10 Prompt Engineering Tips and Best Practices for LLM: Part 1

As software developers and technical professionals, prompt engineering plays a crucial role in our work. With the rise of Large Language Models (LLMs), it becomes even more important to optimize prompt engineering for effective communication with these powerful models. In this article, we will explore 10 tips and best practices specifically tailored to LLM prompt engineering. Let’s dive in!

1. Define Clear and Specific Prompts

When working with LLMs, it’s essential to provide clear and specific prompts to guide the model’s behavior. Clearly define the task and desired output to obtain accurate and relevant responses.

prompt = """  
Translate the following English text to French:  
   
"Hello, how are you?"  
"""  

AI systems lack the perception and intuition necessary to comprehend the needs and desires of users. Prompts need to be carefully worded, formatted, and contain clear details. The slang, metaphors, and social subtleties that people often take for granted in casual conversation should frequently be avoided in prompts.

2. Experiment with Different Prompt Formats

LLMs can process prompts in various formats. Experiment with different styles such as question-answering, completion, or instruction-based prompts. Test different formats to find the most effective one for your specific use case.

prompt = """  
Translate the following English text to French.  
Input: "Hello, how are you?"  
Output: [French Translation]  
"""  

3. Include Context and Constraints

Provide relevant context and constraints to guide the model’s response. Incorporate specific instructions or limitations to ensure the generated output aligns with your requirements.

prompt = """  
Translate the following English text to French.  
Context: Formal conversation  
Input: "Hello, how are you?"  
Output: [French Translation]  
"""  

Context

It is common in prompt engineering to add sentences like: “Act like an expert” + “write a paragraph” + “target specific group of people” + “define your ideal format”.

For instance: “Act like an experienced marketer who works for a car dealer. Write an email with a car sale proposition. Email is targeted at customers with medium budgets who make decisions emotionally. Use a friendly and professional format.

Also, try incorporating a domain-specific terminology or jargon.

prompt = """  
Diagnose the following medical condition:  
Symptoms: [List of symptoms]  
"""  

Constraints

A prompt such as “Explain the Newton’s laws for first-grade students” will produce a dramatically different level of detail and length compared with “Explain the Newton’s laws for Ph.D.-level physicists.”

4. Pose Open-ended Questions or Requests

This is the whole point of generative AI: to create. Basic yes-or-no questions are restrictive and will probably produce brief, dull answers.

Asking open-ended questions allows for far greater flexibility in the answers. A straightforward question like “Was World War II the biggest and deadliest war in history” for instance, is likely to generate a similarly straightforward, brief response. A more open-ended question would be “Describe the social, economic, and political factors that led to the outbreak of the World War II”. This would yield a more comprahensive response.

5. Experiment with Temperature Settings

LLMs have a temperature parameter that controls the randomness of generated responses. Test different temperature settings to balance between creativity and adherence to the desired prompt.

prompt = """  
Translate the following English text to French:  
"Hello, how are you?"  
Temperature: 0.5  
"""  

6. Use System or User Prompts for Contextual Conversations

For contextual conversations, utilize system and user prompts to maintain conversation history and continuity. System prompts provide context, while user prompts capture user inputs and questions.

system_prompt = """  
Translate the following English text to French:  
"Hello, how are you?"  
"""  
user_prompt = """  
User: What is the French translation?  
"""  

Users can choose different models -> the one that performs completion task or chat task, however it is important to remember that there may be minimum and maximum character counts for prompts.

While a lot of AI interfaces don’t have strict limits, AI systems may find it challenging to process very long prompts or chat history. It is good to determine whether a particular AI has word count restrictions, and only create a prompt that is long enough to convey all necessary information.

7. Preprocess Input for Optimal Results

Preprocess input before feeding it to the LLM to improve the quality of generated responses. Tokenization, normalization, or filtering can help remove noise and enhance the prompt’s effectiveness.

import nltk  
   
def preprocess_input(text):  
    text = text.lower()  # Convert to lowercase  
    text = nltk.tokenize(text)  # Tokenization  
    # Other preprocessing steps...  
    return text  
   
# Preprocessing example  
input_text = "Hello, how are you?"  
preprocessed_text = preprocess_input(input_text)  
prompt = f"Translate the following English text to French: '{preprocessed_text}'"  

This technique may be used to: avoid conflicting terms, use punctuation to clarify complex prompts, paraphrase, setting output length goals or limits.

8. Evaluate and Refine Prompt Performance

Regularly evaluate the performance of your prompts. Analyze generated outputs, assess their quality, and refine the prompts based on user feedback or desired improvements.

prompt = """  
Translate the following English text to French:  
"Hello, how are you?"  
"""  
translation = LLMModel.generate(prompt)  
   
# Evaluate and refine prompt  
evaluation = evaluate_translation(translation)  
if evaluation != "High quality":  
    prompt = """  
    Translate the following English text to French:  
    "Hello, how are you?"  
    Please provide a more accurate translation.  
    """  
    translation = LLMModel.generate(prompt)  

9. Experiment with Prompt Length

Explore the impact of prompt length on the generated responses. Test different prompt lengths to find the optimal balance between providing sufficient information and avoiding overwhelming the model (or user).

# Short prompt  
prompt = """  
Translate to French:  
"Hello, how are you?"  
"""  
   
# Long prompt  
prompt = """  
Translate the following English text to French:  
"Hello, how are you?"  
Provide a detailed and accurate translation.  
"""  

10. Document and Share Prompts

Documentation is crucial for prompt engineering. Maintain a repository of effective prompts, along with their performance and any specific considerations. Share knowledge and best practices with your team to foster collaboration and efficiency.

By following these prompt engineering tips, you can maximize the effectiveness of your interactions with Large Language Models. Clear and specific prompts, along with context, constraints, and experimentation, will ensure accurate and relevant responses. Remember to preprocess input, evaluate prompt performance, and document your prompts for continuous improvement.

Happy LLM prompt engineering!

Note: The examples provided in this article are written in python language for illustrative purposes. Actual implementation may vary depending on the specific LLM framework, version or library used.