From Idea to Cold Emails: Full AI Workflow Tutorial (n8n)

What's going on, guys? Michael Elliott here. Today I'm going to be running you through a simple but extremely effective flow that's going to assist you in your building techniques along with some core concepts of how to map ideas out on paper. So very educational video with full deployment of the flow at the end of it. So let's just jump in. This is what the flow looks like. This is ultimately using multiple different nodes, a logic logic node, two logic nodes, a sub workflow node, this as, as a multiple LLMs with system prompts in it. We have quite a lot going on here for the level of the. That you're probably at. If you're, if you're a beginner, this is probably like an intermediate flow. So there's going to be a couple of things from this that I take. I will have the flow attached below in a gumroad that you can download. Feel free to download it, pause it, download it now import it and then build alongside. I mean that's exactly what I'm going to do in this video. That's totally up to you. But yeah, I'm going to hit command S just to save it. What we're going to come out with is we're going to come out with research analysis of individual people depending on what we ask for. We're going to have their company description, we're going to have their tech stack and then we're going to have an icebreaker that uses that information to be able to. To do this. Now we're using. Well, I am using Sonar Pro and I'm using Open Router in order to. To use that model. It is a paid model, so if you haven't got a paid model, just stick with the normal model that you've got and see what outputs you get. You'll find that the quality is less than Sonar Pro for research requirements. Nonetheless, let's start the build. Okay, so I've got a pinned message here that says please give me 10 SaaS, businesses out of the USA that are most active on social with less than 100 employees going to copy this. And I'm just going to put this somewhere, just going to put this somewhere safe for me. So I, so I've got that question. Now we can unpin it. So I want you to hit tab. Let's get rid of that so it looks fresh. I want you to hit tab and I want you to hit to add a chat trigger, new chat trigger. Then we're going to add an LLM. So search for basic LLM chain and Here is where some sexy stuff happens. But first we need to escape out and add a model and I'm going to go through open Router's AI model and I'm going to search for Sonar Pro and there we are, we have this Sonar Pro LLM. If you can't use Sonar Pro and you using the, the, the, the LLM we connected last time. Gemini, whatever, Open AI it, you'll be fine. But it's just that that Sonar Pro is extremely good at deep, deep research and analysis. It's extremely strong. Now when it comes down to the structure, what I didn't want to waste time doing was going backwards and forwards with a AI agent whilst we were on this call to get the system prompt and the enrichment type. So what I'm going to do is I'm going to copy these over and then I'm going to head to AI studio. So inside of, let's do it in order inside of here. What, what do we want? Well, we want to give really clear and concise instruction to the machine. You are an elite tier senior business intelligence intelligence analyst and data enrichment specialist. You're a, your call sign is Pathfinder. Your entire purpose is to conduct exhaustive multi source online research on a given company and compile a detailed intelligence dossier. You operate with extreme precision prioritizing official sources like the company website, official LinkedIn pages, Crunchbase and reputable tech publications, e.g. techCrunch, built in etc. You are methodical, relentless in your search for accurate data and an expert in structuring your findings into a perfect JSON format. If you were to look at an introduction for a system, a good system prompt, that is what assist a system prompt would, would, would look like. Okay, that's exactly what the, the system prompt would would, would be like. Now when you are in a position of generating it from scratch like, like you have, you're going to want to. What I do and what I advise you do is I always take a good prompt that's worked for me before. That's logical. This is, this is good core directives and mythology probably needs cleaning up a little bit with these but yeah like you know, actually identify them. I know already that this is a, this is a good prompt so I'm gonna use this in a moment and we're going to get it to, to rewrite it. Technology stack analysis. You see we've given it some specific requirements so we need the, the tech stack contact vector acquisition for each identified leader, find their professional email use standard corporate email patterns so it's got really keen and strong advice, which is what you need when you, when you're looking at getting the information that we want to. And it's best practice when you're doing even smaller builds strictly here to this JSON structure. This is to try and get the strict. The JSON to be output properly. Anyway, we're gonna, we're going to have a fallback for that regardless. But I'm going to copy this and I'm just going to show you the way that you would go about this is you would head into AI Studio. I'm using AI Studio. So