Command AI's Identity: System, Contextual & Role Prompts (Session 5)

Good afternoon. My name is Dan Gayende and for the last decade I've been the trusted adviser for executives from startups to SMPPS to enterprise scale organizations. I found patterns in the architecture of those that lead their field. I know because I've designed and developed their systems and processes using AI platforms, models, and frameworks to positively affect billions of users. Now I bring my insights to you. Welcome to session five. Last time we learned how to show the model what we want using oneshot and few shot examples. Today we're going to explore another powerful way to guide the model by telling it what to be. We're going to give the model a job, a purpose, and even the personality. By the end of this brief session, you'll be able to first define the and differentiate between system, contextual, and role prompting. Second, provide examples of how each technique guides an LLM's response. To do that, we'll cover three distinct but closely related techniques: system prompting, contextual prompting, and role prompting. It's important to note that while these can be used together, they each serve a slightly different primary purpose. If you're following along on the white paper, the this subject is between pages 17 to 24. Let's start with the big picture. System prompting. A system prompt sets the overall purpose for the language model. It defines the fundamental task that it should be performing. Think of this as giving the model its core job description. This could be translate this text or very powerfully return your answer in a very specific JSON format. For example, on page 20 of the white paper, uh we show a prompt that asks the model to classify the movie review, but it adds a crucial system instruction to return the output in a specific JSON schema. The system command forces the model to provide a structured predictable response, which is incredibly useful in applications. Next, we have role prompting. This technique assigns a specific character or identity for the language model to adopt. This frames the model's output style and voice, adding a layer of personality. An example of this would be where a prompt begins with the simple instruction, I want you to act as a travel guide. Now, this immediately frames the entire response and the model knows now to provide suggestions from a perspective of a knowledgeable guide. You can even combine this with a specific style like asking the output to be humorous. Finally, there's contextual prompting. This technique provides task relevant background. A contextual prompt gives the model immediate task specific details to help it understand nuances of what's being asked. For example, a prompt could start with a clear instruction of context colon. You are writing for a blog about a retros8s arcade video games. When the prompt later asks for three topics to write an article about, the context ensures the suggestions are highly relevant and not just generic blog ideas. It focuses the model's creativity. [Music] The real power comes when you layer these three techniques. Let's create a prompt that uses all three. Imagine you want a customer service response. For the role prompt, you could say you are an empathetic and authentic customer service representative. For the system prompt, you define the task as your job is to write responses to tickets letting the customer know a person will respond in 48 hours and acknowledge the problem or issue they've raised. And for a contextual prompt, you provide the specifics. The tickets are for a new productivity app or something like that. By combining them, you've set the tone, the format, and the specific subject matter. And this result is going to be far more targeted and interesting than the generic request. [Music] Let's do a quick check. Imagine you want an AI to draft an email for your team about an upcoming project deadline. You want the tone to be professional but encouraging and you need to mention the project name is project Phoenix. How would you apply these three techniques? You could use a role prompt of act as a supportive and professional project manager, a system prompt, draft an email to my team, and a contextual prompt. The email is about the upcoming deadline for project Phoenix this Friday. The main takeaway is to think of these three techniques as layers of instruction. System is the what to do. Context is the what it's about. And the role is who you are. So far, all the techniques we've discussed involve giving the AI a direct problem to solve. But what if the problem is too complex for a single step? In our next session, we'll learn about a powerful technique to help the model reason through complexity called stepback prompting. Thank you and I'll see you in that.