Transcription
You may have seen that loops or slash goal have been catching some traction lately. But here's the thing, this isn't new. If you've never heard of a loop, it's basically the Ralph Wiggams technique, which has been around for a while. It's just that now it's gone mainstream and it's caught a serious second win.
And the biggest thing I'm seeing is that people are actually using it wrong for two simple reasons, context and capability. Because almost every video you'll find is about building software. But what about growing a business? What about outbound?
So in this video, we're going to run something completely live, completely experimental. I genuinely have no idea if this is going to burn through my whole subscription. We're starting from a fresh project, no skills, no nothing. And we're going to do onetoone signal-based outreach from absolute scratch. Full end to end all the way to a campaign that's ready to launch. Let's go ahead and dive in.
All right. So, firstly, in order to even just understand what an agent loop is, I would go none other to the actual makers themselves, which is Cloud Code or OpenAI, Grock, whatever you are using. I would definitely use their documentation as the single source of truth because things are always changing, right? This could change literally in a week from now. You just never know, right?
With that said, go to claude.com, which you can see right here, and you can see it says how the agent loop works. Now, agent loops obviously all over the internet, they're talking about running it, you know, Ralph loops, things like that, in order to build apps or build something where it constantly is just going for hours and eventually get a core product. But for this particular use case, we're going to use it for business growth, for actually doing outreach to more or less say like, hey, I have this vision. I need you to execute it and then just do the thing, right? With as least human in the loop areas as possible.
So, just to kind of go into what an actual loop is, you can see right here, execution loop that powers cloud code. We'll click on that in a second, but the loop at a glance is literally your prompt. Okay, it goes into an agentic loop where Claude says, "Okay, I got you. I understand context. This is what's happening. saying this is what I need to do. It does a bunch of tool calls, right? They're called batch commands. It may call CLI tools or even MCP really just external tools which we talk about a lot. But it will call the tools and then it'll evaluate the actual output and it runs into this loop right where it's evaluating and getting it done based on the objective in order to reach a particular goal and then when no tool calls are actually done you get final answer.
This is a very simple way and in software development let's say you push code to GitHub for example right let's say you're shipping something well when you push code it may have bugs it may have some issues maybe some vulnerabilities you never know so then what happens is that agent is going to receive that or the the review agent really and it may catch a bug and it may say okay we need to resolve this and it could do two things could either fix it and redeploy or it can send it back to the original one and actually fix it. We're going to save the app building for another time.
If we actually click on execution loop that powers cloud code, you will see an even better diagram. And this is actually a true Aenta loop cuz you'll see where it says, okay, it's gathering context. It takes action. Verify results. And then really what usually has happened is it will say, okay, this is done. From start to finish, from your prompt, one loop, if you will, actually goes back to done. then you're basically like, "Okay, the next point is actually the human in the loop to say like, okay, what's next?" Right? It just kind of works, but when you're actually in the workflow, it's actually slowing you down, unless you really need to have a human loop to actually verify it, right? Other than that, if you give good context, which with these latest flagship models, they're actually good enough with good steering and good direction, right? You can kind of watch it like a fireplace and just make sure, you know, it doesn't get too big or too small or things don't crash and burn, right?
Oh, with this, sometimes you want to actually steer or interrupt the agent while it's working. The goal here is to be as minimal as possible because agents obviously if they run into a challenge, they're going to basically do whatever they want and go on a wild goose chase or what have you in order to get the objective and figure it out. That could be good but bad because the result may not be exactly what you want. You get derailed, right? So this is where you actually do what you call steering or you may say no like do this or yes do that right you're kind of steering it and so that's where you kind of interject here and so for the purpose of this use case when we're actually doing an agent loop especially when it comes to business growth or outbound you want to think of a loop as something like a conversation to an employee of yours.
So imagine if you're a CEO, which you may or may not be, but let's just pretend that you have a team, right? Because in agent land, you actually have more employees than you can probably handle. You're actually an orchestrator, kind of like a band, right? I may have my flutes over here. I may have my trombones, I have my percussion, I have my clarinets, and you know, I got my wand, right? My daughter will probably like, "It's not a wand, Dad." But you sit there and you're like, "Okay, you got some humming going. You got the drums going." And then you may be like the flute over there. Like you're off, dude. you better fix that or get a little louder or something like percussion. You're offbeat, right? That's where I'm kind of steering that agent and you're orchestrating, right?
