experience
- Company: project management tool, 50 customers, $30K MRR
- I value: directness, speed, practical over theoretical
HOW WOULD YOU LIKE CHATGPT TO RESPOND:
- Match my tone: casual = casual, formal = formal
- Show reasoning before complex answers
- Include: what could go wrong
- Keep under 300 words unless I ask for comprehensive
- Never use: "delve," "leverage," "synergy"
Technique 2: Verbosity Control
GPT-4 can be wordy. Control through language:
For concise output:
"Explain X. Be ruthlessly concise."
"Give me the 3-bullet version."
"Pretend I'm in an elevator, 30 seconds."
For comprehensive output:
"Give me a deep dive on X. Be thorough."
"Walk me through step-by-step, assume I know nothing."
Technique 3: Multi-Turn Sculpting
Instead of perfect first prompt, use rapid turns to sculpt:
Turn 1: "Draft cold email for enterprise sales."
[Review]
Turn 2: "Too salesy. Rewrite as peer reaching out with value."
[Review]
Turn 3: "Better. Cut by 40% without losing impact."
[Review]
Turn 4: "Perfect length. Stronger opening that shows research."
[Done - 5 minutes total]
Technique 4: Code Optimization
For coding, GPT-4 needs direction:
Expert request:
Write Python function to validate email addresses.
Requirements:
- Use regex for validation
- Handle: plus addressing, subdomains, international domains
- Return: (is_valid, error_message)
- Include docstring with examples
- Type hints required
Then:
- Explain regex pattern choice
- Note what valid emails this might reject
- Show 3 test cases
✍️ PRACTICE EXERCISE 3.1: GPT-4 Sculpting
Task: Product announcement via GPT-4
Steps:
- Write basic prompt, get output
- Use 3-4 sculpting turns to improve
- Document what each turn accomplished
Notice how much faster than writing "perfect" first prompt.
Claude (Sonnet 4.5) Mastery
Model Characteristics
- Strengths: Long-context reasoning, nuanced analysis, following complex instructions, structured thinking, precision
- Weaknesses: Can be formal, sometimes cautious
- Best for: Analysis, research synthesis, complex reasoning, detailed documentation, large documents
Claude Optimization Techniques
Technique 1: Extended Context Exploitation
Claude excels with large context. Feed everything relevant:
I'm uploading 4 documents:
1. Q1-Q3 board presentations (63 pages)
2. Customer interviews (24 interviews)
3. Product roadmap (15 pages)
4. Competitive analysis (8 companies)
Task: Identify gaps between customer asks and roadmap.
Recommend 3 Q4 priority features.
Requirements:
- Reference specific quotes and slides
- Note contradictions between documents
- Include: customer evidence, competitive context, effort estimate
- Flag assumptions due to missing info
Technique 2: Artifact-Driven Workflows
Claude creates "artifacts" you can iterate on:
Create PRD for [feature] as artifact.
Include:
- Problem statement
- User stories (3-5)
- Technical requirements
- Success metrics
- Open questions
Then critique as 3 stakeholders:
1. Engineering Lead: Technical feasibility
2. Customer Success: User adoption risks
3. Product Marketing: Positioning challenges
Update PRD after each critique.
Technique 3: Meta-Cognitive Prompting
Claude responds well to prompts about its thinking:
Analyze [business problem].
Before your analysis:
1. What framework is most useful here? Why?
2. What key information is missing?
3. What's the strongest counter-argument?
4. What assumptions are you making? Rate confidence (1-10).
Now provide analysis.
Technique 4: Structured Thinking Frameworks
Claude excels with explicit structures:
Evaluate European market expansion.
Framework:
SECTION 1 - OPPORTUNITY:
Market size, growth, competitive landscape
SECTION 2 - CAPABILITY:
What we have, what we need, gaps
SECTION 3 - RISK MAPPING:
Regulatory, operational, financial, competitive
Rate: Likelihood (1-10), Impact (1-10)
SECTION 4 - SCENARIOS:
Best case, base case, worst case
SECTION 5 - DECISION:
Go/No-Go criteria
✍️ PRACTICE EXERCISE 3.2: Claude Analysis
Task: Complex business scenario with documents
Use meta-cognitive prompting for deeper analysis.