turn the temperature down, go into your LLM. Just tell it that you want a system prompt. But this actually has system instructions. You're a system prompt generator for NAN Basic LLM LLM nodes, using Sonar Pro for research. And then you create prompts similar to the below. And I've pasted in the one that I've got, which you will have if you've downloaded that. And now I'm going to say, can you please create me a prompt that researchers. Researchers Research that. Reese that. Please create me a prompt that researchers in depth. The company, the people of power to abstract business information and contact info will be using the research to produce 1. Produce icebreakers for cold email and hit run. So if you didn't have this base that I've got, you would come in here and you would do this. The problem is you'd probably have to go back and forward two or three times to get something workable, which is why I'm sticking with, with what we've got. So this is, this is, this is, this is workable. This is what, what works in terms of the, the flow. So I'm going to add a prompt, system prompt, and I'm going to paste in the message that we've. That we've got. Okay. So, yeah, I mean, let me just paste a merch style. There you go. It's bugging me now. Should have done it before the video frustrating me. Yeah. So, but it works. And that's. And that's what you want. So we have the system prompt and the input is obviously where, when the message comes through. So we, we have that. We've got our first prompt in. We've got the, the flow ready. And now I want to take us through a little bit of an exercise when it comes down to how you actually deploy ideas. Because this is a really key and core technique. I might do a video on it on its own. In fact, I almost, almost definitely will. When it comes down to filling in the dots here. This is a bit of a cheat because I've already built it, obviously, to make sure that it was functioning. But when it comes, let me turn this output parser off because we're going to use a node. But like, just simply put, one of the biggest things that happens to beginners is they get a brain block. So you know about building it, you know about, let's say you know about how to use. If nodes, switch nodes, you know about this. Handling data is difficult. You're still getting your head around Jason a little bit, but you have these ideas in your head, but you can't, you can't bring them to life. You look at templates, nobody builds exactly the same thing as what you need, so you can't find it. Well, the way that you overcome that, right, is by doing this. It's by sitting in a canvas like that and you give yourself a strict five minutes. You literally, a strict five minutes. Set a timer if you need to, a strict five minutes to build out, like, what happens. So I know that from my experience that this, this needs passing, this Jason needs passing, but you wouldn't maybe know that. So, you know, you want to do some research with a chat, with a, with an LLM, you want to start it with a chat trigger, you add it and then you connect its brain. So then what do you want to do? Okay, so what I actually want to do then is I want to make, you know, I don't know, I want to have a, a router and an if statement. If, if email exists. If email exists. Go this way. Okay, so then I want it to send a message to Slack. I wanted to create a channel in Slack and then I want it to send a message inside of Slack, letting my team know, send. Where is it? Send a message. And then maybe we want to send an email off. Send. Send a message. Obviously, we've got that now in the Slack channel, but now if the email doesn't exist, we want to send an email off and then we want to update a row in Google Sheets with the contact number and I don't know, maybe you want Twilio to make a call because the email doesn't exist. Right. All I'm saying to you is, and what I'm teaching you here is the key and core concepts of building and deploying a flow. When you get that brain block, when you, when you're building NAN flows and the brain block comes over, it's a real thing. No, I'm not John Cena. But it is a real thing, right? It leaves you in a position where you, you're stuck and you're like, how do I proceed? Well, I'll tell you how to proceed. Just dump everything that's in your brain on a canvas so that it looks something like that with no credentials added. It might be something for a business that you're working with, it might be for yourself. It could be anything. But map all of that down on a bit of paper, just start throwing nodes in and out and you'll very, very quickly come to a stage where you realize that actually, you know, I don't need this, I need that. I want it to do this. This. The if statement shouldn't be based on that, it should be based on this. But the point is, is that you will slowly become more and more efficient at building flows by using this, this technique. And then you can do whatever you, whatever you want from, from here. You know, like data can be merged and you can, I don't like it when they cross over, you know, and then you can like, I don't know, send a panda doc not part of this. I don't know, you have to make a thing. But yeah, then use, I don't know, an AI assistant to do something. The point I'm making is just like you build you, you