Creating Advanced Prompts

 Working with Large Language Models (LLMs) is like giving instructions to a very smart helper. To get the best results, you need to guide it by writing clear and useful prompts. This is called prompt engineering.

 

The Two Main Steps to Writing Prompts

1.   Create the Prompt
Start by writing a prompt that gives enough details for the LLM to understand what you want. The more specific and clear your prompt is, the better the response will be.

2.   Improve the Prompt
Once you see the response, think about how you can make the prompt better. Maybe it needs more details or a different way of asking the question. Keep improving it until the response is just right.

 

Moving Beyond the Basics

Basic prompts are simple, like asking a question. But to get the best answers, you need to go deeper. This means understanding why some prompts work better than others. By testing different prompts and comparing the results, you can learn how to write prompts that consistently give better answers.

 

Simple Tips for Writing Great Prompts

·       Be Clear: Write exactly what you want. Avoid confusing words.

·       Add Context: Give background information if needed so the model knows how to respond.

·       Experiment: Try different ways of asking the same thing to see what works best.

·       Assign a Role: Sometimes, it helps to tell the model what role it should take (e.g., "You are a teacher explaining...").

 

By following these simple steps and tips, you’ll become better at writing prompts and getting more accurate and helpful answers from LLMs.

 

Example Prompt Template to get SQL query (for BigQuery)

You are an advanced SQL generator that understands BigQuery syntax. Based on the given metadata and user requirements, generate a valid and executable SQL query for BigQuery.

      ### Metadata:
      %1$s

      ### Dataset Information:
      Dataset Name: %2$s

      ### Table Details:
      %3$s

      ### User Instructions:
      %4$s

      ### Constraints:
      1. Use only the provided dataset and table(s).
      2. Ensure column types are respected (e.g., do not compare integers with strings without casting).
      3. Respect primary key constraints and unique identifiers in joins or filters.
      4. Incorporate NULL handling if the user instruction implies potential NULL values.

      ### Output:
      Strictly analyze the provided user instructions against the table details. If the user instructions do not match or are unrelated to the table metadata, return:
      {"query": "", "title": "PROMPT DO NOT MATCH TO THE TABLE SELECTION"}

      If the instructions match, generate a valid BigQuery SQL query and return in the following JSON format:
      {"query": "constructed query, do not include any explanation", "title": "Title For the Insight, do not include any explanation"}

      Make sure the query is optimized for BigQuery.

 

Previous                                                    Next                                                    Home