But in the context of a business, let's say you're the CEO, you could very well have a head of sales. Your head of sales is now your main agent. So, your number one or number two, like your main gal or your main dude of your business [clears throat] to say, "Hey, look, like this is happening. Fix it." You're now orchestrating and delegating a specific piece of the fulfillment off to now an agent, right? who's like your main person. And so what you do is you use your main harness such as claude code or codeex and you can basically say this is the goal slashgoal is another example of it. And then from here you want to think all the way through the workflow. And what I mean by that is you want to say okay look I have this problem or I have this goal and I want to do it in this way. So then what you do is you actually map it out or you plan it out to say like I want to do this. This is how I want to do it. Here's a good example. Here's a bad example. This is the steps in order to do it. And then you tell them based on that one thing how you would actually delegate it to other employees like junior level managers basically all the way through top to bottom from literally you as a CEO to a director manager to individual contributors to actually do the work, right? And then it will go back up the channel to go back to the CEO to say, "Hey, it's done." So that is kind of what I'm talking about. And each one of those hierarchy conversations are literally loops.
And if you think about it, each conversation is nothing more than kind of like a context window. Context windows are really finite. They're actually pretty small when you start doing a lot of agent work or anything like that. The context window can actually get quite bloated. And you want to be precise. You want to be in a way where the context window is empty enough but not quite full enough. And you want to steer the AI agent to be focused. So what I mean by that is you want to be in a direction where it's covering one objective and have it delegate based on that one focus. Because if you say something and then you go off to another end, you want to be sniper focused based on the workflow compared to zigzag where you're like, "Oh, but do this. Oh, but do this. Oh, but do this." Right? You're starting to zigzag your AI agent. And just like driving, if you drive in zigzag, you're going to have a bad time. You're going to get sick and it's not going to be a good time. So try to think of it like a sniper shot, right? If I can direct an agent and then it were to launch other agents to do the job. Think of like employees, I think your results will actually come out pretty good and we'll see what that happens, right?
So what I'm going to do from here is I'm going to start with a brand new project and we are going to start from literal scratch. Like I literally did not set up anything for this purpose because I want to demonstrate the results what we can actually come up with. And I don't even know if this is going to exhaust my cloud subscription or who knows what. But you know what? Hopefully it's good entertainment. At least hopefully take away something. What we're going to do is we're just going to open up a folder and I'm just going to go to developer folder. I'm going to start a new folder and we're going to say claude loop demo. And then we're just going to go ahead and open. And now I have an empty project. We're actually going to go ahead and toggle cursor agent. Cloud code should be booting up any minute now. There it is. And you'll see here I have actually two claude pinned here, right? So, one of them you'll notice actually is the one I like to use the most, which is I can basically stand up enough tabs that I want. And then this other one is really just kind of the left pane. We're not going to use that today. So, you can see here there are no folders. There's literally nothing that is in this at all. And so, we are going to set up our cloud instance to actually do some work.
So, if I hit forward slash, you can actually see there's quite a lot of toggles here, right? So, we can go account usage. And I'm looking pretty good now. It did just reup the month based on one of my subscriptions. We're going to leave it to default model because I want to specifically direct it. We'll see what does, right? Because I'm not going to say default's the best, but I want to make this as easy as possible. So, we have a number of different things here. It's just really good to be familiar. Okay. Then, we have slash commands, and one of those things is going to be loop. So, we're going to go with a loop command here. And we're actually not going to start just yet, but you can see right here, this is the autonomous loop default that dynamic pacing mode. I ran the check now. There's no prior conversation to work to continue. And this directory isn't a git repository. So, there's no PR branch because it's thinking about software development, right? Schedule a recheck in 30 minutes. Nothing to do for now. I'll look again then.
So, I'm actually going to start a new chat just cuz I don't want to confuse it here and I don't necessarily want it to do anything. What we're going to do is firstly I'm going to use spokenly which is my speechto text application because again I have yet to meet a single person that can type faster than you can talk. So the best way to do it is to actually just talk to your agents. This is a brand new project in which what I would like to do is to start with a fresh context window of understanding what we do as a business and running a cold email outreach campaign in order to gain customers and gain clients in order for business growth. The company that I run is called Topfunnel which is topofunnel.com. My name is Brandon Charlesson and I run the agency. We are a growth marketing agency to where we build custom agentic systems and automations around business growth specifically for sales and marketing but usually always bleeds into other departments as well. And what we do is we assist and facilitate growth leaders such as founders, sales teams, marketing in order to increase topline revenue in such a way that doesn't necessarily need additional headcount. But it allows for the existing personnel or employees that are part of the company in order to maximize their leverage and their ability to send value propositions or their offer to their targeted ideal customer profile. This is what we do. So we are largely eating our own dog food in terms of how we approach our own customers because we build the systems that we not only use but we also build the systems that we deploy to our customers.