image. Like image is a big thing for the brain. It will unlock a lot of things that you already know and remember when you actually start building. And I think that is really critical now today we aren't going to be using any of those things, so let's delete them. And what we're going to be doing is we don't want. The reason I haven't put the output parser on is because in every single flow that I use in production, I have a structured output parser that's pre built. It's a template by initm by a guy called Lucas Perrin who's, who's a great bloke, I know him well and I don't know why I've just gone to YouTube NAN products templates and then if we just, we find the Lucas no JSON. So this here process AI output to structure JSON with robust parser used for free import to the cloud workspace. I didn't need to do that because I've already got it. But yeah, so you import there, you command S. This explains to you what it does. So it explains that the workflow expects an input item to be text. It explains the core logic behind what this is in terms of the JavaScript. Don't worry about any of this. AI wrote all of this. Don't worry. That is absolutely fine. You don't have to worry. But basically all you need to know is that this just ensures that the, the data is passed 100 accurately and you have your values correctly correctly put out. You know, any strings that you need passing will get passed if you use this and I strongly urge you to save it and use it in the majority of flows that you build moving forward. It's a big, it's a big thing. So let's close these down because I don't need it because I've already got it inside of here. We'll have a look at what that came back with later. But then what we're going to do is we're going to execute a sub workflow and this is something that you might never have done previously. You'll notice that that other workflow that I've just added and we have. You've just added to your, your sort of quiver right of workflows now is available when we trying to execute a workflow. So what do we want? Safe JSON parser. We're just mapping for now. So we don't put the text, we don't have to put the text through at this stage and we look at moving forward from here in terms of what would be. So let's just try and map it out in our, in our brains. Hunter. We use Hunter IO to verify the deliverability of an email address. Yeah, we do that. We then update Google sheets and we append or update a row in sheets. We then use another basic LLM chain. So we add all information to sheet Generate Icebreaker model. We'll let it use the same one Sonar Pro again Choose whatever you want here. Command S I'm going to rename as you should always do with every node. Safe JSON parser Generate Icebreaker. Then we've got Update Google Sheets. I always do that. Google sheets. Append or update a row. Append or update a row. Okay. Right. So this looks, this is what we've got now this looks interesting but there are a few things that we are missing. So one thing that comes to mind straight away and, and as you look at this now try and think about what, what is, is an issue here. And it all comes down to logic, which is the area that most people get stuck on when it comes down to the way that they move data. Well if you look at this now, if we go into email validation, it's asking for an email but we have, we have emails coming through that we've asked inside of the system prompt that I've built. At least I've asked for three emails at once which is a lot and is not going to perpetually do that. We don't want to put a loop in we. Because it's not what where that's required. We just want to add a split out node and this comes from experience, not from. Not that you would know this right now but if I'm thinking in my head maybe that would work there and from a perspective of after the validation we're going to certainly want to truncate whatever happens inside of it. Okay, so now auto align. Shouldn't have done that. That was a bad idea. I just sort of gave the game away. You see I've just by revealed a little secret here. But this isn't an if, this isn't an and if an if statement at all. It's just a filter. That's all it is. But this is what I mean about how you learn with your brain and you, and you look at them and you go through and you do that. Add information to sheet filter. Split out, split out emails question mark. We'll find out when we give it some data. Now what I'm going to do is I need to check the, I need to check the LLM chain, the original one. So this text message from the user message which gives it as an, as an example I'm going to paste this in again. This will be, this is a user prompt but this will be part of the download that you can, that you can grab me. Just make sure I copied and pasted all that. Yeah. Okay, you see so this is the example that I gave it from when I was speaking to the, the AI and it gave me this output because we want three individual people from said company at the same time. And we can do some real cute stuff in the icebreaker by asking it to cross reference another member of staff or owner or whatever we want to do. Yeah, so I think we're, I think we're good. Yeah, so that's an issue. Message type text. Yeah, that was DAF to me. Load this out because it looks horrible. Paste and match style. Happy