Transcription
Most NA10 operators focus only on telling AI the actual task, but they totally skip the two layers of prompting that actually prevent your AI agents breaking in production.
The truth is everybody has access to these tools, chat, GPT, Claude. Your competitors have them, your clients have them. So what actually separates you from 99% of the other operators? It's not the tools that you use, it's the three layer prompting system that makes those tools reliable at scale.
[music] I've generated over $25 million in revenue through automations for my clients. And in this video, I'm breaking down my complete prompting system, the one that cuts costs, prevents mistakes in production, and generates [music] clients results that they can't get anywhere else. That being said, let's jump in.
All right, guys, let's get into it. So, like I mentioned in the intro, there are three layers that need to be included in your AI agents. So the first layer is the system prompt. The second is the user prompt. And third is how to define the tool usage. Without all three in your agent, it's just not going to get you the outputs that you want.
If you're not using any tools, then you don't need to give tool instructions. But in this video, I'm going to show you an agent that is using some tools. And that's really where the AI agents can trip up when it comes to your tool usage. So that's where your prompting and your workflow architecture plays a huge role.
So think of it like this. If you're hiring an employee, they need a role. They need a list of tasks to complete and they need training on the different types of tools that they'll be using in their job.
Now, before we get into some examples and some templates that you can use, I want to give you a bit of context of why this is important. I strongly believe that prompting is probably the most valuable skill right now because whenever you interface with any kind of AI, the quality of the outputs are determined by the quality of your inputs. So that applies to your AI agents. It applies to using GPTs. It applies to using Claude and any other LLM. Whether that's for for content, whether that's for videos or photos. It's all in the prompts.
So here in my opinion, the most important prompt design fundamentals. Firstly, being concise because extra words means more tokens, which means a higher cost and that's going to risk truncation. So if you have a huge prompt and you're trying to generate a massive output, then certain LLMs might not have the context window. So for example, if you're pulling in a YouTube transcript, then you might want to use a Gemini LLM because it has a higher context window. The second thing is to prioritize certain information, and I'm going to show you how to do that with Markdown. Third, avoid redundancy. You don't want to repeat the same thing in your prompt, even if you're saying it in a different way. You got to be concise. And finally, strategic context. And this includes examples of a successful input and output. And an example of an error or an edge case that the agent should be able to handle.
So I mentioned markdown. Why is markdown important and why should we use it? You've probably seen this in prompts that you found online. It's the the hashtags and the asterisk. And when we say markdown, it's essentially how Word documents and Google Docs are formatted. We have headings, we have bullets, and if you actually look at the code behind that, you'll see that it's uh hashtags and asterises that are responsible for that formatting. So, when we add markdown to our prompts, we're basically just adding signposts in the embedding space. And the reason why we have the roll at the top of the prompt is because order matters. The context window scrolls from left to right. So, the earlier tokens in your prompt get more weight. So this is where we put the role, the goals, [snorts] constraints and output format as early as possible.
So with that being said, let's go in and take a look at some examples. So here I have a very simple straightforward template that you can use. We have a role. So you are a expert direct response marketing copywriter that helps write worldclass video sales letters. That could be an example. We have some rules of what you should do and what you shouldn't. And in the rules section, we have rules of things that we should do and the things that the agent shouldn't do. It's very important that you create that distinction. We give instructions for the tools that are available. We specify the output format, whether we want JSON or plain text, for example. We give examples, so a sample input and the output that we should expect with normal behavior. We can also give an example of an error case that's going to help us improve the prompt in our agent. So that is a simple template. You can get started with this.
I can give you something a bit more advanced if you want. This again is just a template. We're going to get into a specific example in a second. But you can see here there's a little bit more additional information. We've got background context. We've got validation requirements. And specifically with the tools, we've got the purpose, when to use, and the required parameters. This is important for getting the best out of your tool usage. We've got input format, output format, error response structure, decision logic, examples, quality assurance, escalation protocol. So, some of this stuff might not be relevant for the use case that you're trying to build for.
But I'll show you an example of a chatbot that I've got that basically books people in for calls if they come through to my website. The chat bot appears, they chat to it, maybe they ask some questions, and the prompt has been designed to move people towards booking a call. You can see here we've got access to a couple of different tools. Now, I'm not using a sub agent to call these tools. I don't think it's necessary for this particular agent, but if I had more tools that the agent had access to, then that's what I would definitely do. If you're adding more than a couple of tools to your agent, you're probably going to want to separate those out into sub aents and have the AI agent as the CEO and the orchestrator and then the sub agents as individual workers.