So what I want you to do is to understand context first before we continue with anything. And I want you as the main claude harness to research what top ofunnel does. I also want you to know best practices around cold email and the skills at which I want you to get is around what it takes to actually have a good cold email. A good cold email is usually a short message of around 75 to maybe 100 words at most. It should be able to be read on a mobile device and it should be able to be read in maybe 5 to 7 seconds at most. The value proposition should be so good that it is a no-brainer to actually say yes to it. It needs to be factual. It needs to never be fake or lead the prospect into a way that is disguising, but it needs to be truly insightful, valuable, and a good piece of information that leads the prospect to believe that how did they even know about this message? It provides a signal that we know something about their situation that they may not be already aware of or if they are aware of it, it's due to a specific signal or something that we noticed in which we can help them resolve for that pain point or that problem.
One of the biggest things is AI adoption, especially at enterprise level. One of the biggest things on that is to be able to safely and securely deploy agentic enterprise systems especially around non-technical departments such as sales and marketing in order to grow the business. A lot of times engineering is focused on products. They're focused on security. They're focused on a lot of roadmap things in terms of building new products for organizations. And a lot of the sales departments and marketing is often left running in silos to the point where they selfs serve and they sign up for tools of all types of different SAS companies and enrichment tools and they often find challenges and waste time in regards to actually moving the needle and closing deals, which is connecting humans to other humans. Everything else in between is really just a workflow where you're taking a piece of data and you're moving it along in order to actually come up with a good message or a good touch point for your prospect that you're trying to reach out to.
The thing that I generally see that I want you to take into close context is that a lot of AI slop is around usually where it is immediately obvious that AI has written this message. Some of those things include saying I hope this email finds you well. Some of these things include m dashes. Some of these things also are very generic small talk types of emails. We want to do none of those. I repeat, we want to do none of those. The intent here is to actually reach out to the person. And the first sentence, which is the hook, should be immediately obvious and immediately relevant in order to capture the person's attention at first glance because that will indicate and that will only be our opportunity in order for them to actually read the message all the way through the end of the message, which is the call to action. And then the next sentence should be ideally the body or the bridge between the hook and the call to action. And the bridge needs to be in such a way that actually makes sense for your first sentence. And I'll give you some good examples and some bad examples.
Some good examples are, "Hey, I saw on your website that there were some areas that had broken links, which leads me to believe that your website has been neglected or recently went through some sort of downtime or migration that wasn't already addressed. We actually help deploy and fix better websites and we've worked with many other companies. I can draft up a comparison on what that new looks like. Would you mind if I send it?" So, I was completely spitballing that, but another example would be in a practical scenario of if you're driving, right? So, I know you're an agent, but if you were driving down the road and you notice the car next to you at a street light has a flat tire and let's say you know how to change a flat tire and you thought, "Wow, this person is driving on a flat tire." That is dangerous, not just for them, but for others, right? So, if you were to roll down your window and capture their attention, you don't even know who that person is. But if you said, "Hey, I noticed your tire is flat. I have changed many tires before and I can get this done in 4 minutes. Do you mind if I help you out with that?" It would be really hard for that person to actually say no to that because you've immediately noticed a painoint that they need to address. they bought into you to know that you could actually change it very quickly and you offered help without actually saying, "Hey, I charge $20 for this or whatever. You literally just offered them something that actually helps them." And the actual transaction or the reward or whatever the monetary exchange, the value exchange is a byproduct and it should never be the core thing because anything that is ever done by humans as a result of benefiting somebody else usually comes with money. But the thing is is you shouldn't always expect it. But it should be in such a way that it's just naturally known that in business that's what you usually do. So in the context of the spare tire or the flat tire, the principle is the same. And you are immediately noticing a signal. You show to the cold prospect that you can be trusted with case studies or anything that bridges the gap. That's why I call it a bridge. And then you do a call to action that is no-brainer to them to actually say sure or yeah, that would be great.
Since this is a fresh window, you got to give it context. So I'm literally going to see what this does. The other philosophy that I have when it comes to actually coming up with great email copy that I really want you to generate as a skill for your own purposes is really called the sew that loop. And what I mean by so that loop actually it's more of a ladder. So we're going to call it the ladder. The so that ladder means that if you ask yourself what you do in this business that's just layer one. So if you said I sell Agentic Enterprise Systems so that we can help businesses automate their sales and marketing at a scalable fashion so that teams can get more sales calls in order to grow revenue. We're at level three now. So that your CRO or your CEO is not coming into the room thinking why have we not had any sales calls? Do we need to hire more sales reps or dial, smile, and dial more so that ultimately the business can grow? In that just example right there, we have already gone at least three levels. We've gone about four, maybe even five levels. The sweet spot is usually at least the third, if not the fourth level. But the reason I say so that is to continue the sentence so that people can obviously understand in the context of their environment. You are not just pitching your product because you ask so what like so it does that so that your offer can benefit them so that extended benefits so on and so forth.
