Transcription
I spent the last three years running my AI automation agency, in which I've scaled to over 100K per month. And during that time, I've figured out exactly what works, and I'm going to give it to you, the entire playbook, within this 10 minutes of this video. No fluff, no theory, just the exact framework that I use every single day within client projects. So, without further ado, let's get straight into this video.
So, here's what we're going to be building today. This is the prompt structure that I use, and it's: role, task, context, format, examples, and constraints. And that's kind of how we format all of our prompts.
Now, the first one is role. That's step one. That's giving the AI prompt a role. This is where you tell the AI exactly who it is. Set the expertise level and change how it thinks about the problem. This is our prompt skeleton. So, we have role, task, context, and format. And for the role, this is what sometimes it can look like for a smaller prompt. "You are an expert email marketer specializing in B2B SaaS." Some other example roles are: "You are a customer service specialist." "You are a sales copywriter." "You are a data analyst." This is completely customizable to whatever AI agent you are building.
Now, if you just write, "You are a helpful AI assistant," then expect to just get these mediocre results that we spoke about in the intro. If you want to build these real AI agents that you are given to clients and you want them to perform at the highest level, then you have to spend the time to give it the exact correct role.
For step number two, we define the task that the AI is going to do. Now, with this, we need to be specific and not vague. One clear objective, tell it exactly what to do. So, for example, we have down here: "Your task is to write a follow-up email to a lead who downloaded our pricing guide but hasn't booked a demo." Now, a bad example that we could have used in that situation is: "Just write an email." Now, that's a terrible example. And to be honest, I'm guilty of doing this sometimes, and I'm sure if you're watching this video, you're probably also guilty of that. Now, a better example is to say: "Write a follow-up email to a lead who downloaded our pricing guide." Now, these things may seem like just small, tiny little improvements to the text, but in the sense of AI and AI agents, these are massive. These are leaps and bounds better in the results that you're going to get out of them.
Now, for step number three, we have add context. This is where we give the background info that the AI needs. This is like stuff about your business, your audience, the constraints that the AI agent may have. This is the why behind the task. But why exactly are you doing something? Now, an example of this is: "We're an AI automation agency targeting small businesses. Our tone is helpful and educational, not pushy. The lead downloaded the guide 3 days ago." So, you can see exactly what we've done there. We've given it some background info, which is: "We are an AI automation agency targeting small businesses." This is our background info up here. We've given it the business and the audience constraints. "Our tone is helpful, educational, and not pushy." So, there's your constraint and your audience. And then we've say the why behind the task is because the lead has downloaded the guide 3 days ago and hasn't booked a demo, which kind of ties in with the task up here. So, what to include here? Who you are, who the audience is, the tone and brand voice, and any relevant details to that task that the AI is going to partake for you.
Now, step number four is specify the format. So, what exactly do we want the AI to output? Do we want it to output a paragraph? Do we want it to output JSON? Do we want it to output bullet points? This is where you're going to do this. So, how you want the output, the length, structure, and style of the output. And the idea of this is to prevent rambling responses. We've all received these type of responses when we're working with ChatGPT, when it comes up with the actual answer, but before there is a mess of words and paragraphs that we have to sift through. And this is what we're trying to deter in this step here. So, you can see down here, we say: "Keep it under 150 words. Include a clear CTA. Use short paragraphs." Now, other examples are: "Give me three options." "Bullet points only." "One paragraph." "Casual tone." And obviously, this is an example, and this slide that I'm showing you is just some basic examples. And if we were to do this for a client, all of this prompt that I've showed you already is going to be a lot more detailed. But for example purposes, I just want to make this extremely simple for you.
Now, that's basically the core of building a good prompt. That's kind of all you really need to know. So, if you clicked off this video right now, you wouldn't really be missing too much. You would kind of have the general gist of how to build a decent prompt and most likely be better than before. But we can make this a good bit better in the following slides.
Now, down here, we have a power move, which is giving the AI example. So, as you can see here, this is example one, in which I give the AI, and this is to level up the output of the AI. Why this works is because it shows the AI exactly what good looks like. It's a massive quality jump, and it's especially good if you want a specific brand voice. So, you can see down here, we have our previous prompt, but we add a little example in here, saying: "Here's our brand voice." And we want the email to look like: "Hey, [Name], notice you grabbed our guide. Quick question. What's the biggest bottleneck in your current workflow?"
Now, power move number two is constraints. So, the first one was example. The second power move is giving the AI constraints. This is a real level up, and this is setting the guardrails for the AI. Now, why constraints matter is because we tell the AI what to do, we prevent common mistakes, and we keep the output on brand. So, some constraint examples can look like: "No jargon." "Keep it under 50 words." "Avoid being salesy," etc. And in a prompt, the example could look like something like this, where we say: "Don't use corporate jargon. Don't mention pricing in email. No emojis," etc.
So, when your AI agent has tools, like most of your AI agents in any den or make.com, wherever you're building them, this is what you'll do whenever we need to tell it the tools that it has access to. So, firstly, why does this work? It's because in a prompt, when AI can take actions like send emails, search the web, update a CRM, call APIs, you prompt the goal. The AI picks the tools, and this automates complex workflows. So, that's the idea of having a tool within your AI agent. The goal here is to delegate the task execution, allowing you to focus on the strategy. So, guiding your AI agent through this process is: define the desired outcome, not the specific tool commands. Establish guardrails to control the behavior and prevent errors.
So, you need a prompt like this, like we've been using previously. I might give it tools like this, where you have an email sender tool, a CRM updater, a web search, and a scheduling tool. And we'll basically give it a description of each tool. So, as you can see in here: "Email sender: Sends personalized emails. Used for direct communication with the leads." Then we have also down here: "Log interactions and updates lead status. Use after communication or task completion." We also have: "Gather current market data or competitor information. Use if specific, up-to-date external info is needed." And for the scheduling tool: "Book meetings or demos. Use when a lead expresses interest in a next step." So, we give it a very brief example on how to use these tools and the description of these tools.
So, you can see down here, when we give it some guardrails: "Do not use the send email tool until I approve the draft." "Only use the CRM updater if the lead status changes significantly." Or even down here, we say: "If a lead has booked a demo within the last 24 hours, do not send a follow-up email." So, we give it constraints on when to use these tools.
Now, these are kind of like some before and after. So, you can see here, a previous prompt may look like this, where it's just a block of text, and we say: "Hey, I need you to build me a lead follow-up system that automatically checks if somebody downloaded our ebook in the last week, and if they did, sends them a personalized follow-up email about booking a demo with our sales team, but make sure it updates the CRM when it successfully blah blah blah." You get the point. It bundles multiple vague requests into a single command, lacks clear guidance, forces the AI to guess the actual intent. It leads to inconsistent and undesirable outputs and does not effectively automate the desired workflow.
Now, on the other hand, we have the prompt that we just went through here, where we give it the task, context, the tools that it has, the instruction, and the guardrails to create a proper prompt that we could now use in production for one of our clients. You can see here, we clearly define the AI's role, specific task, and relevant context. We outline the tools available, provide precise instructions, include crucial guardrails, ensure predictable, high-quality output, and we leverage AI's full potential of the intelligent automation process.
Now, that's everything for this video. That's three years of my prompting experience combined into 10 minutes to hopefully make your life a little bit easier when you're building client projects or just prompting AI in general. If you like that, leave a like below. I'm going to I'm going to leave this presentation also below as the second link, and the first link is going to be the link to my AI automation community. If you want to learn how to scale your AI agency to over 100K per month like me, then click that first link in the description.