So, we're still on layer number one, which is the system prompt, and you can find that just down here in the option section. Now, if you don't see the system message, you can go down to add option and add that in there. And I'm just going to open this up. Now, if you do have some dynamic expressions in here, in this center panel, you're going to see the expression and on the right hand side, you're going to see the completed expression using the dynamic data.
And so, I'm not going to take you through every section of this, but you can see we've got the key elements that I mentioned in the template. We've got the role, the mission, because it's a chatbot, we're adding the tonality, which I think is important. We're giving it access to specific knowledge. So, you can tell the agent to tap into, for example, I'm using Spin Selling, which is a very popular book. It's going to have access to that knowledge in its LLM database. Let's say it's going to have access to this knowledge. So, if you want to pull in information from different techniques and books, you can do that.
In this section, we've got strict scope limitations, forbidden topics. So, if somebody comes to the chatbot and asks, hey, can you write me an essay on climate change? It's going to say no. It's not going to write the essay and take up a ton of tokens. So, this is important if you want to add some guard rails to your system. We've got examples. Booking call protocol, very important. Gathering essential information. Step three, determine the meeting type from the context. Check availability responses based on the availability. And then required fields for the calendar event. So, we're asking for specifics that we can then input into the calendar event that we schedule in the calendar. We've added some mandatory parameters for the email invites, time zone format, if what happens if the booking fails, pre-booking, verification, confirmation, error handling, troubleshooting, calendar issues, and then at the bottom down here, we've got FAQs, information about my company, the offer, the pricing, all that stuff. So, we don't really need to tap into any kind of knowledge base. Now, we do have access to a Pine Cone store there, but for the most part, it's not actually tapping into that.
So, if we open the chat and so if we open the chat, I've already had a conversation here to show you how this works. I've asked how much is the mentorship. It's given us an answer and it's asked, you know, can we get you booked in for a call? So, I've said, sure, how about 11:00 a.m. tomorrow? It's asked me about my time zone. I give my email address and it's all sorted. It's been booked in. So, if we go over to my Google calendar, you can see the meeting has been added correctly with the Zoom link and all the information that that you'd want, including the email address and the invite for the attendee. And you can see in my workflow in the executions, you can see it's it's used the check availability tool. And you can see in the last execution that I ran for this agent, you can see that it's checked the availability and it's created a calendar event. So, it's actually used those tools.
Now, if you don't see these tools light up in green, then the agent is not using them. So, how you would troubleshoot that is you can come into the logs and you could see what's actually happening with the tool usage and what the agent is actually doing. So, if you do have an issue maybe in the create calendar event tool, then you can see what the input is and if something is missing here, then it might not give you the output that you want. It might not have enough information to create the calendar event. And that's certainly the issue that I had when I was troubleshooting this. This is absolutely not the first draft. It took quite a while to get this up and running and booking as I wanted it to. So, this is just one of the ways that you can actually troubleshoot and fix things.
Now, one last thing I will mention here. We do have markdown. We've got these dashes for bullet points. We've got the double hashtag for subheadings. Got the main headline up here. You can put in emojis if you want. You can see we've got three important sections. We're using the astericses to bold that section. So that's layer one in a nutshell. It's the prompts that you're going to do the most iteration with to get your agent working as you want it to work.
Now, there is a tool out there that you can use that I think is pretty cool to get you a first draft of an agent prompt. I am going to give you these templates below in the video. So there's going to be a link below. You can grab these. You can also just speak to chat GBT and voice note in. One of the tools that you can use is prompt cowboy. And so what I did is I voice noted in this initial prompt. So can you write me a system prompt designed to pre-qualify leads for my mentorship program. So it ended up giving us this pretty solid system prompt. So ro task context and instructions. Now if you want to improve this, you can just it's got some questions here that you can answer to help improve the prompt. So this is a pretty cool tool that you can try out if you want to try something that's a little bit more structured. And you can just come in for new prompts and there's different types of prompts that you can create for agents, deep research, etc. And you just come down to the agent prompt and see see what it gives you. You can try different methods.
So, moving on to layer number two. This is a little bit more straightforward. This is the user message. So, with some chat bots, you might have this agent connected to the chat trigger. And in that case, you're just going to select it'll be selected by default. It's connected chat trigger node. and that's just going to pass the message that comes through from the chatbot. Now, if you're using an agent for different purposes and you want to add dynamic content in there, maybe you're pulling some information from a database, then you can add that into the user prompt. Typically, what you do in the user message is that's where you add all your variable information that's then going to give the context for your system prompt to work. So, in this case, I am just passing the chat message and adding in the current date and time, which is important. If you're trying to book a call, it's important to give the LLM the specific date because sometimes it can think that it's 2023, which is definitely not the case. So, not much else to say on that.