Here's what I want you to do. I want you to firstly set yourself up in a loop fashion where you as the main clawed agent need to be on Opus. And your task and goal is to come up with a large lead list of prospects that is verified emails. We have context of each lead as if you researched them for at least 30 minutes. And our goal is to do onetoone outreach to each and every one of them and for us to come up with a message just as I explained that is so good and so relevant and the offer is also so good that they saying sure. So, with that said, I want you to plan this out strategically. And then what I want you to do is to use instantly CLI because that is our cold email platform as well as the leads that we're going to be looking for. Do not use any other tools for this use case because this is going to be a test to see how far we can get. And then from there, do not hit send. But I want you to at least create a campaign inside of instantly. I want you to upload all of the leads to this campaign. Each of them have a onetoone message to them. I want this to be an extremely impactful and good campaign. So much so that it's actually successful. So your objective is to really make sure that you do every bit of execution and resource that you can do in this for being able to delegate and do this swiftly and in a reasonable amount of time. I want you to spawn sub agents using sonnet and I want you to execute these swarm of agents in a small batch fashion in order to further research in parallel each one of these prospects. Some of the details include scraping their website. Some of the details include understanding context around who the actual ICP person that we're reaching out to that is relevant to our approach. This doesn't need to be extravagant. It just needs to be enough to start a conversation. We're not trying to obsessively research them on how they went to Disneyland with their family or they love boating or that their website is amazing. That is all small talk nonsense and BS and nobody has time for small talk in this scenario. This is a cold email. It needs to be direct. It needs to be to the point. But most importantly, it needs to be human. And if we can so much as use any type of relevant good taste humor that's appropriate, I would encourage that because it actually strikes emotion and we usually want people to laugh versus get angry or any other emotion.
With that said, install instantly CLI. We're going to go at instantly AI CLI. We're going to do my forked one. We're going to go via instantly npm by Instantly CLI. There is the official one as well. Definitely encouraged to use that. I like to also continue momentum here. So install instantly CLI and we're going to do it uh at the project scoped fashion within this project and not global. I will provide API key. Let's go ahead and do that. Go with API key. We're going to go integrations and we're going to say all right claude loop demo. By the way, this is how you get an API key. The API key will not work after this. This is just for demonstration purposes. put this API key in the ENV file within this project and ensure it does not go anywhere else. It will be rotated upon this project being launched. I want you to remain focused again. Start with scraping topofunnel.com and understand what we do and how we benefit customers. We're going to go loop and let's hit enter. So, we're going to let this rip. And I know with that feedback loop, we'll see what happens.
All right. So, I watched it for about 5 minutes and it has written a whole lot. So, it says map instantly capabilities and it went through a bunch of things here. So, we have our first uh human the loop which is fine. And the reason is I've mapped out the full pipeline before I spend enrichment credits and import leads into your live top offunnel workspace. Two decisions materially shape everything downstream and they're genuinely yours to make. I actually think that's a pretty good call. You know, that's subjective, but I actually appreciate that. And then we can see right here. Okay. Spawn sonnet sub aents, which is exactly it. Draft one to one cold email per lead. Create instantly campaigns and upload leads and no send. That's exactly right. So, which ICP? Yep. B2B SAS founders. Sounds good. And we'll choose that. How many leads for this first campaign? Each gets 30 minute deep research and a true 1:1 email. That's right. Recommended for first test set. Okay. Tight, tight, high quality. I want to fully lean on you for this objective. So, let's go with 2,000. We're just going to let it rip. Usually, I'd like to do a sample set of 10, maybe up to 50, and then we kind of scale from there. But, they're just emails, right? So, honestly, cold email is kind of like your unlimited first impressions if you think about it. Because how many cold emails do you remember getting, right? So, if I were to reach out to you or somebody reaches out to you, let's say they reach out to you tomorrow or a week or 6 months from now, you're not going to remember. They're just emails at the end of the day. See here, it says locked B2B SAS founders, one honest approach. Then I execute 2,000 genuinely 1:1 emails and a system build, not 2,000 literal 30 minute manual searches. We'll see what the next step is.
