E60: The Claude Code Era: Mastering Autonomous GTM Agents (Jordan Crawford, Founder @ Blueprint)

Jordan Crawford is an advisor to hypergrowth startups like Clay and Fullenrich. He's the founder of Blueprint GTM and one of the most impressive GTM AI practitioners out there.

I have given Claude Code a goal and it will continuously work to accomplish that goal. And by the way, hands-free.

Jordan joins us to explain why we are moving past the ChatGPT and Clay eras and entering the Claude Code era where the constraints on what you can build simply vanish.

You can remove all those things that could be orchestrated at a much better level by building your own tools. You now essentially have have an engineering department for any problem.

But this episode isn't just about tools. It's a warning to every CRO and VP of sales.

If you would like a job, you must understand these tools because all of your organization could be rewritten with these tools. If you're not able to deliver an 80% improvement in a quarter, there's someone that's willing to take your seat that will.

And this is what's most valuable about the conversation. Jordan's at the cutting edge, so he sees what's coming more clearly than most.

It is an alien intelligence that will swallow my job. It will swallow your job, you will be much, much, much better with it, but not if you don't.

Welcome to the Revenue Leadership Podcast. Momentum is a powerful tool for turning sales and customer conversations into go-to-market intelligence. Using GenAI, it extracts, analyzes, and automates customer intelligence across your GTM org. The best part, it's not a new platform for your team to adopt. It integrates seamlessly with your existing stack, or it can straight up replace your conversational intelligence tool. At Owner, we ripped out Gong and went all in on Momentum. It writes back to our Salesforce directly, captures forecast and churn risk, autofills CRM fields, shares product signals, and tracks sentiment. Companies like Cursor, Zscaler, Ramp, and 11Labs use it every day. Check it out with a free trial at momentum.io. All right, today's guest is Jordan Crawford, uh, coming on the podcast for the second time. Was an episode that I got a lot of messages about and people want more from, and I learn from Jordan all the time with his content and through my incessant text messages. For those of you don't know, Jordan is the CEO, founder, chief GTM engineer slash cod— code wizard at Blueprint GTM, his company. And prior to Blueprint, he founded Scout, which was postcard marketing, and he advises a bunch of startups that all of us would have heard. Of Clay Fullenrich Tenner. Uh, he helps me a ton at owner.com as well, uh, and his expertise is really in outbound sales strategies, data-driven GTM operations. He is on the absolute forefront of how AI is reshaping GTM, um, and we are going to dive into what the heck is Cloud Coat because Uh, it seems like we really hit an escape velocity on Cloud Code over the last couple of months. Opus 4.5 plus the Cloud Code update. I don't know whenever, whenever that was, it feels like it hit some tipping point to where the non-technical world has really gotten pulled in. I spent a bunch of my holidays cooking some stuff in Cloud Code that, uh, I'll eventually need your help fixing, but, uh, we're going to unpack like, what is Cloud Code? What are agents? What are subagents? Skills? Like, should CROs be doing this and like spending their time in Clockcode and other tools? And how does it fit into this landscape? This is not a paid episode despite our collective enthusiasm.

I know those bastards. We should send us some money. I'm sure we got plenty of it.

Yeah, I really should have thought of that before we started recording, but I'll pretend we haven't done it yet and I'll send them a message. Um, so what did I miss on your background? Like maybe just not even like your career background, but what type of stuff are you building today? I think it's really valuable for people to get a sense of like, you run a fairly sizable business now. Uh, you have a bunch of customers and are doing like a lot of work and you're a one-man show. It's just, it's just like you and Claude. And so like you give people a sense for the art of the possible on how you can, how much you can. Agents do on your behalf?