Compare to simple prompt output quality.
Perplexity Mastery
Model Characteristics
- Strengths: Research, current information, source synthesis, citations
- Weaknesses: Shorter responses, less creative, more constrained
- Best for: Research, fact-finding, comparing sources, staying current
Perplexity Optimization
Technique 1: Research Query Structuring
Weak:
"What's happening with AI regulation?"
Strong:
Compare AI regulation: US, EU, China (2024-2025):
Focus:
- Major legislation passed/proposed
- Definition of 'high-risk AI' per jurisdiction
- Enforcement mechanisms and penalties
- Industry response and compliance challenges
Prioritize: government docs, official statements, policy papers
Include: company examples affected by each approach
Technique 2: Source Quality Control
Research [topic].
Source requirements:
- Prioritize: Academic papers, gov reports, company filings
- Avoid: News aggregators, opinion pieces, marketing
- Date range: Last 6 months
- Include contradicting viewpoints if they exist
Cite sources with publication date.
Technique 3: Iterative Research
Turn 1: "Overview of quantum computing commercial applications 2024-2025"
[Review]
Turn 2: "Focus on top 3 nearest-term applications. Leading companies?"
[Review]
Turn 3: "For quantum optimization: technical blockers, timelines, expert predictions"
[Review]
Turn 4: "Find case studies or pilots testing quantum optimization. Results?"
Technique 4: Comparative Research
Compare [Tech A] vs [Tech B] for [use case]:
Structure:
1. Technical maturity (evidence/milestones)
2. Current adoption rate (numbers)
3. Cost considerations (pricing range)
4. Key limitations (technical, not marketing)
5. Expert predictions 2025-2026
For each: cite sources, note if experts disagree
✍️ PRACTICE EXERCISE 3.3: Research Challenge
Choose research topic for work/project.
Use Perplexity with:
- Well-structured initial query
- Source quality requirements
- 2+ rounds of iterative refinement
Document insights vs simple Google search.
Model Selection Framework
Which Tool for Which Task?
Use ChatGPT when:
- Brainstorming and ideation
- Creative content generation
- Conversational interactions
- Code generation and debugging
- Quick iterations and sculpting
Use Claude when:
- Deep analysis of complex problems
- Working with large documents
- Structured thinking required
- Detailed documentation
- Nuanced reasoning needed
Use Perplexity when:
- Researching current events
- Finding and citing sources
- Comparing multiple viewpoints
- Fact-checking and verification
- Staying up-to-date
Cross-Model Workflow
Experts often use multiple models in sequence:
Example: Product Launch Strategy
1. Perplexity: Research market trends, competitor analysis
2. Claude: Synthesize research into strategic framework
3. ChatGPT: Generate creative campaign ideas
4. Claude: Evaluate and refine best ideas
5. ChatGPT: Write final copy and content
🎯 MODULE 3 CHECKPOINT
You now understand optimization for:
- ChatGPT - Creative, conversational, needs sculpting
- Claude - Analytical, handles complexity, loves structure
- Perplexity - Research-focused, needs specific queries
Integration Exercise:
New product launch needs:
- Market research (which tool?)
- Positioning strategy (which tool?)
- Marketing copy (which tool?)
Task: Map each to optimal tool and write expert prompts.
MODULE 4: TROUBLESHOOTING LIKE AN EXPERT
Diagnosing and Fixing Common Problems
Even expert prompts sometimes produce suboptimal outputs. The difference is that experts know how to diagnose and fix problems quickly.
The Expert Troubleshooting Framework
- IDENTIFY THE GAP - What specifically is wrong?
- DIAGNOSE ROOT CAUSE - Why did the model produce this?
- APPLY THE FIX - What modification addresses it?
- VALIDATE - Did it work? If not, iterate.
