AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff

Introduction

This tutorial provides a comprehensive guide to AI prompt engineering as discussed by Sander Schulhoff in his recent podcast episode. It outlines effective techniques for crafting prompts, common pitfalls to avoid, and insights into AI security and red teaming. Whether you're new to AI or looking to enhance your skills, this guide will help you understand what works in prompt engineering and how to apply these techniques effectively.

Step 1: Understand the Importance of Prompt Engineering

• Recognize its role: Prompt engineering is crucial for guiding AI models to produce desirable outputs.
• Acknowledge its evolution: Techniques have changed significantly, necessitating a modern approach to prompting.

Step 2: Explore Effective Prompt Engineering Techniques

Sander identifies five effective techniques for prompt engineering, including:

1. Few-shot prompting:

   • Provide a few examples within your prompt to guide the AI.

   Example:

   "Translate the following sentences into French

   1. Hello, how are you? 2. What is your name?"

2. Decomposition:

   • Break down complex tasks into smaller, manageable steps.
   • Ask the AI to solve each part sequentially.

3. Self-criticism and context:

   • Encourage the AI to evaluate its own responses.
   • Include context to guide the AI’s reasoning.

4. Ensembling:

   • Combine outputs from multiple models or prompts to improve accuracy.
   • This can help mitigate errors from individual responses.

5. Thought generation:

   • Prompt the AI to generate ideas or brainstorm solutions.
   • Use open-ended questions to elicit creative responses.

Step 3: Identify Prompting Techniques to Avoid

• Role prompting: Avoid trying to make the AI act as if it has a specific identity; this often leads to irrelevant or nonsensical responses.
• Threatening the AI: This method is ineffective and can produce unpredictable results.

Step 4: Differentiate Between Prompt Types

• Conversational prompts: Use for engaging with users in a dialogue-like manner.
• Product/system prompts: Tailored for specific tasks or functionalities, usually more structured.

Step 5: Learn About Prompt Injection and AI Red Teaming

• Understand prompt injection: This refers to techniques that manipulate AI models to produce unintended outputs.
• Explore AI red teaming: Participate in competitions like HackAPrompt to learn about vulnerabilities and improve AI security.

Step 6: Implement Defense Strategies

• Common defenses that fail: Be aware of ineffective strategies that don’t protect against prompt injection.
• Effective defenses: Focus on robust validation and input sanitization processes to safeguard your AI products.

Conclusion

Prompt engineering is an essential skill for anyone working with AI. By understanding effective techniques, recognizing what to avoid, and implementing security measures, you can create more reliable and secure AI systems. For further learning, consider exploring Sander Schulhoff’s resources, such as his AI red teaming courses and guides to deepen your understanding of this evolving field.