days. That's a good example. We're now cooking so we can do this but we want to go through it one step at a time and we need, and the reason we need to go through it one step at a time is because we absolutely definitely need a Google sheet that that's created. We know that what Else do we know? We've got to get a system prompt for the generating icebreaker. Okay we know that and we need to set up the credentials for Hunter IO. We know that and we need to find out what the filter is. We know that but what we don't know is what the output of the current LLM that we've built and given a system prompt to is going to look like. So let's open chat, let's open chat here and I'm pasting in that message. Please give me 10 SAs. Should we change that? 1010 renovation businesses out of the USA. It's very broad that has less than 50 employees. I want you to find the business decisions makers for me and ensure people found are likely to have a valid, a valid email address. So we fire that off sonar programs and does it does its thing whilst we do do that. You see this response here that I got from it. It gave us a pretty similar, a pretty similar prompt because we gave it. We prompted it to give us a similar prompt. So that's how you keep the quality and that is something that you, that you should make sure that you keep. This is actually trying to bring in more. More values which is good. Like more. More is fine. Let's see what the, let's see what the. The LLM has spat out. So it looks like we've had success now as expected, the, the structure because, because I haven't given it a structured output, it's going to look like that. So that's fine. So now what do we do guys? We pin data. We pin. We pin. We connect. We connect. We execute this step which is the safe JSON parser. Now look at the output, look at all that lovely stuff that we can go after. So it's now in a nice digestible format and importantly, importantly we have the JSON correct. Now obviously we drag this text over because this is the input for the parsing node to connect that to make that work. But this was 100% correct. But I don't think it's the emails we need to split out. I put split out emails looking at this. It's the companies that we need to split out. No other field. Let's just disconnect this and see what happens if I execute that L to hide the logs so we can see a little bit more. Let's see, split the company out. Let's have a look. Split out contacts, splitting out the companies. We have an issue, we have an issue with data here because we have the, the, the it hasn't Targeted the US which is one thing United Kingdom it's gone global. Sweden, Australia, Indonesia. So the prompt needs editing so we need to go back unpin go into here critical mandates all of your research should be in the USA. You can use international information but the company you research must be US based. Okay now we're going to run that back. I'm going to see maybe I was right the first time. We'll find out as we as we go through this as we start to generate the icebreaker I'm going to go in here I'm going to clean this up because this is my email and what I advise that you guys do now is head over to. I'm just going to delete all that head over to sheets New and just title your spreadsheet something that you can select from the drop down and we'll decide on what what we want to put in our in our in our sheet a little bit later on. Company name, email status score LinkedIn description City Tech Stack and Icebreaker seems like a pretty good list so we can do that but let's see what we what we got to with the output now Jason does it say doesn't say category see through it Australia that is the. That is the. That is the LinkedIn URL though let's have a look at what the output the passing output is Australia it does not want to pull headquarters that are only in the UK sorry only in the US so your strict strictly bound to researching US based businesses only where it's strange strange how it ignored that first individual let's execute can't execute that we need to come back out unpin execute this step one more time. Maybe it's just renovation companies get all over maybe that's maybe that's what it is such a growing game that the tariffs might have caused you know huge disruption there with that this is where you get to when you when you are prompt engineering properly and you will and why you need safe JSON passing outputs why you need the mapped out system in your mind and on canvas and also why you should be in a position where you can be comfortable editing this this information. Yeah Headquarters Country USA so I saw a USA in there. Let's pin that now let's execute this see what we get New York USA USA Brooklyn USA San Francisco USA USA USA so giving it the the system prompt we learned a lot there. You've learned a lot about system prompting and how you have to make changes on your feet and dynamically and we have to do that but we can see from, from this we're pulling all of the correct information. Now the problem is this. All of the individual contacts are grouped together and inside of our Hunter IO which you need to go and connect by hitting create a new credential, go to Hunter IO API, create a new key. It's free of charge up to a certain amount of credits, amount of our credits here as we go through