Sure, sure. I will say this. I also, in the last year, I have created the Cannonball Substack. So top 40 Substack in under a year because people want to see real work. And that's what you get to see on the Cannonball. So what am I doing now? Well, let me actually— I'm— I know you may hate me for this, but I'm going to dive real deep and I'm going to talk to you about like an actual use case that a customer had. Well, they wanted to basically do 12 unique enrichments, and this was just basically web searches, right? Like if you were to Google it yourself, you could Google it and you would know what to do and what to look for. And so I said, well, what I'd like you to do is take 5 example customers and go do each of these 12 researches yourself on them and tell me what you found. And at every point in the transcript, when you've done with one type of enrichment, just say the word alligator. Um, because, you know, we have to talk to our AI overlords, and, and if we say alligator and not paperclip, they like us better. So, um, the, the nice thing is that, uh, I fed that 2-hour transcript in, uh, uh, in, you know, I pulled the transcript and I said, hey Claude, in this case Claude Code, uh, here is a transcript of my customer doing manual enrichment Can you go help me invent a prompt to go do this reliably and test it on 10 new customers? And so if you sort of think what's happening there, the first thing is it can go and find the transcript locally. So I have it saved to a file, then it can go test that so it can deploy a bunch of agents and then it can evaluate the results. And so because remember, it has like the actual client talking And so if you think about the capabilities here, these might be long-running sessions, Kyle Norton. So this is not try this, then do the next thing, then try this, right? It's like I have given it a goal and it will continuously work to accomplish that goal. And by the way, hands-free. So I just talked to this thing and it keeps going. And so the transition—

As long as you click. Accept all.

No, no, you can, you can skip that. Bypass permissions. Engineers hate when I do this, but you can. There is a dangerously skip permissions flag in Claude that will just— so I, I shouldn't say this publicly because someone will find a way to take advantage of me, but I dangerously skip permissions. So it just, it just rocks. And, you know, I actually, I sort of joke about this, but you hear all these doom— these engineers are like, don't use it, it's dangerous, etc. Everything I'm doing. Yeah.

Yeah.

I, I'm just doing everything locally. So it's just like writing code locally. It's like, you know, calling some services, but it's not like it's on the internet. It's not like it's, you know, I'm not, and also I'm not trying to rebuild salesforce.com. Like we're not, there's a difference between, um, uh, what I'm talking about here, which is just like, I, I want help automating a particular task and like go recreate this 100 million line code of some big SaaS company, right? It's like, that's not what we're talking about. And when you think about this, the capabilities of what a tool like Cloud Code can do, it's as if ChatGPT could take more actions. That's what you should think about it as. Like, if ChatGPT could do more, if it could send emails as you, or— and that's a bad example, but if ChatGPT could—

It can, and mine does.

So yes, yes, yeah, yeah. So But that's like a scary thing. I don't want to scare people.

Can you explain how, like what is, what's happening under the hood in Claude Code that's not happening in Gemini and ChatGPT, for example?

Okay. That's a great question. So basically what it's doing is it can write code. And what the code can do is it then can interact with things like APIs, which are like structured information. It can pull things in and out of its context window. So like for example, Kyle, you have, let's say you have 100,000 transcripts. Well, you can't just paste those in a ChatGPT, right? And let's say you're like, you know what I want to do is I wanna find anytime someone talks about a competitor. Well, what it can do is it can write code, find all the competitor keywords. It can go search online to find your competitors and then say, okay, great. What I'm gonna do is I'm gonna pull just the 5 paragraphs around every, every mention of competitors across all of the transcripts. Well, first, the transcripts are too large to house anywhere except locally, right? Maybe they're hundreds of megabytes, and ChatGPT can take a 250-megabyte file. You also— here's a, here's a quick tip for ChatGPT cheaters here. You can zip up files and upload a ChatGPT, and they'll unzip them. So if you want to get a file that you can't get into ChatGPT, just like make a zip of it and upload a ChatGPT, and ChatGPT will do it for you.

Or just use Gemini with a bigger context window.

Well, it's, no, no, it's not the context window, it's the file size. It's like literally uploading. So if you want to upload it like a zip file of all of your transcripts, you can get more transcripts in. And by the way, what ChatGPT is doing in that case is it is writing code. It's not pasting it into its context window. But so when you use Cloud Code, the benefit is that it can determine what it needs to do. Should I pull this into a subagent? And it sends a bunch of subagents out. Should I pull this into my local context window? Should I write code and go use code to go determine what to do? And so it basically has access to more tools. And by the way, it can sort of run forever. So, um, you might not know this, but Claude has this concept called compacting. So if you're vibe coding for a very long time, what it does is it summarizes the last session and brings it into a new context window. And why this matters to you is that just like you or I, if we're working on a task for a very long time, we get exhausted. Well, if we could take a rewind back to, you know, the time after we had finished our workout or whatever and we had more energy, we would do that.

