Providing examples

Prompt Engineering: Example-Based Prompting

• Core Concept: Providing examples within the prompt is a highly effective technique for guiding the model's output.
• Terminology:
• One-shot prompting: Providing a single example (Input → Ideal Output).
• Multi-shot prompting: Providing multiple examples to improve consistency and handle complex scenarios.
• Implementation Structure:
• Clearly define the task by providing a sample input and the desired ideal output.
• Use XML tags to wrap inputs and outputs (e. g. , <input>... </input>, <output>... </output>) to structure the prompt and clarify the model's role.
• Key Use Cases:
• Handling Corner Cases: Examples help the model recognize nuances, such as sarcasm, which might otherwise be misinterpreted.
• Complex Output Formatting: When generating structured data (e. g. , complex JSON), providing a sample input/output pair ensures the model adheres to the exact required format.
• Reinforcing Quality (Advanced): To further guide the model, include the reasoning or justification for why the output is ideal (e. g. , copying the grader's explanation from a prompt evaluation report).
• Application: This technique is particularly valuable when performing prompt evaluations (prompt evals) to guide the model toward high-scoring responses.

Takeaways

• Example-based prompting guides model output by providing input → ideal output pairs.
• One-shot prompting uses a single example, while multi-shot prompting uses multiple examples for consistency.
• Use XML tags to structure the prompt and clearly define the model's role by wrapping inputs and outputs.
• This technique is essential for handling complex output formatting (e. g. , JSON) and recognizing nuances like sarcasm.