Problem 1: Generic / Bland Output
The Symptom
AI gives advice that sounds like corporate handbook or generic blog.
Example:
Prompt: "How can I improve team productivity?"
Output: "To improve team productivity:
1. Set clear goals
2. Foster communication
3. Provide right tools
4. Recognize good work
5. Encourage work-life balance"
Useless—everyone knows this.
The Diagnosis
Root cause: Model is pattern-matching to generic training data. Giving "average" response.
The Fix
Add specificity and constraints:
Improve team productivity. Specific situation:
- Team: 7 engineers, 2 designers
- Current: Shipping features but missing deadlines by 30%
- Constraints: Already 50+ hour weeks, morale decent
- Tools: Jira, Slack, GitHub
- Recent: Switched to 2-week sprints 3 months ago
Requirements:
- No generic productivity tips
- Focus on root cause, not symptoms
- Consider time/effort is maxed
- Suggest what to STOP doing
- Include validation methods
What's actually going on?
Problem 2: Factual Errors / Hallucinations
The Symptom
AI states something confidently that's incorrect or makes up details.
The Diagnosis
Root cause: Models predict plausible text, not truth. Can't distinguish fact from fiction.
The Fix
Strategy 1: Verification Loops
Research [topic].
After response:
1. List every factual claim you made
2. Rate confidence in each (1-10)
3. For claims with confidence <8, mark [VERIFY]
4. Suggest how I could verify [VERIFY] claims
Strategy 2: Request Sources
Research [topic]. For every factual claim, cite source.
Format: [claim](source URL)
If uncertain, say "I'm uncertain about [X] because [reason]"
rather than guessing.
Strategy 3: Cross-Validation
- Get response from AI
- Ask: "What might be wrong about this?"
- Verify key claims independently
- Feed corrections back if needed
Problem 3: Not Following Complex Instructions
The Symptom
Detailed instructions written, but AI ignores parts or gets confused.
Example:
Asked for: 5 sections, each with examples and data
Got: 3 sections, no examples, generic statements
The Diagnosis
Root causes:
- Instructions too long/complex (working memory overload)
- Ambiguous instructions
- Conflicting requirements
- Model lost track mid-generation
The Fix
Strategy 1: Chunking + Confirmation
Task in 3 steps. I'll describe all, you confirm, then we execute one at a time.
STEP 1: Analyze data, identify 5 key trends
STEP 2: For each trend, find 2 supporting examples
STEP 3: Create summary table with trends, examples, implications
Confirm: What are you doing in each step?
[Wait for confirmation]
Good. Start Step 1.
Strategy 2: Checklist Method
Create market analysis report.
Checklist (confirm you'll include all):
- [ ] Executive summary (2-3 paragraphs)
- [ ] Market size with numbers
- [ ] 3-5 key trends with evidence
- [ ] Competitive landscape (4+ companies)
- [ ] SWOT analysis in table
- [ ] 3 strategic recommendations
- [ ] Sources cited
Confirm checklist, then begin.
Strategy 3: Simplify + Iterate
Instead of: "Do A, B, C, D, E all at once"
Try: "Let's start with A. Once done, we'll do B."
Problem 4: Wrong Tone or Style
The Symptom
Output is too formal when you want casual, or too casual when you want professional.
The Diagnosis
Root cause: Without explicit guidance, models default to "safe" professional tone.
The Fix
Strategy 1: Tone Specification
Write email to colleague Jake about project delay.
Tone requirements:
- Talk like to a friend, not formal business
- Use "I" and "you," not "we" or "one"
- Contractions fine (we're, didn't)
- Direct—no corporate jargon
- Conversational but professional
Think: over coffee, not quarterly report.
Strategy 2: Provide Reference
Write [content].
Tone example (different topic):
"Hey Sarah—quick update on mockups. Pushed to your
review stack, but heads up: mobile needs work. Thinking
we simplify the nav. Thoughts?"
Match this: casual, direct, brief, human.