And so with all of the learnings that you got through that session.

Yeah, yeah, exactly.

So I don't know, we kind of meandered there a little bit, but so the question was like, okay, so like what is Claude Code doing under the hood that the other models aren't really doing? Doing, and it's writing code to accomplish tasks, sometimes building things that you might use repeatedly. And we'll talk about skills and sub-agents. And it has more access to different things that I find their ecosystem is much more open.

Yeah. And I'll say another thing is that the way in which it's designed is more aligned with work than, um, the web-based tools. And by the way, I'm talking about Cloud ChatGPT and Gemini on the web. So like when we talk about these models, like it's probably a misnomer to be like, what's ChatGPT not doing? Well, ChatGPT has a version of Claude code that they call Codex. And so, so, um, I just want to be really clear here. Um, and the thing is that when you pop into ChatGPT, every new session is sort of brand new. Now both of them have concepts of memory, but that's not really how you would work because that exists in a model where it's like, I'm going to take everything you've ever said to me and try to help you with task, which is like unreasonable, right? People don't need to know what you bought your daughter last Christmas, but like your Gemini or ChatGPT has like that history, right? But that's not helpful if you're accomplishing a task. So when you set up Claude locally, you're generally saying you're creating a project, and in this case is a folder, and you're saying in this folder, this is what I'm trying to get done. And in my case, I deliver campaigns. So campaigns have a couple of components, which is like, I need the closed one transcripts. And by the way, the transcripts are different per campaign. So if I pasted other transcripts into historical ChatGPT, like, those may not be relevant for the task at hand. So this is sort of the benefit of doing this stuff in Cloud Code is you're saying this project is for this type of task. In my case, it's campaigns. So all of the context is only relating to that problem. And that can compound, right? That context gets better and better and better.

Yes. And you're controlling it because it's all in the same folder locally on a computer. Yes.

Yeah. And the folder has, by the way, like, I can pull in all the transcripts, say read— so I can have a session that says read through all these transcripts and create a summary of the thing, the reasons I win and the reasons I lose. And then it writes that to a a document. This, they call it— there's a markdown file, which is just a fancy way of saying it's a text file with some formatting in it. And that way, the next time I pop open Cloud Code, it doesn't have to go read all of those transcripts again. It can just read the closed one reasons, the analysis. And so that's where my context compounds, Kyle. So I go from just like unstructured information to structured context, and then I can take that context and accomplish other things with it.

And it's creating files within this local environment that then can be cross-referenced and you can say, oh, look here, don't look there. And so, so I want to zoom out. I want to do a couple more zoom out questions before we get into some details and maybe actually walk people how to do some stuff. Um, what do CROs really need to understand about this? So like, this is a fairly— we're now like fairly out on the sophistication bell curve. And, but you see really forward-thinking go-to-market people, tons of AI-native companies, like sort of living in this, in this like cloud-centric world. But if I'm a CRO at a Series C SaaS company and I really want to bring more AI to my company, like, How deep do I really need to go on this?

Well, I'm kind of of the opinion that— I mean, I can talk about eras here, or what I can do is I can talk about how I think that the role needs to be realigned, which is a better path.

Cool. Let's do it.

Which, which of these paths is the question?

Oh, I thought you said the second one was a better path. Um, let's talk about the eras. I think that's better landscape setting.

Okay.

So we'll probably talk about both. Yeah.

So generally I think of the, um, sort of the go-to-market world in sort of three eras. There's the, the, um, ChatGPT era, the Clay era, and the Claude Code era. Um, and so, uh, and, and these are just like tools that kind of, uh, bookmark where we're at. and they're helpful to think about because they give some context as to the problem that we are able to solve. And as our tools can do more, it's important to know what they can do because then we know what to give them and what not to give them, like what we shouldn't expect from them. And so the ChatGPT era, uh, was defined by, well, now I can take unstructured tasks and I can just accomplish tasks and I can accomplish in my browser. And that might be research a company. It might be something— and it also can be more impressive than this, which is like, you know, write me a Google Sheets formula, or, or take these 5 files and merge them, right? This is like a task.

