5. Prompting Part 1: Introduction to System Prompts and Prompt Engineering

Understanding Prompts in LLMs

What is a Prompt?
- A prompt is the input text provided to an LLM to guide its response.
- Purpose: Prompts direct the model's output by setting context, defining roles, and outlining the style or tone of the response.
- Types of Prompts:
  - User Prompt: Direct input given by a user, usually a question or command (e.g., “Summarize this article”).
  - System Prompt: Pre-set instructions that shape the model’s behavior consistently across all interactions (e.g., “You are a helpful assistant”).
System Prompts: Defining Model Behavior
- Definition: System prompts are specific instructions embedded in the LLM to set a consistent tone, behavior, or focus throughout an interaction.
- Example Use Cases:
  - Professional Assistant: “Respond formally and professionally, addressing users as 'Mr./Ms.' or 'Sir/Madam' when appropriate.”
  - Friendly Assistant: “Respond in a friendly and conversational manner, using informal language.”
  - Industry-Specific Assistant: “Provide answers based on healthcare best practices, and respond with medical terminology when relevant.”
- Benefits of System Prompts: Ensure the LLM’s responses align with a desired role or behavior across sessions, providing a cohesive user experience.

What is Prompt Engineering?
- Prompt engineering is the practice of designing effective prompts to achieve desired outputs from an LLM.
- Objective: Tailor prompts to increase the model’s response relevance, clarity, and tone based on user needs.
Key Principles of Prompt Engineering
- Clarity and Specificity: Ensure prompts are clear, concise, and specific to guide the model effectively.
  - Example: Instead of “Explain this concept,” use “Explain the concept of gravitational force as if explaining to a high school student.”
- Context Provision: Include necessary background information so the LLM can generate informed responses.
  - Example: “You are an AI language model tasked with helping users fill out government forms.”
- Iterative Refinement: Adjust prompts based on output quality; small changes can significantly improve relevance and tone.
Prompt Structure and Components
- Task Instruction: Directs the LLM on what to do (e.g., “Summarize,” “Explain,” “Translate”).
- Context Setting: Provides the background or domain knowledge needed for accurate responses.
- Tone Specification: Instructs the LLM on formality, friendliness, or professionalism.
- Example for Structure: “Summarize this medical research article in a friendly and accessible tone, suitable for a layperson.”

Examples of Prompt Adjustments and Their Effects
- Simple Prompt: “Summarize this text.”
- Enhanced Prompt with Context and Tone: “Summarize this research article on climate change in a clear, layperson-friendly tone, highlighting key takeaways.”
- Example with Iterative Refinement: If the output is too complex, refine to “Summarize this research article in 3 main points, in language suitable for a general audience.”
Role-Playing Prompts
- Use Case: Creating a virtual agent that acts as a customer support assistant.
- Prompt Example: “You are a customer support assistant for a booking platform. Answer questions on cancellations, refunds, and booking changes in a polite, helpful tone.”
- Iterative Refinement: Adjusting tone based on output, e.g., “Respond in a concise and empathetic manner, avoiding technical jargon.”

System prompts define the LLM’s foundational behavior, ensuring consistent tone and focus.
Prompt engineering techniques allow for fine-tuning responses to suit specific needs or roles.
Iterative refinement helps improve output quality, providing flexibility in real-world applications of LLMs.