The art and science of communicating with AI to achieve professional outcomes
In the age of Large Language Models (LLMs) like ChatGPT, Claude, and Gemini, prompt engineering has become as essential as:
Prompt Engineering is the practice of designing and refining inputs (prompts) to generative AI systems to produce desired, accurate, and useful outputs.
| Week | AI Type | Key Capability | Prompt Engineering Role |
|---|---|---|---|
| 2-3 | Predictive ML | Pattern recognition in data | Data preparation, feature selection |
| 4 | Causal ML | Understanding cause-effect | Hypothesis formulation |
| 7 | Generative AI | Content creation | Prompt engineering |
| 8 | Prompt Engineering | Controlling AI outputs | Today's focus |
Moving from correlation to causation to generation requires increasingly sophisticated human-AI interaction
CRAFT Method - Developed for business analytics
C: "Analyze the customer churn patterns in the telecom dataset"
R: "As a data scientist specializing in customer retention"
A: "Identify the top 5 features that predict churn and explain their business significance"
F: "Present as: (1) ranked list with coefficients, (2) business interpretation for each, (3) actionable recommendations"
T: "Use professional language suitable for C-suite presentation"
Asking the AI to "show its work" by reasoning step-by-step, leading to:
Few-Shot Learning: Providing examples of desired input-output pairs to guide the AI's responses.
Issue: AI can perpetuate or amplify biases in training data
Action: Explicitly prompt the AI to check for and mitigate bias
Issue: Never input confidential or personally identifiable information
Action: Anonymize data before using in prompts
Rule: Assume everything you input could be seen by others
Issue: AI-generated content may inadvertently copy protected works
Action: Review outputs for originality; cite AI assistance
Issue: AI can "hallucinate" false information confidently
Action: Always verify facts, statistics, and citations
Rule: AI is a tool, not a source of truth
When AI generates false or misleading information presented as fact.
Prompt engineering is iterative, not one-and-done:
Initial Prompt → AI Response → Evaluate Quality → Refine Prompt → Better Response
| If the output is... | Then refine by... |
|---|---|
| Too generic | Adding specific constraints and examples |
| Too verbose | Specifying length limits or bullet points |
| Missing context | Providing more background information |
| Wrong format | Explicitly stating the desired structure |
| Inaccurate | Requesting step-by-step reasoning and sources |
| Off-topic | Narrowing the scope and focus |
Scenario: Weekly sales performance reports
Key Elements:
Scenario: Analyzing 1000+ customer reviews
Key Elements:
Scenario: Interpreting ML model outputs
Explaining feature importances to non-technical stakeholders with business interpretation and retention strategies
Scenario: Evaluating dataset readiness
Provide data quality score, identified issues, impact assessment, and cleaning recommendations with code
"Tell me about sales"
Problem: Too broad, no context or goal
Fix: Specify what aspect, time period, and desired insight
[Paste 10 pages of data] "What does this mean?"
Problem: No guidance on what to analyze
Fix: Provide context, specific questions, and analysis goals
"Explain why our sales dropped last quarter"
Problem: Assumes facts not in evidence
Fix: "Analyze our sales data to determine IF sales dropped and identify possible causes"
"Write some Python code for data analysis"
Problem: No specifics on data, goal, or methods
Fix: Specify data structure, analysis type, libraries, and output format
"Analyze this data and tell me everything about it"
Problem: Too many possible directions, unfocused
Fix: Ask specific, targeted questions one at a time
| Platform | Strengths | Best For |
|---|---|---|
| ChatGPT (GPT-4) | Strong general knowledge, code generation | Versatile business tasks, Python coding |
| Claude | Strong analytical reasoning, large context | Complex analysis, document processing |
| Google Gemini | Integrated with Google Workspace | Data analysis with Sheets/Docs |
| GitHub Copilot | Code-specific, IDE integration | Software development, coding tasks |
| Perplexity | Research-focused, cites sources | Fact-finding, research |
Best Practice: Never input confidential or sensitive data. Understand each platform's data usage policies.
| Stage | Human Role | AI Role |
|---|---|---|
| 1. Problem Definition | Define business question | Help refine and scope the problem |
| 2. Data Preparation | Gather data, understand quality | Suggest cleaning approaches, generate code |
| 3. Exploratory Analysis | Direct the exploration | Generate visualizations, identify patterns |
| 4. Modeling & Analysis | Select approaches, interpret results | Implement models, explain outputs |
| 5. Insight Generation | Apply business context and judgment | Draft explanations, suggest implications |
| 6. Communication | Tailor message, make decisions | Help structure reports, refine language |
Key Principle: AI is a collaborator, not a replacement for human expertise, judgment, and accountability.
Combining text, images, data, and voice for holistic analysis
AI that can plan and execute multi-step tasks with less prompting
LLMs fine-tuned for specific industries (Finance AI, Healthcare AI, etc.)
AI systems that evaluate other AI outputs with automated fact-checking
Assessment 2 (35%): Generative AI Startup Project - Due Week 9
Prompt: "Explain what each column in this dataset likely represents and suggest appropriate data types and potential quality issues."
Prompt: "I need to perform predictive and prescriptive analytics on [describe data]. Suggest an appropriate analytical approach."
Prompt: "Generate Python code to [specific task] using [libraries]. Include error handling and detailed comments."
Prompt: "Explain these model results in business terms suitable for a non-technical stakeholder."