And, and it's really more about answering questions though, like some tasks, but like basically word tasks, word and, and like data tasks, I would say.

Yeah, yeah, that's right. Yeah. So like you could upload a CSV to it and say categorize this so you wouldn't have to use Excel formulas anymore. You could upload a CSV to ChatGPT and say, what are the reasons I'm losing based on these fields or whatever, right?

Yeah.

And so, yes, you're right. Word-based tasks is probably a pretty good way to say it. The Clay era was more about workflow automation, and these are deterministic workflows. So when this happens, then do X, Y, and Z. And even within a deterministic workflow, you can do non-deterministic things, which is what the models are good at. So you could say, go research and tell me if this is a good fit for me, yes or no. And here's all the criteria I use for a good fit. So, um, a thing that would be really perfect for the sort of the Clay era is, uh, go when a new deal comes into my CRM, run this agent and define these qualifications. And if it is qualified, write yes. And if yes, then go push X, Y, or Z to HubSpot, and that can go trigger an SDR to make a call, for example. So this is like, this is an example. It's a, it's a very defined workflow. And what you would do, and most, I mean, most, I guess most people are probably still in the ChatGPT era. There's like a good number of growing people in the Clay era. And the way in which you would get benefit here is you would stack these kind of brittle tasks. And these brittle tasks And, you know, things could change and you had to keep, you know, so the way in which you got value here is you said, well, these 50 brittle tasks are going to solve a lot for me and this is going to be very valuable to me, but I have to go in and change something. If I change my schema or whatever. And by the way, those workflows don't know me. They don't know what I'm trying to do. They have no understanding about my contacts, they don't know that I'm a CRO. They know when HubSpot comes in, do this thing, run this AI agent, push this here. And this is like, it's, it's really good for observability. You have a very clear understanding about what's happening.

Um, and they can know some stuff, but you have to like put that context in the prompts in the table basically.

Yeah, yeah. In, yeah, exactly. In the prompt. And by the way, when you go create the new thing, you gotta go fetch that context yourself. So you have to go, and by the way, if you want to improve that prompt, you had to copy and paste that into ChatGPT and say, improve it, fix it, fix it, fix it. And actually, it's kind of funny, I just released autoclaygent.com, which creates claygents for you in Claude code. And this is actually a pretty good segue to the sort of Claude code era. So I should give a, we'll put a promo code for a discount or something in the show notes.

Autoclaygent.com, you said?

Autoclaygent.com. Yeah. So, and this, this is a pretty good way to introduce you, and it's really, uh, it's a gateway drug so people could get into the Claude Code era. So a Claygent is basically Clay's AI agent, and the way in which you would build them, the way in which I used to build them, is I would say to ChatGPT, I'm trying to do X, Y, or Z, help me. And then I would say, here are some examples, go improve the prompt And then I would paste that into Clay and then I would run that Clay agent.

And it would fail and you would go back and you fix it and it would fail.

Yeah. Or even if it didn't fail, what I would do is there's a little button in Clay that says copy all the JSON. And what that is, is all of the actual steps that it took and the outcome. And I would feed that back into ChatGPT and say, how did it do? And it said, well, it did okay. Here's where it changed. Here's where it could have done better. And so the autoclägent tool is sort of a perfect example of where I think this Claude Code era can take us. And how the tool works is you get Claude Code set up, and it's really a lot easier than you think, and you just say go. And what it does is it says, okay, Kyle, well, what do you want to do? It's like, well, I'd like to qualify accounts. It's like, okay, well, do you have examples of qualified or disqualified accounts? And you can just dump a bunch of information in the folder. Here's things you could talk to it. And it says, great.

I downloaded a, I downloaded a whole report from Salesforce and then it's in this folder of like things you should know folder.

