Skip to main content

Lily's Lab

Back to all posts

Best Practices for Pro Prompt Engineering

Artificial-Intelligence

Best Practices for Pro Prompt Engineering

Becoming a pro at prompt engineering takes practice, but following these best practices from the sources will get you on the right track:

  • Provide Examples: Using one-shot or few-shot examples is a highly effective way to guide the LLM. They show the desired output and help the model understand your expectations.
  • Design with Simplicity: Keep your prompts clear, concise, and easy to understand, for both you and the model. Avoid complex language and unnecessary information.
  • Be Specific About the Output: Clearly define what you want the LLM to produce. Provide specific details using system or contextual prompting to focus the model.
    • Example: Instead of "Write a blog post about video game consoles," try "Generate a 3 paragraph blog post about the top 5 video game consoles. The blog post should be informative and engaging, and it should be written in a conversational style.".
  • Use Instructions Over Constraints: Focus on telling the LLM what to do rather than what not to do. Positive instructions are generally more effective.
    • Example: Instead of "Do not list video game names," try "Only discuss the console, the company who made it, the year, and total sales.".
  • Control the Max Token Length: Be mindful of the length of the LLM's response. You can set a maximum number of tokens in the model's configuration or explicitly request a certain length in your prompt.
  • Use Variables in Prompts: For reusable and dynamic prompts, use variables that you can change for different inputs. This is useful when integrating prompts into applications.
    • Example: "You are a travel guide. Tell me a fact about the city: {city}" where "{city}" can be replaced with "Amsterdam" or any other city.
  • Experiment with Input Formats and Writing Styles: Try different ways of phrasing your prompts (questions, statements, instructions) and different writing styles to see what yields the best results.
  • For Few-Shot Classification, Mix Up Classes: When providing examples for classification tasks, ensure you mix the different categories in your examples to avoid biasing the model.
  • Adapt to Model Updates: Stay informed about new model versions and features and adjust your prompts accordingly to leverage these improvements.
  • Experiment with Output Formats: For non-creative tasks, try requesting the output in a structured format like JSON or XML for easier parsing.
  • JSON Repair: If you're working with JSON output and encounter incomplete or invalid JSON due to token limits, consider using tools like json-repair to automatically fix it.
  • Working with Schemas: Use JSON Schemas to define the expected structure and data types of your input, providing the LLM with a clear blueprint.
  • Experiment Together: If possible, collaborate with other prompt engineers to generate and test different prompt approaches.
  • CoT Best Practices: For Chain of Thought prompting, put the answer after the reasoning and set the temperature to 0 for more deterministic results.
  • Document Your Attempts: Keep a detailed record of your prompts, model configurations, and results. This helps you learn what works and track your progress. Use a template like the one provided in the source.

By understanding these techniques and following these best practices, you'll be well on your way to becoming a proficient prompt engineer. Remember that prompt engineering is an iterative process, so keep experimenting and refining your prompts.

![[LilysLab-Logo.png]]