Strategy 3: Iterative Adjustment
Turn 1: [Get output]
Turn 2: "Too formal. Rewrite like texting colleague, not memo."
Turn 3: "Better, still stiff. Sound like how people talk."
Problem 5: Too Long or Too Short
The Symptom
Wall of text when you wanted summary, or bullets when you needed depth.
The Diagnosis
Root cause: No explicit length guidance.
The Fix
Strategy 1: Explicit Constraints
Explain [topic].
Length: Exactly 3 paragraphs, ~150 words total.
Structure:
- Para 1: Core concept (2-3 sentences)
- Para 2: Why it matters (2-3 sentences)
- Para 3: Common misconception (2-3 sentences)
Strategy 2: Compression/Expansion
Turn 1: "Explain [topic]"
[Review]
Turn 2: "Cut by 60% without losing key insights."
OR
Turn 2: "Expand with 3 concrete examples and detail."
Strategy 3: Reference Length
"Write this in about same length as:
[paste example of desired length]"
Problem 6: Lacks Specific Examples
The Symptom
All theory, no practical examples.
The Diagnosis
Root cause: Default to abstract explanations unless prompted for concrete examples.
The Fix
Strategy 1: Explicit Requirements
Explain [concept].
Requirements:
- For every principle, provide concrete example
- Examples must include specific numbers, names, scenarios
- No abstract examples (bad: "like a company might do X")
- Good: "like how Slack uses X to achieve Y"
Minimum 3 examples total.
Strategy 2: Example-First Structure
Explain [concept].
Structure:
1. Start with specific, concrete example
2. Extract general principle from example
3. Show 2-3 variations
4. End with counter-example (when doesn't work)
Problem 7: Doesn't Challenge Your Thinking
The Symptom
AI agrees with everything or gives what you asked without pushback.
The Diagnosis
Root cause: Models trained to be helpful and agreeable. Won't naturally play devil's advocate.
The Fix
Strategy 1: Request Challenge
Here's my plan: [describe]
Your job:
1. Identify 3 weakest parts
2. Present strongest counter-argument
3. What am I not considering?
4. Why might this fail?
Be direct. I want critique, not validation.
Strategy 2: Red Team
I'm proposing [decision/strategy].
Act as three skeptics:
1. Pessimist: What's wrong?
2. Competitor: How exploit weaknesses?
3. Experienced advisor: What mistake am I repeating?
Each provides perspective, then I'll revise.
The Expert Troubleshooting Checklist
When output quality is poor:
Content Issues:
- ☐ Too generic? → Add specific context and constraints
- ☐ Wrong? → Add verification loops and sources
- ☐ Lacks examples? → Require concrete, specific examples
- ☐ Too abstract? → Force practical application
Structure Issues:
- ☐ Ignoring instructions? → Chunk task, use confirmation
- ☐ Wrong length? → Set explicit length constraints
- ☐ Poorly organized? → Provide clear structure
Style Issues:
- ☐ Tone wrong? → Specify tone with examples
- ☐ Too agreeable? → Request challenge and critique
- ☐ Too cautious? → Reframe request
✍️ PRACTICE EXERCISE 4.1: Diagnostic Practice
Three problematic outputs. For each: diagnose, fix, test.
Problem A:
Asked: "How should I price my SaaS?"
Got: "Consider value-based, competitive, cost-plus pricing. Research market and test."
Problem B:
Asked for 150-word ML explanation
Got: 450 words of dense technical content
Problem C:
Asked: "Critique my go-to-market strategy"
Got: "Solid strategy with clear focus and good channels. Timeline reasonable, budget sensible."
Write improved prompts, test them.
🎯 MODULE 4 CHECKPOINT
You now have systematic approaches to:
- Diagnose why outputs are poor
- Apply specific fixes for common problems
- Iterate efficiently to quality
- Build quality controls into prompts
Integration Exercise:
Take a prompt you use regularly that doesn't always work. Apply framework:
- What specifically goes wrong?
- What's the root cause?
- What fix addresses it?
- Test and validate
Document before/after results.