Yeah. Yeah. And it's like, and maybe you just have like, look, I don't even know what you're like. I don't even know what qualified or disqualified looks like, but here's, here's accounts we lost. Here's accounts we won. You figure it out. And then Claude will go say, okay, well, let me learn about owner.com and then let me have an understanding about like why they lost. Let's like merge that with the transcripts. It's like, okay, great. I have now gathered all the context I need. I'm going to ask you questions, Kyle Norton. I think this is the case. You're like, well, that's not really true. This changed, this changed, and you could talk to it.

And then what is that mostly when you're in plan mode? Like, is it just to go—

No, this is, this is the tool that I built that, but, but it's a perfect example of what Claude can do because I've set it to be that Claude is prompting you, not the other way around. Mm-hmm. Um, and so it's like, okay, give me this context, give me this context. And, um, and then from that point, um, it will go and it says, okay, click these 2 or 3 things in Clay. And then it will go run 10 agents and send them back to Claude, send all that context back to Claude automatically. So it's like, okay, this is what Clay did. And then Claude can say, how did it do based on all the context I already have? And then Claude can say, well, it didn't do this right. It didn't do this right. It took too many steps here. This isn't the right way to, and then it can do it again. And so you have this loop that can happen without, you know, go off and have a sandwich, Kyle Norton, um, or a salad because I know you're a health nut. Um, but, uh, but like that, that type of, loop means because it has your context, it can make judgments and it can improve itself. And so this is why the cloud code era to me is about making autonomous judgments within very, very narrow lanes. And so this is why, like, in my lens, it's like a campaign is a unit, is a lane, is like sort of a unit of building that I want. And every time I build a campaign, all of those primitives— because I— if you say a fancy word, people love fancy words, so primitive sounds like a real fancy word. Um, so all of those primitives of all the campaigns, every time I improve the concept of the campaign, it gets better for the next campaign. So it's like—

because this is going and writing that back to the skill or the local repo, right?

So it's like this, or the folder, the repo repo is a word that engineers get paid $500,000 a year. Regular people call it folder. The only difference between a regular person and engineer is— and $500,000 a year— is the word repo. So it's just a folder. People call it—

that might actually be true now that we all have Cloud Code.

Yeah, yeah, exactly. It's like, uh, there's the— there's an old joke which is like, um, uh, like the difference between a great engineer and like a bad engineer is like knowing what questions to ask on Stack Overflow or something. And it's like, it's like kind of true. Yeah. So this is the cloud code era.

So campaign is a unit of building. The cloud code era is about like recursive improvements, more autonomy. And what— anything else that like defines anything?

Yeah. And this is why, you know, you sort of ask what is the CRO of the future? Well, if you were to understand the task of maybe an SDR or something, or even, you know, we talked about this before the call, but like even deal review. Right? Like, that's a task that you probably have a heuristic that you as a human, you're like, look, these are the 40 things that people do wrong, like, very consistently. And so, and every day what you do is you like go into Gong, you like ask for the transit audio or whatever you're using, Momentum, um, and you say, okay, great, do X, Y, or Z. You look at that and the, the, your context on that particular task like isn't crazy. You're like, these are the standard things that I look for. And you just— and if you could do it every day, you know, you might have better pipeline or whatever. You could, you could offload that task using a tool like Cloud Code because it can connect into your systems. You build the context that you want. You say, pull this information, use this skill, evaluate according to my rubric. And then when you're done, push to an email or push to some other sort of tool. And, and by the way, if it fails, if it failed in the sort of Clay era, you'd have to go back like, oh, what did I do on the table? What is it? And you'd have to like muck with the user experience. You got to click a bunch of things and like, look, Clay's a beautiful tool, but it just is not designed to do it like this. But with Cloud Code, you could basically say everything is totally screwing up. Here's the 10 things that it's doing wrong. Can you go rerun these 10 things and check to see if it's doing right according to my rubric? Which is the crazy thing to be able to talk and have it do those things. And Cloud Code can.

Just explain for folks, like, what is it doing? Like, I think it's important for people to understand sort of like, what are the units of action that's happening? Because you're like, Cloud Code just does this thing and it seems like this weird magic box. But just explain what's actually happening under the hood.