this, I have to buy some, copy the key. Well, create a key, copy key, come back, paste it in, click save. Our Hunter IO is now set up and structured. We have that in place. We have the data coming through from the USA and we're in a good, we're in a good position right now. There's, there's some. So inside of here I've gone, gone in. You drag companies over, split out companies. There's only three items. Why is there only three items? Well, I've basically got ahead of myself with the build because what we need to do is we need to add the code node and this node deserves a round of, of applause. Right, I'm going to take the, the code node from that I had before the code from the code note ad before. What this is basically telling you to do is companies when, when, when the constant companies equals input item JSON.com companies return companies.map and it maps the companies together. Return JSON company so it's all grouped in one and this is the power. But you can ask AI to group to group. Let's do it in, let's do it in our flow, right? Let's do it in our flow. Let's just do it in here. I want to group the 10 companies together. I want to group the 10 companies together. Let's just hit generate code and see what it does. Right. This is, how powerful is this going to work? Code generation. Your workflow is too large for pro. Simply simplify the data. Okay, well maybe not then. This is, this is what it should, it should return now. This should in theory group, group companies together. Okay, so not only have we worked with logic nodes, we've also worked with the code node now which is really advanced. Right? The code node is incredibly advanced and you. But the thing is, is because it's JavaScript inside the code node, AI and LLMs are fantastic at writing it and often you'll find you can cut like four or five nodes out if you're using the code node very well. I don't want to go into too much detail about it today, but let's execute, see what, see what happens now. You see, we've got a perfect, we've got a perfect output of 10 companies that are grouped together. And we didn't have that before, we just had an array of companies. Now they're all grouped together as 10 items. It's much more, it's easy to read, it's easy to handle and it's the contacts that are the issue that we need to, need to branch out, which is exactly where the split node came from earlier on. So in here I'm going to change this to contacts. Contacts, split out contacts. Let's execute this step. And what this number should be should be somewhere in the region of 30. No contacts, 22. That means that because we told it to find a minimum of 2 per, per email and up to 3, if we go back to the previous node and look at its output, you will see that some of them have three emails, some of them only have two and that's why the number is lower. A lot of them do just have two. So that's where that number came from. That's absolutely fine. Now this failed. Why did it fail? We're going to pin this data. We're going to pin that data. Do I need to go into any more detail about this? Whenever you see a code node, take it over to what your favorite AI system. Turn the temperature down. If it's, if it's, if it's possible, please can you tell me exactly what the below JSON does? The below JSON that was inside of a code node in NAN does. You'll find a lot of templates that you'll probably be looking for and running through. Do have this like code nodes inside and you don't need to learn to write JavaScript. You don't, I don't know, I don't know how to write it properly. I use AI all of the time. Like the best way to get the quickest job done is be as lazy as possible, which is a really weird concept. But in high level summary, the code takes a list of an array of companies from a single incoming item and splits it up so that each individual company becomes its own separate item for the next nodes in the workflow. You know exactly what our JavaScript node does just by literally asking an LLM, you know, I'm too busy trying to make money and win projects to go and learn how to write JavaScript. There are no code backgrounds for platforms for a reason and we should keep it that way in my opinion. So now that's where the, the split outcomes. We've got 22 items. Now we've got 22 items. We can just pull in the email and execute. This is now going to verify with Hunter. I just hope I don't run out of credits. This is. Is 9 or less, otherwise it's going to fail. Grouping the companies together was such an important place because we're in a position where we need to have individual isolated entities like we need it. In some cases we wouldn't have needed it, but here we. We certainly did. And we. That's why we've been able to do what we. What we have done further down the road. And when we get to the email, email Icebreaker, we'll pull from split out contacts all of the information that's needed, which it will. But what's been provided? Company description. Description. And then we're going to see here the valid emails. How many of them are valid, hopefully in the next 25 minutes. Whilst that does its thing. Whilst that's doing its thing, let's get the sheet made. So master sheet. Thought I was gonna say something else then. Yeah, Master sheet. And inside this master sheet I'm gonna select append or update