Then, finally, layer number three, which I touched on earlier, was instructions for using the tools that your agent has access to. This can be a tricky part of the iteration process when it comes to your prompts. tools can be notoriously tricky to configure because your agent is then choosing what to do with the information that it's got and which tools to call and then what information to pass to those tools. So what I've done here is I've taken the prompt for my agent and I've [clears throat] asked Claude to highlight the key elements where it's instructing the agent to use those tools. So this is what it's got. So we have a calendar booking protocol and then we have a check availability section. So we're giving an order for the agent to use the tools, right? All right, we're giving it a step-by-step process. Very important. So, we got the tool execution and then create the booking mandatory fields required. So, the chatbot is then going to ask for these fields that it needs to complete the booking. We're then sending mandatory parameters for the email invites. Very important. This is something that I figured out during testing that just wasn't working for some reason. We've got the pre-booking verification. So, do we have all the information that we need like email address, the name, the context, etc. We've got some error handling and troubleshooting calendar issues.
Now, is the troubleshooting important including the prompt? You can test it. This is just what's working for me right now. But there is one way that we can test whether the prompt is solid or not, and that is by putting it in your AI agent and seeing if it gets the outputs that you want. One of the other things that you can do is jump into Claude. Right? So, what I've done here is I've put in the prompt into Claude and I've asked it to rate on a scale of 1 to 10 of how clear it is, how concise it is, and how likely it is to complete the actions with the tools that I've specified. So, we can see here there is some room for improvement, right? Nothing's perfect, but clarity, eight out of 10. I'm pretty happy with that. Conciseness, okay, could be more concise. We can definitely make some improvements on that. The tool execution, which is really the goal of this prompt, is for it to use that booking tool correctly. Eight out of 10. Pretty happy with that. And you saw earlier that this workflow does actually call the tools and takes the actions that are required.
So, one final thing I will mention about tools before we wrap things up for today is how we define the fields. So, one thing that you can test is a tool description. If it's not working how you want, try set manually and give a more detailed tool description. for the create calendar tool that we've got here. Letting the model define the start time and end time because in the prompt I've told it how long the call should be. The attendees again we're detecting with the AI description detect by the model. The location is hardcoded and the summary is also added dynamically by the model. And so it's important that we mention in the system prompt how to define these particular fields that we pass through to this tool. Last thing I will mention is when you're mentioning the tool in your prompt is making sure to reference the exact name of the tool. So if it's got underscores between the characters, use those underscores.
Now there are some more advanced techniques that we could use here. We could use chain of thought, we could use critique prompting, we could use self-reflection, and we can use meta prompting. But this is all stuff that I go into in more depth in my mentorship.
But this is all stuff that I go into in more depth in my mentorship.
Now, there's one final little golden nugget that I want to share with you here because you might be thinking, well, if I'm running a huge prompt, how can I track the token usage? So, what we can do here is we can swap out this open AI node. We can pop in open router. We can select the model that we want. You've got access to all of the models. So, I'm just leaving this 4.1 mini. And in open router, we can create an API key for anything. So, we could create one for each specific workflow. We could create one for a client for all of their workflows and we're able to track the usage and the cost. Right? So if I come into here, I can see the activity for this particular API key. You can see all of the different runs that the agent has done and you can see the cost of that. So when you test and run your agent and you want to see how many tokens that agent is using and how much that costs, this is a great way to track that.
And if you really want to push your agents to the limit to see if it breaks, when it breaks, you can test it with edge case. You can input erroneous data intentionally and see what it does. Like if I asked my agent here to write me an essay on global warming, I'd want to make sure that it's not actually going to do that, right? It's going to take up too many too many precious tokens.
So there you have it. That's my complete prompt engineering system for AI agents split into those three layers. Now, most operators, they understand these concepts. They've seen them before, but they tend to freeze up when it comes to actually building [music] and testing these AI agents. They're not really sure which layer to start with, whether to do single or multi-agent workflows. And when they do set things up, the prompts just [music] break. The AI agent doesn't give them the outputs that they want.
So, if you want me to work through this with you, review your prompts, help improve the architecture of your workflows, and take you from zero to a top 1% AI operator, there is a link below in the description to join my mentorship. And if the [music] intake is currently open, then you should be able to apply and book a call to speak with me. And if you're wondering how to position yourself as a premium AI operator and land clients that actually pay big money, then [music] watch this next video on the screen right now.