Prompt Engineering Techniques
📘 Introduction
Prompt engineering is the art and science of crafting inputs that guide Large Language Models (LLMs) to produce desired outputs. While anyone can write a prompt, effective prompt engineering requires an understanding of LLM behavior, configuration tuning, and iterative testing.
Based on Google’s 2024 whitepaper, this guide breaks down strategies, parameters, and real-world examples to help you get the most from any LLM.
⚙️ LLM Configuration Essentials
| Setting | Description | Tips |
|---|---|---|
| Temperature | Controls randomness in output | Use 0 for deterministic, 0.9+ for creativity |
| Top-K | Sample from top K tokens | Lower K for focus, higher for diversity |
| Top-P | Sample from top tokens within cumulative probability P | 0.9–0.95 is a good balance |
| Token Limit | Controls max length of output | Impacts cost and clarity |
✅ Recommended defaults: temperature=0.2, top-P=0.95, top-K=30.
🧪 Prompting Techniques (with Examples)
🔹 Zero-Shot Prompting
Use: Simple tasks where the model can generalize well.
Prompt:
Translate this to French: "Where is the nearest restaurant?"
🔹 One-Shot Prompting
Use: When the model needs guidance on format or tone.
Prompt:
Example:
Input: "What is the capital of France?"
Output: "The capital of France is Paris."
Now, answer this:
Input: "What is the capital of Japan?"
🔹 Few-Shot Prompting
Use: Tasks with variability; adds consistency by showing patterns.
Prompt:
Q: What is 5 + 3?
A: 8
Q: What is 12 - 4?
A: 8
Q: What is 9 + 6?
A:
🔹 System Prompting
Use: Guide output format, tone, or persona via system-level instruction.
Prompt:
You are a JSON API assistant. Always respond in valid JSON format.
User: "Tell me the current weather in London"
🔹 Role Prompting
Use: Assign a personality or function to steer response style.
Prompt:
You are a helpful personal finance advisor.
What’s a good way to save for retirement in your 30s?
🔹 Contextual Prompting
Use: Provide real data or background to ground the answer.
Prompt:
Here’s the company’s 2023 HR policy: [insert excerpt]
Based on this policy, can an employee carry over unused vacation days to next year?
🔁 Advanced Prompting Strategies
🔸 Step-Back Prompting
Prompt:
Let’s first reflect on the broader question:
"What factors should we consider before choosing a new CRM platform?"
Now, given those, which platform is best for a mid-sized SaaS startup?
🔸 Chain-of-Thought (CoT)
Prompt:
Q: Jane has 5 apples. She buys 7 more, then gives 3 to her friend. How many apples does she have now?
A: Let's think step by step...
🔸 Self-Consistency
Approach:
- Run the same prompt multiple times.
- Use majority voting to find the most consistent answer.
Prompt (run 3x):
What’s the next number in this sequence: 2, 4, 6, 8, ?
🔸 Tree-of-Thought (ToT)
Use: Let the model explore multiple branches of reasoning.
Prompt:
You are solving a puzzle. First, generate 3 different strategies to solve it. Then evaluate which one is most effective and explain why.
🔸 ReAct (Reason + Act)
Use: Combine reasoning with tool use.
Prompt:
User: What’s the weather in Tokyo right now?
Assistant Thought: I need to look up the weather using the weather API.
[Call: GET https://api.weather.com/tokyo]
Action: Retrieve weather info
Observation: It’s 22°C and sunny
Answer: The weather in Tokyo is currently 22°C with clear skies.
💻 Prompting for Code Tasks
Prompt engineering also works well with LLMs like Gemini or Claude for tasks like:
- Writing Bash scripts
- Explaining code
- Refactoring
- Translation (e.g., Python to JavaScript)
Prompt Example:
Convert the following Python list comprehension to a standard for loop:
[ x**2 for x in range(10) if x % 2 == 0 ]
✅ Best Practices Summary
- 🎯 Be clear, concise, and direct
- 🧱 Use examples where helpful
- 💬 Keep format structured
- 🧪 Test and iterate
- 🧰 Abstract common prompts with variables or templates
- 🔒 Stay aligned with model safety and bias guidelines
📌 Final Thoughts
Prompt engineering is your superpower when working with LLMs. It's part design, part trial-and-error, and part understanding the model’s training behavior. With the right strategy, even complex workflows become simple, reusable, and reliable.