And now after about I think 20 minutes or so, we have a series of things that has gone through, right? So at the end of the day says status checkpoint massive progress in this iteration here's where things stand foundation understanding API key stored in the get ignore good to go confirm for instantly CLI research top of funnel which it captured information about our business and stored it into context codified your cold email standards into cold email playbook lead sourcing via instantly super search. So, we discovered Super Search. No industry. Okay. We did a lookalike for Gong. Good deal. Locked in a filter. So, it sourced leads right within Super Search for almost 2,000 leads right there. And we have 1766 verified work emails and domains. That's a good start. Full pipeline proven end to end. Okay. We paused the draft campaign that was created. So, it already created the campaign and it says cannot send cuz I do want to verify that with a custom variable sequence and two warmeders. Okay. Well, we'll add more emails there. A swarm of sonnet agents scrapes each site. Finds a real signal. Writes 1:1 hook bridge CTA. Drops non-fits. 82% fit. Build a deterministic sanitizer that kills M dashes. Sonic kept emitting them and ban phrases. 38 leads now in the campaign with genuinely good 1:1 copy running now in the background. So I deployed a workflow in researching and writing copy for the remaining 1721 leads. On completion, I'll sanitize, upload, and rerun any failed batches, do a final QA check, and hand off the paused campaign for your review. I'll be woken when the workflow finishes. Setting a fallback heartbeat so the loop is resilient if that notification is missed. Schedule wake up. It basically says, "Okay, go to sleep and wake you up when it's done. Right. Next wakeup is scheduled for this checkpoint complete. Research copy storm is running in the background and I've set the fallback heartbeat. When it finishes for the time or the timer fires, I'll automatically resume to sanitize. You can now watch live progress with forward/workflows or tell me stop anytime.
So I'm just going to let this rip and we'll see where the next checkpoint is. So, we are finally done with the loops. So, it's gone quite a bit. I honestly didn't time it, but I would venture to say it went on for maybe about an hour or so. So, from here, uh, it says its final state is 1962. Again, it's a good start. You know, when you're doing any type of volume, it's definitely not a lot. But the key here is the goal here. So, you start with a very small sample set of could be 10 to 50, making sure that the copy and everything is dialed in because as you steer the LLM, it's getting, you know, a lot more context around what actual good copy looks like. So, if you start with a sample set of say 10 to like 50ish where each and every one that could be human reviewed is good to go ideally when you do this, but in a loop, this is obviously experimental, then we go for much more bigger sample sets. So you start to roll in like 300, a thousand, that kind of thing. For the context of this, I just wanted to see what 2,00ish looked like. And here we are about to review it. So we can see final state 1 one paused. Good to go. Cuz we don't want to just yet arbitrarily just press send. I mean, that could be bad, right? Without us uh actually reviewing it. 1962, not a flat 2000. The high fit lookalike cohort capped at 1988. Effectively, it has done all the steps that it needed to. And then it stopped, which basically said the campaign is now complete, optimized, and send ready. I've reached the point where only the remaining action, which is activating, and sending is yours, not mine. So, I'm ending the loop rather than spinning on make work, which is really good.
From here, where it stands, we have 1962 leads. All right, we have a three-ouch sequence where we're doing a one one and then a bump, then a breakup. That's really interesting. Text only stop on reply. Again, this right here, I haven't even looked at the copy, but this just tells me the types of training that these LLMs are getting is very legacy and boring, right? I mean, the things of doing a bump this to your inbox. I mean, let's be real, that kind of crap doesn't fly anymore, right? And then the breakup, like seriously, why would you send an email to remind somebody they didn't reply to you and then add pretend as if you had a relationship in the first place and just arbitrarily annoy them? So, I know for me, I don't reply to that stuff and I for sure as heck am not going to send emails with that kind of notion either, right? That's for the amateurs.
So, what we're going to do is we're actually going to review it. This is again a project that I started from complete scratch. So, in the context of iterating and building skills and when you have something that is a really robust and focused skill around your business and what good copy actually looks like without actually relying on the LLM, that is where fine-tune training actually comes into play. So, with that said, let's go ahead and take a look at the copy. So, we see here we have the campaign which was actually created. We have a sample set of 2,000 leads loaded. That's great. And let's go ahead and look at it. So, we have subject line, high first name, hook, bridge, call to action. This is the type of copy that I like because that means everything is custom. So, if we hit preview, we can see already that the first name does not actually populate. What we could do is just take a look. Well, we see right here this first name is actually incorrect. There's several types of actual common casing. And I think I've talked about this before on previous YouTubetubes, but we have different types of popular casing. So, just not to go off course here, but you really want to be keen on these types of actual casing because it's going to populate incorrectly or not, right? So, this is actually called snake case. So, if I were to even type it out, snake case, that's what that actually looks like. Whereas the actual variable that instantly accepts is called camel c case. So if we did first name, you notice that the N right here is actually uh capitalized and therefore it's called camel case. The other one is called kebab case. Kind of like you know you have your meat and vegetables on a stick. So we have kebab case, we have snake case and then we have camel case. Think of like a camel with their hump. You know what I mean? So when you're looking at these variables here, you can see that these variables are very simple. But in the context of first name, these are the kind of things that depending on which tools you use, you want to be very very focused on actually doing the attention to detail, right? So if I were to even just change this to first name, you can see right there that it actually changed it to camelc case. And if I were to preview it, we could see Michael actually populates now, right? So we're going to go ahead and save that.
