The Art & Science of Prompt Engineering

Prompt engineering is the highest-leverage skill in AI development. Well-crafted prompts improve output quality by 50-80%, reduce token costs by 30-50%, and often eliminate the need for fine-tuning entirely.

Core Techniques

1. Zero-Shot Prompting

Direct instruction without examples:

Classify the sentiment of this review as positive, negative, or neutral:
"The product arrived late but quality is excellent."
      

When to use: Simple tasks, quick prototyping. Accuracy: 60-75%.

2. Few-Shot Learning

Provide 2-5 examples of desired input-output pairs:

Extract key information from customer emails:

Email: "Hi, I need to return order #12345. It's too small."
Output: {"order_id": "12345", "intent": "return", "reason": "size"}

Email: "When will order #67890 ship?"
Output: {"order_id": "67890", "intent": "shipping_inquiry", "reason": null}

Email: "I want to cancel my subscription immediately."
Output:
      

Results: 75-90% accuracy with 3-5 examples. Optimal for most tasks.

3. Chain-of-Thought (CoT)

Ask model to explain reasoning step-by-step:

Problem: A store sold 48 apples. 1/3 were green, rest were red. How many red apples?

Solve this step-by-step:
1. Calculate number of green apples
2. Calculate number of red apples
3. Provide final answer
      

Results: 30-50% improvement on complex reasoning tasks.

4. Role Prompting

Assign specific expertise to the model:

You are an expert Python developer with 15 years of experience in data engineering.
Review this code for bugs and suggest improvements:

[code here]
      

Impact: 20-40% improvement in specialized tasks.

5. System Messages (ChatGPT/Claude)

System: You are a helpful customer service agent for TechCorp. Be concise,
professional, and always provide order numbers when discussing purchases.
Never make promises about refunds without manager approval.

User: I want my money back!
      

Use case: Set consistent behavior and constraints.

6. Output Formatting

Request specific output format:

Extract company information and return as JSON:

Text: "Apple Inc., founded in 1976, is headquartered in Cupertino, CA."

Output format:
{
  "company": "company name",
  "founded": "year",
  "headquarters": "location"
}
      

Advanced Techniques

ReAct (Reasoning + Acting)

Combine reasoning with tool use:

You have access to: [search_web, calculate, get_weather]

Question: What's the total GDP of countries with population > 100M?

Thought: I need to find countries with population > 100M
Action: search_web("countries population over 100 million")
Observation: [list of countries]
Thought: Now I need their GDPs
Action: search_web("GDP of [countries]")
...
      

Self-Consistency

Generate multiple reasoning paths, choose most common answer:

• Run same prompt 5-10 times with temperature 0.7
• Use majority voting on final answers
• 20-30% accuracy improvement on complex reasoning

Parameter Tuning

Temperature

• 0.0-0.3: Deterministic, factual (data extraction, classification)
• 0.7-0.9: Balanced creativity (content generation, conversations)
• 1.0+: Highly creative (brainstorming, creative writing)

Top-p (Nucleus Sampling)

• 0.1: Very focused (use with low temperature)
• 0.9: Balanced (default for most tasks)
• 1.0: Maximum diversity

Max Tokens

• Set generous limits for reasoning tasks (1000-2000 tokens)
• Constrain for simple tasks (100-300 tokens)
• Monitor cost vs quality tradeoff

Model-Specific Tips

GPT-4

• Excels at complex reasoning and coding
• Use system messages for consistent behavior
• 32K context: leverage long conversations and documents
• Best for: code generation, analysis, complex reasoning

Claude 2/3

• 100K-200K context: process entire books
• Strong at following detailed instructions
• More conservative, less prone to hallucination
• Best for: document analysis, content moderation, safe responses

GPT-3.5 Turbo

• 10x cheaper than GPT-4
• Use for simple tasks: classification, simple Q&A, summarization
• Requires more explicit prompts than GPT-4

Common Pitfalls

1. Vague Instructions: Be specific about format, length, tone
2. No Examples: Add 2-3 examples for 30-50% accuracy boost
3. Wrong Temperature: Low for factual, high for creative
4. Ignoring Context Window: Monitor token usage, summarize long conversations
5. Not Testing Variations: A/B test prompts, measure quality

Testing & Optimization

1. Create Test Set: 50-100 representative examples
2. Iterate Prompts: Test variations systematically
3. Measure Quality: Accuracy, completeness, format compliance
4. Monitor Costs: Track tokens per request, optimize length
5. Version Control: Track prompt changes and performance

Real-World Examples

Customer Email Classification

• Technique: Few-shot (3 examples) + output formatting (JSON)
• Model: GPT-3.5 Turbo
• Result: 92% accuracy, 0.5s latency, ₹0.02 per classification

Code Review Agent

• Technique: Role prompting + chain-of-thought + system message
• Model: GPT-4
• Result: 87% bug detection rate, comparable to human reviewers

Legal Contract Analysis

• Technique: Claude 2 (100K context) + structured output
• Result: Process 50-page contracts in 30 seconds vs 2 hours human time