Exercise on prompting

Prompt Engineering Concepts: Improving Output Reliability

• Goal of Prompt Engineering: To improve the performance and reliability of a Large Language Model (LLM) by refining the input prompt.
• Evaluation Process:
• Run the prompt and evaluation to generate an output report (e. g. , output. html).
• Analyze the report to identify common failure points in the model's output.
• Techniques for Prompt Improvement:

• Simple and Direct Prompting:
• Be extremely clear and direct about the exact desired output format.
• Example: If you need a JSON array of strings, state this explicitly in the first line of the prompt.
• Structuring with XML Tags:
• Wrap input data (like a passage of text) in specific XML tags (e. g. , <text>).
• This creates a clearer connection between the instructions and the data being processed.
• Step-by-Step Instructions:
• Define a sequence of actions for the model to follow (e. g. , "First, examine the text. Second, identify topics. Third, format the results as a JSON array. ").
• This increases specificity and guides the model through the required reasoning process.
• Advanced Techniques (Optional):
• One-shot or multi-shot prompting can be added to further improve performance.

• Core Task Example:
• Input: A single paragraph of text (content property).
• Desired Output: A JSON array of strings containing all key topics mentioned in the text.

Takeaways

• Prompt engineering aims to improve LLM performance and reliability by refining the input prompt.
• Be extremely clear and direct about the exact desired output format (e. g. , JSON array).
• Structure inputs using specific XML tags to clearly separate instructions from the data being processed.
• Use step-by-step instructions to define a sequence of actions and guide the model's reasoning process.
• Advanced performance can be improved using one-shot or multi-shot prompting.