But we have the first name, the hook, the bridge, and the call to action. So we have, hi Michael. Your positioning at Vudit's enterprise tier for CFOs and FinOps leaders, but the only sales motion on the site is a colony link. All right. Is that relevant? Sure. Right. Only Michael will be the judge here. That gap, a PLG signup doing the work a full enterprise outbound sequence should be doing is exactly where we put AI agents for sales teams. All right. Good deal. This is saying we ran a similar outbound system for a SAS company and a 28.8 reply rate. I know this is hallucinating because earlier in this video I never actually said any case studies. So regardless if this is actually true or not, this is not right. I mean I have done those kinds of reply rates, but I didn't give it specific examples. Therefore, I know it's BS, right? So we want to steer it and call like it is. You know, I'm not here to sugarcoat anything. But uh that is not good copy, but it's a good start. We have a greeting because we're all human and we like to be called by our first name, at least most people. We have a relevant hook which is really your only chance to actually capture their attention in literally 1 second which will actually be up to your demise on if they actually read the rest of the email. And then the actual call to action. The call to action obviously with the offer should be in such a way that is immediately obvious to your prospect and actually be like sure or yeah go ahead and send it or absolutely. The biggest thing is you're asking for time or you shouldn't be anyway unless you have like a million dollars or you're selling Ferraris. But you should be at a point where you're just really trying to peek open the door and it's doing a soft call to action to say, you know, something like if I could give you a quick tear down of what a outbound sequence looks like specific to your offer, would that be useful? Or something like that. So, to the point where you're actually reached out to me, regardless if you use AI or not, but you're reaching out for something that's relevant based on timing or my situation or your prospects, I should say. And then the body or the bridge, what I like to call it, you're actually bridging the gap between relevance and your call to action. So, the bridge here should actually be something that is recognizing the environment or the problem and then you're building some sort of relevant trust to actually show that you understand and you can be trusted. That is the biggest thing and the biggest takeaway here. So without even looking at any other ones, I mean honestly that was kind of not the best. So I'm just going to refine it here. And then from here once I actually validate that a few samples are actually good. Then you know you can start to comb through and really see to make sure that the AI model is actually generating the best ones. And then we run it at scale. So, ideally, we'd run, let's say, a couple sample sets, you know, of anywhere from 5 to 10,000 leads. And then, if we start cooking, you know, start getting some replies. That's when you want to add fuel to the fire, right?
So, we're going to go ahead and prompt Claude and uh see if they can fix that for us with the instantly CLI. So, this is a good start. The hook is initially good to go when it comes to relevancy and timing. and I think you did a great job in terms of deploying the sub agents in a way that actually resonates with them. With that said, there is one thing that I want to call out in this is that the first name you actually pushed to instantane as snake case. I just want to make sure it actually knows it right here. We're going to go here and first name when it should be camel case which is first capital N. I went ahead and corrected that. So in the future, please ensure that any future campaigns that are pushed that the first name and any other standard ones like company name are standard to be camelc case. With that said, I did notice that your first sentence is good to go. So that is a strong relevant hook in which we can test to see if it actually resonates with the ICP. Again, I want to remind you that we should sound human. We should not use any type of typography or words in such a way that does not immediately sound human. And I also want to make sure we do not use any m dashes. Regardless of the fact if it is AI written or not, the overall thing here is we need to ensure that we are sending something that is insightful. We send something that is immediately obvious and valuable to the prospect and that most importantly it is accurate. In a lot of the examples that I went over into the copy in the bridge, you are specifically calling out that you did 20.8% reply rate on a particular data set of sends in which I did not give you any context on what case studies that we've done. Although that claim is a bit ambiguous which it doesn't actually say what the campaign was or the nature of it. So regardless if that metric is true or not, which we have done campaigns like that, I would not necessarily want to be in a position to actually show that because it could be confidential information and it could honestly lead the situation into a situation where we're not really opening the door to the prospect with the best intentions in mind. Basically making a simple claim. The goal here is actually much bigger than that, where it goes past the reply rate. So the reply rate is also not the keen and the the best metric to actually go off of. The reply rate is actually a vanity metric if you think about it. So reply rates honestly do not matter in this whole scenario. The goal here is to get impressions and eyeballs around our email. Here's an example. If I were to send an email to thousands of people, let's just for happen stance say that we are landing in the inbox and we're not even asking for a reply. We're just basically asking if they would be able to provide feedback or we can direct them to our website in a funnel scenario where they are not really interested in replying in the first place. Most of these prospects, if you think about it, well, there's three scenarios. They will either reply yes or no, or they will reply and say go kick rocks or unsubscribe, or they will not reply and take no action. And finally, they will not reply and they will instead do selfserve research through their own means such as sniffing you out and