Yeah. And actually a good way to talk about this is the new Cloud Cowork. Which is available on the $100 and $200 a month plans, because what Claude is doing is— oh, sorry, what Anthropic is doing— Anthropic is the company Claude's— what Anthropic is doing is that they're saying the world will look more like people who use Claude Code, but not in Claude Code. So they're trying to, like, bridge these two worlds of, like, you just can go into chat and it has all the power of Claude Code, but you, you as a user won't know the difference eventually. And that's kind of what Cowork is intended to do. And so, like, for example, one of the things that it— that Cowork can do, and we'll, we'll sort of go into a deeper thing about how Claude Code works, but— and Cowork is basically access to a lot of Claude Code-like things in a— in just using the desktop version of the Claude chat app.

So it's the desktop app in Mac. But it's basically Claude code because normally you access Claude code in your terminal and we can fire up a terminal and show people how to like install it and whatever. But this is like a more friendly interface to Claude code because the terminal takes some getting used to. It's like a CLI.

Not quite, but it's getting there is my point. You can see these worlds merging. There's a lot, like it can't write code. So, which is like actually kind of important to this. CoWork can't? Uh, no, no, no. It's like sub-agents, professional outputs, long-running tasks. Like, I don't think you can write code. I mean, this was released at 5 PM yesterday according to Windward, so it's like I haven't, I haven't played with it, but everything I've read is like it can't actually write code. Um, so, but, but let's talk about some of the things that it can do that are available in Cloud Code, and then we'll talk about what Cloud Code can do now that Cowork can't, and I think eventually they will merge. So creating a plan, this is really important, right? Sometimes you have tasks, right? And everything in the past was just a prompt. But now there's like kind of these two concepts like plan and do. And in plan mode, what it can do is basically says, okay, I, I need to— this is a complex task that you've asked me. I need to go figure out all of the components of the complex task and break them out into subtasks. And create a, you know, generally these are like 1 to 2 page docs that are like, here, all the, here's the 78 things I'm going to do. And the nice thing about this is they have these like subtasks or task-based agents. So it says, great, I can go send off. Basically, you could think about this as like another ChatGPT window, another Claude window that can go do this other thing. And when it's done, bring it back to me, right? And so there's these other pieces that it can go do autonomously. And this is like kind of invisible to you, but it just means that it can do more things. And so that's really, really helpful because now what would take you 5 different chats to do, Claude can do it in this sort of plan mode and then deploy those. You could call them chats or just independent agents if you want. Subagents. Yeah, is the right word. And then bring that back to the main agent that coordinates all of that, right?

So you have this like kind of army of tiny little ChatGPTs that are off and can do things for you and are connected to more things and can run these local commands, these like bash commands that can do things and actually like take— make work happen.

Yeah, I think take actions like search, search the web, you know, and we haven't talked about this, but Cloud Code has a Chrome extension too that can literally open up a browser window basically what it's doing is taking a screenshot every time and loading it back in the cloud. It's like, where do I click now? Where do I click now? Where do I click now? And so you can imagine that if you, you're like, I wanted to go shop on Whole Foods and I wanted to go do X, Y, or Z, it can go take screenshots and make those. And this is a silly example, but, or, you know, load a Salesforce report while you're logged in. There's a better way to do that, of course. But, but just to give you a concept that it has the notion of click-based work too that it can go execute on your behalf.

Still a little bit work in progress, but—

Oh, totally. Yeah. It's not there yet. I mean, yeah.

