Exercise on prompting

Prompt Engineering Techniques: Improving Output Reliability

• Goal of the Exercise: To improve a prompt to reliably extract key topics from a scholarly text passage and format the output as a JSON array of strings.
• Initial Problem: The starter prompt was vague, leading the model to fail the required JSON output format (initial average score: 2. 8).
• Technique 1: Simplicity and Directness:
• Be extremely clear and direct about the desired output.
• Explicitly instruct the model to return a "JSON array of strings" containing all extracted topics.
• Result: This simple change significantly improved the score (up to 9. 5%).
• Technique 2: Structuring with XML Tags:
• Wrap the input content (the text passage) using XML tags (e. g. , <text>... </text>).
• Purpose: This creates a clearer structural connection between the prompt's instructions and the actual data being processed, helping the model distinguish input from instruction.
• Technique 3: Specificity via Step-by-Step Instructions:
• Instead of a single command, structure the prompt using a "Follow these steps" format.
• Example Steps:
• Closely examine the provided text.
• Identify each topic mentioned.
• Add each topic to a JSON array.
• Respond only with the final JSON array.
• Advanced Techniques (Optional):
• One-shot or multi-shot prompting can be added to further refine the output, though they were not necessary for the final successful prompt in this exercise.

Takeaways

• Be extremely clear and direct about the desired output format, explicitly instructing the model to return a JSON array of strings.
• Use XML tags (e. g. , <text>) to create a clear structural separation between the prompt's instructions and the input data.
• Structure complex tasks using "Follow these steps" to guide the model through a sequential process.
• Advanced refinement can be achieved using one-shot or multi-shot prompting.