a row sheet within document. Lead enrichment YouTube. And the sheet is sheet one. It's gonna fetch what the information I've got and this is what I have for the company. So Hunter IO is rolling. So it's time for us to look at our sheet to make sure that lead enrichment YouTube. I've called it that. Call your sheet whatever you want. Company name, email Status, score company. LinkedIn, company description. City. Don't know what. That's what that column is meant to be. Maybe it'll be in the data somewhere. The validation's done, which is good. Let's have a look Verified. So score Result Deliverable, undeliverable. So in here the result is equal to deliverable. We want to drag the result over. Is equal to deliverable. You can see there, that's it. We execute that step. It's only kept seven items out of 15, so we would have been in trouble if we'd emailed all of those. And now I've got a sheet. Oh, it's already mapped. It must be the old one. When it comes down to this. Might tell me what's missing. City. City. Tech stack. There you go. Tech stack. Tech stack. Now, don't worry, I know your screen might look like that, so I'm going to do it with you again. Clear these out and effectively they are. They are just mapped. So what I want to show you is something interesting again. We're Going to write some manually. Okay. And when you look at these screens, when you look at these screens, you're going to have an option of column to match on. And that needs to be a unique identifier. Generally speaking, it's an id. In this particular case, we don't have an id, so we're going to. And the businesses are mixed. You can see how this would cause a problem if we use the same company, LinkedIn for example. So the email is always the, the same is always unique. No one, no two individuals have exactly the same email. So we're going to use the email and we can see from here. So we've got email, we've got score status, we've got status. Other than that, we haven't got anything. But that's the beautiful thing about these three tabs. JSON shows you the JSON output. You can copy it by clicking that, which is important as well. Table view is exceptional for clarity. Shows you here, shows you the array, shows you everything that you need to see. Like, you know, glance of the businesses and then you have, we have, we have schema and schema is interesting because if we close that tab, we can access any of the individual items inside of our workflow from here at any point, any variable. And that's why we're going back to split out contacts and we going to drag company name, their name, their actual name. Where was that split out contacts name, company name, individual name. So like for company LinkedIn, let's do this. Yours will be probably be fixed already. So if yours are fixed already, press the equal sign or just press the express expression there and then you go double quotes. You go down to split out contacts.item.JSON.company LinkedIn done. So now we want the company description. This is, this was already a, a expression. If yours isn't, just click it. We go down split out contacts.item.JSON. company description, same thing. Now the only reason I'm doing that is instead of dragging and dropping it is to just showcase to you how you, how you do it. Dragging and dropping it is by far the easiest way. And here we want to pull the whole array, not just an individual element of the array. We want to, we want to pull the whole array so we have the whole text stack. It's going to append or update a row execute step. The column names were updated after the notes settled. Let's refresh company. Any of us. Yeah, well that's okay. There you go. Tech stack. So we need to refresh this, refresh Fetching columns, tech stack, pull over the whole array, hit execute. You won't have had that issue because I messed around with the column headers on accident. The seven items have been output. Are they in the sheet? They are, but there's seven below because I've got a writing in there. You see that? I've got. I've got an obscure thing. So what I'm going to do is I'm going to delete that. Your sheet will probably look bang on, but I'm going to delete it and I'm going to make sure that it does it again. We're going to rerun that, run it back from an execution standpoint. Execute step. We've missed a key and core point of pinning the validation. There we go. Now we're cooking. So we have, we have this options to us. We could filter by score. Like we could have a filter if the score is below 80, then done and done and leave it. You know, we could have another filter there or could have added another condition and score is greater than. Is greater than. It's a string though. Yeah. But it now is a number. So. Yeah, number is greater than 80. You do that. Commands set, command S, save. That means all of our items will have come through anyway because we've got 88. But it stops you getting any ones that are lower. But they do exist. Our sheet is now updated with the correct information, which is brilliant. Okay, and now we're going to do something pretty cute. What we're going to do is we're going to. I. Well, I am going to join along. I think this is going to be a good idea. Let's. I'm going to zoom out just a little bit and I'm going to get a screenshot of all of this information and we're