Transcription
There's an AI tool that can build your entire cold email lead generation system. You might have heard of Claude code, but it is radically changing everything that we do inside of our business. From coming up with campaign ideas, pulling lists, enriching the leads, even setting up the campaigns completely end-to-end.
In this video, I'm going to show you how we use Claude code to one, share SOPs throughout the entire team through things called skills. Two, how we use it to plug into databases to build lists and enrich things very quickly. Three, a secret hack that we use for sub agents to filter for ICP and write personalized lines. And then four is a new kind of lead generation research category that when we say Claude code is changing cold email lead generation forever, that last fourth thing is going to be the most important thing so you want to stick around for it 100%.
Let's first talk about why Claude code is such an absolute game-changer. So, for starters, if you all you've ever used is just a chatbot, ChatGPT, Claude on their web browser up until this point, you ask a question, it gives you an answer back. Maybe it will search the web to find live information, but that's really it. Claude code is so amazing cuz it breaks, let's call that sandbox of the chat window. When you enter something into a chat and you say, "Hey, look at files on my computer." it's not connected to the files on your computer, it has no shot of being able to do that. Claude code though is connected to the files on your computer. It can make tool calls and make API connection requests between two softwares. It can store data so that if you write an SOP once you're always using that SOP going forward. Just a completely paradigm shift in comparison to a regular chatbot.
And you might think that this is really complicated to set up, but that's not the case. I've never coded a day in my life and it is so easy. All you have to do to set up this platform is go to code.claude.com and then you can click on the terminal, you could click on desktop. If you've never set it up before, you probably don't want to use VS code or JetBrains. I personally use the terminal. All you have to do is just copy and paste this into your terminal and you're ready to go. Anytime you have a problem with the setup of any of these AI tools, I just take a screenshot of it and then go to the chat version on the web and just say, "How do I fix this?" and it always comes out to be the right case.
So, now that you have Claude code set up, let's talk about the first game-changing feature, which is called skills. Previously before skills, people would train an AI in order to make its own custom use cases. You might remember custom GPTs or being able to fine-tune models. All of these things, they worked okay, but they weren't fantastic. And so all a skill file is is literally a text file that you can plug into Claude code and it will remember everything that if you worked on this long workflow and you were to say, "Hey, first we do step number one and then we do step number two and then we do step number three." instead of telling it that every single time, you can set the rules, the boundaries, and the guidelines inside of skills.
And so I think for lead generation, there's two things that Claude code does so amazingly with skills. One is you can make a process yourself and work on it back and forth with Claude code eventually just get to the point where you say, "Great, this is exactly every time we want to run a report to see which domains are most likely landing in spam, we want to run this process every single time." The other thing is you can use other people's skills and instead of putting in all of the time to become a great front-end designer or putting in all the time to figure out exactly how an API works and all the edge cases, you can just download somebody else's skill. We even have this repository I'm building up for ourselves called cold outbound skills that right now it's got four main skills in it where we already have a skill that grades your cold email copy. This cold email copy grader is based off of every campaign we've run in the last year and the way I trained it is I said, "Hey Claude, guess how successful this campaign would be based on the job titles that are in the campaign data and what the copywriting is and then check how successful it actually was and improve your guessing." and it did that over definitely more than a thousand. And so it literally is just guessing back and forth and it knows what makes a campaign successful. This domain setup, if you're using DynaDot and Zapier, there's little edge cases that of literally just vocabulary in the Zapier API that Claude is going to get confused with on how to set up the inboxes. That's all taken away, you can just plug in the skill right away here. And then if you wanted to scrape Google Maps or you wanted to scrape Prospector searches, we have a skill straight up for that as well as some common outbound lists like I have 12 million US businesses that you could just have here. I have a SaaS company list here. I have every zip code in the United States. And you can just point Claude code literally at this repository and get access to all of it.
So, I'm just going to use WhisperFlow and I can just talk to it and I can just say "Hey Claude, using the prospector.io search skill, can you help me build a list of 1,000 marketing leaders in the United States with a employee head count between 10 and 100 and they had recently raised money. Can you help me build that list?" And so now we're just going to shoot that off to Claude code. It's going to read our skill document and then it's going to put that together. So, you'll see it's even proposing the filters for me. CMO, chief marketing officer, everything that I wanted, location, funding, raised money in the last 6 months, all stages only. And so it's even asking me questions, which is great. Let's make the funding recency 180 days. I don't care about the stage and verified emails are required. We can keep that and there's no industry filter. Thank you so much and when you finish this, open the CSV file on my computer so I can show everybody. And now we're just going to let that run and get kicked off.
So now we got our CSV file from the Prospector search and without clicking around, without doing anything, just using my voice and that Prospector export skill, I now have a list of marketing leaders in the United States at all kinds of different companies. Really just in a flash, just like that. My favorite part about using my voice is then also cleaning up the data. Prospector inherently does not give us the email when we're using the search endpoint. We need to use the enrich endpoint. So, we're going to get everybody's email addresses right now. But then what we're also going to do is we're going to confirm that these companies are software companies and I'm going to show you a secret about how we're going to be able to do that. We're going to switch back over to Claude code and I'm just going to tell it, "Hey, thanks so much for pulling this list for me. Can you also use the Prospector enrich endpoint to find all of the email addresses of the contacts in this? And I also know that the enrich endpoint usually comes with the company description as well, too. Can we confirm using a Sonnet sub agent that these companies are actually software or software-as-a-service companies and then filter this list for only marketing leaders that fit this initial criteria, they have a valid email, and that we can confirm that they run a SaaS or a software company? And again, use a Sonnet agent to be able to figure that out.