checking your digital presence. They will check out your website and only if you are trustable and only if you are offering something that is actually valuable to them, they will likely selfserve and book a demo. So therefore, the reply rate actually becomes a soft metric versus a hard metric. So leading an enterprise customer to believe that you are claiming any type of actual percentage really doesn't mean a whole lot if you think about it. It could be 2% with a killer offer and a killer demo and very high ticket or it could be 20% and just a bunch of noise and conversations that mean absolutely nothing. So with that said, it really doesn't matter on the actual metrics. What actually matters is the outcomes at which generated pipeline and ultimately closes. Without actually making any false claims, I want you to really understand our offer, our value proposition. And we're not really looking to close or send an email and basically make a value proposition. What we're trying to do is to show the possibilities of being able to build enterprise agentic grade workflow automation and being able to actually produce outcomes without increasing headcount in such a way that is secure, scalable and reliable. So with that said and finally the call to action is I want the call to action to be worded in such a way that actually is a no-brainer to actually reply to. And some examples would be if I could provide you a go to market campaign in a three-step sequence that is actually onetoone and something worth considering, you know, for your next motion. Would that be useful to you? The goal here is they should be able to very easily reply. Sure. And even if they don't, we're really not asking for anything from them. We're just trying to provide some additional insights because at the end of the day, a lot of this stuff comes down to good taste, comes down to careful strategy and execution when it comes to articulating offers to a prospective person.
All right. And then let's go ahead and throw some more stuff at it. So, we have step two. Again, we have snake case, not camel case. Bumping this up. the offer to send a quick tear down top of funnels leaking pipeline but it's built specifically when we put it together. Okay, let's call like it is. This is terrible. I'm not trying to sugarcoat anything here. So, we're going to say in the followup is actually awful. I never want to see this again. And let's go for the third one here cuz again they talked about the breakup. So, I'm not even going to look at it. But last note for me, I'm not clearing your inbox. Okay, that's just going to annoy anybody. Here is the last one. Again, it just goes to show that AI models are trained on old clunky campaigns that don't work in 2026, let alone even to the future. On the campaigns for step two and step three, I have provided to you examples of what you have put in here. And honestly, these are very embarrassing and you should be ashamed of yourself. This type of stuff is old and even if it has worked in the past, it would definitely will not work in today's. A follow-up is actually not an opportunity to remind them that they didn't reply to you. A follow-up is actually an additional opportunity to provide additional insights or context around your first message. And here is an example. If I were to send you a message, we're completely cold. You don't know who I am. You don't know who the company is. You don't know anything. And we're already in a battle of deliverability and getting their attention in a very saturated email inbox for wherever the person is checking their email. So we have a split second to capture their attention. So when you are sending a second email, the last thing we want to do is to remind them that they didn't reply. In fact, we want to do the opposite. We want to just provide so much insight and value. Then if we're going to chain on the first email, we need to say something additional that we understand their context, we need to say something additional to provide them even more insights that we did not have enough room in a sub 100 message on the first step. And we need to make it even easier for them to say yes or sure as a result of the first message. The same goes for step number three, if step number three is even warranted. In fact, step number three should be probably an email that is new, meaning we are not using the same subject line. We're actually generating a brand new email, even though it is a third step, but perhaps maybe a week later to
Just send a new email with a completely different approach in almost the same regard to actually reach out to them when a scenario such as this. Most people don't even remember a cold email they got 5 minutes ago, let alone a week ago. So with that said, I need you to revise step two and also step three in such a way. And once again, the first name variable must be camelcase and not snake case. I want you to continue on this loop until these are done. Before the loop ends, I want you to verify for quality and accuracy. And then I want you to stop and pause for me to preview. Once these are done, we will scale beyond 2K to know that our copy good to go. And then once we have our skills and persona all dialed in, I want you to document these skills and learnings so we can continue to use this.
So we're going to send this on another loop. Again, this is a fresh loop. So, as you start to kind of build on context again, we're almost training like a new employee, right? So, I literally took a fresh slate. I've given it a bunch of loops to say, "This is where you're starting. This is where you need to end. Stop here." And we're using flagship LLMs that are very close, but they're still on old training data. So, the intent here is to show you guys like the loop exists. The loop is doing it. It's not like a hey go. The loop definitely continues. And then from here, the leverage here is getting it so fine-tuned, having good skills, good context, very sharp steering when it comes to the LM delegation, no different than what you would do in a scenario where you're CEO with employees. And then once you verify it's good to go, the repetitive and self-learning auto feedback loop, similar to Andre Carpathy's auto research, this actually will start to compound. And that is, my friends, that's how you do it. So, we're going to let this loop continue and we'll see where we start from here. All right.