going to go to AI Studio, going to paste this in. We're going to say, please can you write me a system prompt for cold email Icebreakers. But before that we tell it. You are a veteran cold email copywriter who knows how to make people gasp at the quality of your introductions. Please can you write me a system prompt for cold email Icebreakers. The LLM will have access to the below information. You should strictly output two sentences per email. Yeah, per email. See that? Let's hit it. See what it does. We're giving it, feeding it information that is real life information. That's what's important here. We're trying to get it to go without just copying my standard, standard flows. And we want to see, we want to see it work. So we're going to try and get it, try and force it into this. Let's hope there's no funny characters or anything inside of here. Once this is done, I'm gonna test it. The first email output is too generic. You're better than that. Go and dig deeper to find compelling copy that really stands out. That would have worked, but we're still trying to improve it. This is for you to play around with later really when you have the, the flow built and you're sending messaging. Messaging the models, that's fine. If you run out of Hunter IO credits, just create a new email. Like that's not an issue either. You can do that better. Better. But please make them more humanized human like. Even humanized human like. And then once this is done, we all we've got to do is select the right sheet, append or update a row. I've just come back in here to select the document, select the. The sheet. This will be select column to match on email and then we delete everything except the icebreaker. So let's see what AI Studio has now come up with. This isn't. This isn't. This isn't the best that problem that you'll use for an icebreaker, but this won't be. It won't be the best prompt that you've ever used. I pasted pasted it in the way that looks paste and match style. Okay. And now as a prompt we're going to give it all of the information. So name, name, company linked in company description. Notice the distinct difference between the prompt where I'm putting the user message and where the system message is in an LLM. It's really important that you, that you know that city because you can get in the wrong way around company and probably the important part of all this, their tech stack system. You're a seasoned strategist. Okay, execute step. Let it do its thing. I've got this matching on JSON txt for the. For the fly but I'll do show you now. Gina is interesting to see how sweet and weaves through weaves together react node and segment with service platforms like Zendesk and Intercom. It feels like you're not matching it feels like you're not just matching homeowners and contractors but also dialing in every touch point of the renovation journey. They're all right, but they're not. They're not what you would put out. But this is where your prompt engineering comes in and good practice for you to go away and make them good Put some in the comments of some excellent Icebreakers. The final thing, we need to rename this Update Icebreaker. Update Icebreaker. And this is going to. If I execute that with just the Icebreaker, will it go in order or will it not? Do I need to add column to match on email? Okay, going to the enrichment gene coder. Luke. They're all the right names in the right orders. You've built out a custom system inside of here. Pin that data command S Today you have learned not only you've learned some of the common pitfalls, you've learned LLM system prompting and usage mainly in the Icebreaker section where we, where we used it logic nodes. So there's been logic nodes inside of here. With these we've connected a third party API, which is Hunter AI. We've used the goat of all nodes, which is the code node. You didn't have to write the code yourself, but there was a reason for us to group something together. And I showed you how to do it. We asked AI together what it meant and that gave us the opportunity to be able to, you know, to really, really dive in. You. There'll be a copy of this attached to the video, but this is the second video of three or four that I'm doing in terms of in depth. The next one will be around using custom HTTP requests, curls, things like that. That will assist you and basically just the way that I would want to learn it, that is effectively the way I'm teaching it. So, you know, if you're still here now, you obviously like it. If you, if you, if you're not, if you're still here now and you don't like it, then. Then you need to fuck off. Because. Because this is the way that I wish I'd have been shown how to map inquiries out, how to. How to be able to map flows out. Understanding the leverage that that gives you in terms of the brain power to then go and add in the logical steps the code node. We didn't add straight away. We skipped over it, even though I don't just built it before. And then we added the code node in when we knew we needed to group companies together. Not the simplest of builds, not the simplest of presentations, but I hope you enjoyed it. And the next one will be increasingly more powerful. Once again, thank you for your time. Like subscribe. My name's Michael. Thank you for watching. Thanks.