Skip to content

Tips & Tricks

Large Language Models (LLMs) are powerful tools, but their effectiveness largely depends on how we interact with them. Here are proven strategies to maximize their potential:

Chain of Thought Prompting

Chain of Thought (CoT) prompting is a technique that encourages LLMs to explain their reasoning process. By simply asking the model to share its thought process or solve problems step by step, you can significantly improve the accuracy and reliability of responses. This approach is particularly effective for:

  • Complex reasoning tasks
  • Mathematical problems
  • Logic puzzles
  • Decision-making scenarios
  • Analysis and evaluation

There are several ways to implement CoT prompting:

Simple Thought Process Request

Basic prompt: "Is this a prime number: 377?"

Better prompt: "Think through this step by step: Is 377 a prime number?"

Explicit Step-by-Step Instructions

Basic prompt: "Solve this word problem: If John has 3 apples and gives half to Mary, who then gives a third of her apples to Tom, how many apples does Mary have?"

Better prompt: "Let's solve this word problem step by step:

  1. Start with John's apples
  2. Calculate Mary's initial share
  3. Calculate how many she gives to Tom
  4. Determine Mary's final amount

If John has 3 apples and gives half to Mary, who then gives a third of her apples to Tom, how many apples does Mary have?"

Show Your Work Format

Basic prompt: "What will the weather be like if the barometric pressure drops rapidly?"

Better prompt: "Walk me through your reasoning: What will the weather be like if the barometric pressure drops rapidly?"

Benefits of CoT

  • Increases transparency in the model's decision-making process
  • Makes it easier to identify potential errors or biases
  • Allows for better verification of the logic used
  • Improves the model's accuracy on complex tasks
  • Provides educational value by demonstrating problem-solving approaches

When to Use CoT

CoT is most valuable when:

  • The task requires multiple steps of reasoning
  • Accuracy is crucial
  • You need to verify the model's logic
  • The problem has multiple interdependent parts
  • You want to understand how the model reached its conclusion

Expert Personas

Adopting specialized personas can enhance the quality and relevance of responses. When an LLM assumes the role of an expert in a specific field, it tends to provide more focused, domain-specific information.

Example:

Poor prompt: "How do I fix this Python code?"

Better prompt: "You are an world-class Python developer who specializes in optimization and clean code. Please review this code and suggest improvements, focusing on both functionality and best practices."

Iterative Prompt Refinement

Instead of continuing a conversation when responses aren't ideal, refining the initial prompt often yields better results. Look for an edit button (e.g., a pencil) next to chat bubbles to refine your prompt. This approach:

  • Maintains focus on the core question
  • Reduces context window usage
  • Prevents accumulation of potential misunderstandings

Example:

Initial prompt: "Tell me about quantum computing" (Response too general)

Refined prompt: "Explain the fundamental principles of quantum computing, focusing on qubits, superposition, and entanglement. Include practical applications in cryptography and optimization."

Exemplar-Based Learning

Providing concrete examples of desired outputs helps LLMs understand expectations more precisely. This technique is particularly effective when you need specific formats or styles.

Example:

Poor prompt: "Write a product description"

Better prompt: "Write a product description following this format:

Good example: 'The Ultra Comfort Pillow features memory foam technology that adapts to your sleeping position. With its hypoallergenic cover and temperature-regulating core, it ensures peaceful sleep all night long. Perfect for side, back, and stomach sleepers.'

Bad example: 'This is a pillow. It's soft and comfortable. You can sleep on it.'

Please write a description for our new ergonomic office chair."

Contextual Priming

Before asking the main question, provide relevant context and background information. This helps the LLM understand the scope and requirements more accurately.

Example:

Poor prompt: "How can I improve my website's performance?"

Better prompt: "I run an e-commerce website built with React and Node.js with versions 18.0.0 and 16.0.0, respectively, serving approximately 10,000 daily users. We're experiencing slow page load times, especially during peak hours. Our current average load time is 4.5 seconds. How can we improve our website's performance?"