And after a solid probably couple hours, this thing went down the rabbit hole of loops and is finally here. So a little bit of muscle, a little bit of training, a little bit of tough love at a fresh clean slate. We have finally come to hopefully is going to be some good copy where we can actually press send. So we have batch two complete. And what I did have to do was add a little bit more credits, add a little bit more data cuz a lot of times with any one data source, you really want to integrate a waterfall method that actually you have a lot more leads. So with that, I've come up to about 8,900. So almost 9,000 leads and added in this batch. Cool. So we added different verticals. We have verified information here. Clean copy, real signals. And that's the biggest thing here is we're doing onetoonebased signal-based outreach, which this is in my opinion the best way to go about outreach in the first place. Unless you're offering free candy and million dollars, you know, it doesn't really apply to everybody. So, if you're going specific businessto business outreach, signal-based relevance copy is going to be your biggest thing here. Let's go ahead and check it out. Let's see uh what comes up here. And again, this is what I'm going to do to actually send to people. So, we'll see what it generates. So, we're going to go ahead and hit preview.
We have Michael here. You're positioning VATS Enterprise tier for CFOs and Finops leaders, but only sales motion on the site is a calendar link. That gap matters because CFOs and FinOps leaders rarely find you through a calendar link. A signal-based GTM system could map your full TAM of finance decision makers service the ones actively. Okay, cool. If I were to map out a full go to market system look like for VAT from your TAM of CFOs, would that be worth a look? I say that's pretty good. Sendable, I should say. Same thing doing the greeting, right? Again, the framework says you just want to be nice. You usually always want to say hi, first name if you have it, right? Normalized. And then you can see here there's the hook says that the top of funnel motion is a single get in touch link. The walled garden narrative is a strong wedge. Okay, let's see here. But the studios that have not yet heard of Sanlow are unlikely to find a get in touch link. A GTM system that maps the right game developer ICP and reaches them when they are actively building or expanding their web monetation could change that. If I mapped out a go to market system that could look like for this, would that be worth a look? I'd say that's pretty good. Let's go for Barbara. Poly has Stripe, docuine and octa on the logo wall and 200k companies using the product. So that's some pretty good case studies I should say. But your enterprise growth motion is a schedule a demo button waiting for business or buyers to find you. With 200k companies on the platform and stripe tier logo already in the mix, the enterprise buyers most likely to convert are the ones who share that profile but have yet heard of Poly. A full GTM system outbound could map that segment, detect the right HR and ops decision makers. Good deal.
So what we're doing is we're bringing some insights to say, hey look, you have a calendar link, you have a schedule a demo, people go to your site, you got to be found, right? So the biggest thing is that that is domain capture. So if someone is actively searching for your business, then cool. Like you would think these companies would have that figured out, which is get in touch, schedule a demo, book a call, whatever that is. But the goal here is to actually create a go to market system, at least for what we do, that is signal based and showing like, hey, this is what an actual plan looks like. And therefore for our value proposition, we build the actual system to actually do that, right? And what I'm doing is I'm saying if I mapped out what a full go to market would look like for Poly from your total addressable market of enterprise HR and basically built it around your ICP, would that be worth a look? We'll see where that goes. But I'd say in summary, that is pretty much good, right? So we have step two and step three already good to go. And then we just want to do a last little check. So this one we actually want to do 3 days cuz we don't want to send that right after. And then this one we're going to do 7 days. So that said, I think we're all square and then we are going to test this with an initial data set of almost 9,000 leads and we shall see what comes from it.
So there you have it. Clawed loops after all do take a while. But you've got to anticipate having extremely long sessions. You have to anticipate and articulate every single bit of detail that you can. And I mean take the time with the intention that will only get smarter and better. The more context and the more skills that you actually implement, the better. And for that, I think the goal is actually to run fewer loops and to be more focused on the objective of actually achieving the goal that you set forth. This is effectively no different than, if you think about it, how you run a business as a CEO. We would want their managers and their employees to actually accomplish the task and return it back. So, if you made it this far in the video, I really appreciate you watching. But before you go, everything you just watched runs on one of the core tools that are needed, which is the Instantly CLI. So, click this video right here where I break down the CLI from scratch, what it actually is, how to install it in just seconds, and why it's the way to use instantly if you're building anything with AI. Thanks again and I'll see you in the next.