While that's thinking, the secret sauce that we've been using inside of our agency is normally what we would have to do if we ever wanted to look at a company description and say, "Is this a good fit company or is this a bad fit company? Do they mention this keyword? Are they a good fit contact?" And we wanted to use AI to be able to look at this. We would then say, "Hey, use our open AI key. Here's the prompt. Make this call." But Claude code gives you so much extra usage on the $200 month plan. We are basically circumventing making direct API calls and just using Sonnet sub agents inside of Claude code to do all of our ICP filtering, to write our personalized lines, and that's exactly how we do it. I literally just say, "Prove to me that they are a SaaS company." and use a Sonnet sub agent to do it. It will also probably check with me to make sure that what it thinks is a SaaS company is actually a SaaS company before that runs, but that's the awesome power of using this through Claude code is this is going to be included in my usage to be able to run it like this and I don't have to pay anything extra.
So now let's jump back into our terminal screen and see how it's enriching all of the email addresses and all these things. What I was talking about, classifying batch one as SaaS, classifying batch two as SaaS. This is where it's spinning up those Sonnet agents so that we don't have to pay extra, we don't have any extra token costs, it is not charging me in any other way. I'm using my Claude code usage here to get all of this done. And so it's not done yet, we still need to wait, but this is what I'm talking about where it can do the scoring and the filtering for you straight in here so then you don't have to pay extra, you could just use your usage, a part of your plan to be able to do this. All right, now we just got to wait, I guess.
And we're back. Everything ran. Again, I didn't have to touch a single thing. Just with my voice, we enriched all of the contacts for their email addresses. Their phone numbers stayed masked because we didn't pay extra for them. And then what we could see, the most interesting thing is we have the company description, but then we have the output of is it a SaaS company? Is it not a SaaS company? And we have all of the reasoning in here as well, too. And so I didn't pay any extra for this. Like I said, we use Sonnet sub agents to be able to get this done. You could review it, you could then go back and forth with it and you could say, "Oh, I actually prefer that this is considered a SaaS company or this isn't considered a SaaS company." But that's one of my favorite hacks that we do for our ICP filtering is this Sonnet sub agent thing.
The last thing, in my opinion, is the biggest game-changing thing and the thing that we're working on the most right now for our customers. And it is something that is called auto research. And I think this is by far the most game-changing part about Claude code. Auto research is a repository made by Andrej Karpathy, who many people regard as one of the best AI researchers in the world. And this is actually a really simple idea when you think about it, but when you give it to Claude code, its ability to run this process over and over again autonomously is game-changing and we'll talk about that in a second. All this does is it basically is a repository that you can say, "I am going to give you a bunch of context on how my business works or an experiment that I want to run and I want you to optimize for positive response rate or reply rate." Or if you're using this outside of cold email, you want to optimize it for a Google ad bidding strategy. Whatever the number is, you tell it, "I want to optimize for this outcome to happen."
So then what you can do is you could use this repository framework to then keep running experiments without you touching anything that improves over time your Google ad budget, your cold email campaigns, whatever it might be. Without touching anything, you can have an AI optimizing your cold email campaigns, running experiments, and then learning from what works and what doesn't work, so that you don't have to keep running those experiments, right? Because marketing is just a fancy word for testing. And with this auto research repository, you would be able to just plug this in, and without you having to think about anything other than giving it context up front to basically say, "Hey, don't give our services away for free, and here's some other rules of experiments that you can run." It can plug into SmartLead instantly, whatever tool you're using for sending, change the copy, change the lead list, and run all these experiments for you on your behalf.
And so, this is actually a visualization of a offer campaign loop that we've created as well, too. What this version on our repository does is it first pulls the recent results of the campaigns that we have auto research pointed to. We usually give it context, and we say, "Hey, for our customer, this is the boundaries of the experiments that you're allowed to run, and what kind of offers you're allowed to make, and who we're targeting, and all those kind of things." So, first it pulls the recent results, and it says, "Okay, what happened yesterday? What was the most recent experiment? Did it work better than baseline, or was it worse than baseline?" And then it rereads the context file to then think about, "Okay, what are new experiments that we're allowed to run?" And then it changes the list filter, it changes the copy, and then it relaunches the campaigns, all completely autonomously, and it stores the new experiment that it's going to be doing, so that we learn what's the best list filtering for us to use, what's the best copywriting for us to use, and it recursively learns all of this knowledge over and over and over again.
And so, we've already installed two of these auto research repositories for some of our enterprise customers. We've taken 10% of their overall volume, and we're just dedicating it to this auto research thing. For one of the customers, it's starting to outperform the human-written campaigns as well, too. But I really think that this is the most game-changing thing to happen to lead generation in a really long time. And so, if this hasn't convinced you to start using CloudCode for your outbound campaign strategies, I don't know what will. Hopefully, we'll have another video coming out that will convince you, and I'll see you on that one. Thanks, everybody.