And so I think the important concept here for people to understand, because one of the things that was like a little tricky for me is trying to be like, wait, so if I'm building an agent, does that mean I'm coding an app? And then I'm like building myself other apps that this one app can can click into. And that made it seem scarier and bigger. And I think when I like tried a couple projects, I think the easiest way to think about it is like Cloud Code is basically an agent. So like Cloud Code in your terminal is the Cloud Code product was supposed to be like a place to go help you write code. And then a whole bunch of people were like, oh, like you can sort of do anything. You can just tell it to do random stuff and it will go figure out how to do these non-coding tasks. And so you can follow, if you want, if you want people that like have gone down this path, Alex Finn, F-I-N-N, or Peter Yang are like good YouTube follows of, um, that, that are like in this in-between zone of technical people that are using cloud code to build non-technical solutions and like building stuff for themselves. And so like Claude Code, you could think of like, this is your master agent. I can just like give this master agent a bunch of instructions. I can talk to it in a really unstructured manner and it knows how to use plan mode and think through a complex task and then build these sub-agents to go like spin off and do other work. And so then, and like skills are like a thing that Claude Code as your like master agent could go do. And I know I'm now like bastardizing the use of the word agent. And if like engineers listen to this, they'll be like, that's not what it is. And I, and I, but I think conceptually, especially for my audience, the CRO crowd, just think of Claude Code as your agent. And like you can just speak things in and it will go do things. And the method of it doing things is either running and writing code to like break down a task. So it could be running code to analyze data or running code to move a document or running code to build a file in your folder that then it can reference later. And then eventually, and what we'll talk about maybe right next, or if we've got more stuff on this, like like, uh, set up building tasks, which are basically like these sub-agents or tools you want to give to your master Claude agent. So I think that's my, my non-technical guy version of trying to explain it, even though I know I'm like now— yeah, just I'm disappointing many— anybody who's—

well, well, I mean, well, I mean, the best way for me to talk about this is like in a client deliverable, and the, the best like unit is to think about this as a spreadsheet. And when I generate campaigns, I, I generate a spreadsheet. And in that spreadsheet, there are deterministic things like email. So I'll just use a tool to go get email, right? There's like nothing, nothing earth-shattering there. But, but I may want to write a message, right? And like, oh boy, a message is a hard thing to write programmatically, right? And, and so, you know, in, in the way that it looks like today, it's like, well, I know your title and I know your company, so I I have all the personalization I need to write the worst message I could ever receive. But imagine if that row had a bunch of context, not just about you, but about the situation that you were in. And in this case, like, that row— like, I have a company that I work for, a lot of companies that do in the healthcare space, and there's a lot of public healthcare data. And so when I say, okay, I want anyone that meets this criteria, Well, it can query a tool called Mimi Data Labs, which I love, which is just like a database of all the public health care information. It can pull all the doctors that have X, Y, or Z challenges, you know, that maybe they're— they don't have certification or whatever. And then it can say, okay, great, I've looked at 100 of these and based on your value prop, they fit in these 5 buckets, right? And that's wild to think about. Right, is that it can— and here's a, here's another example of like kind of you could think about what the value of this. I had an opinion, I was like, okay, well, I think that negative Google reviews are a good way to go figure out what we should pitch them basically. What, which, are they a good fit for what we do? And if so, which of our value props are they a good fit for? And so I pulled like some ungodly number of Google reviews, like 200,000 Google reviews, and Claude downloaded them all locally. And by the way, think about how you do this in the pre-world. You just— what are you going to paste those into ChatGPT? Like, ChatGPT can't even probably access all the Google reviews, right, for whatever reason. So I had it call an API, and don't worry about that. I pulled all that information locally, and then it wrote code to analyze, because the reviews are just— they wouldn't fit in any context window, it's too big. And they would say, well, we looked for, we fed a little bit into the context window and we looked for these keywords. And it turns out that only in 2% of cases are there any good, valuable, useful things in the negative reviews for this problem. And it's like, I would have never been able to do that before because you couldn't get access to the data. You couldn't have a way to crunch it. It didn't have, and basically every time it makes a choice, Kyle, it takes my context into mind. So it's like, oh. Okay, well, this is a negative view and it says I left hungry. It's like, okay, well, how does that relate to my product? So we probably don't want the word hungry, you know, and it's making those decisions when it's writing the code. It's like, okay, great, these are the keywords. And so because it has your judgment and it knows what you're trying to do, it can take many more steps on your behalf doing things. And instead of it going off, you know, way far left field, which we've all had that happen in ChatGPT, It's like, nope, that's not what I'm doing here. And that's kind of the benefit of, of Claude Code. Yeah.

So I want to go back to this notion of the CRO of the future. And so I think you said something interesting. You need to be able to see tasks and then eventually like translate that into agents or, or skills. And so just say more about how you think a CRO— so now you have this new sort of magic set of capabilities in this super agent known as Claude Code. So what does a CRO need to know about that and what action should they take to figure out like how to take advantage of this skill set for their— this capability for their team?

Yeah. So if you think about the way we built jobs for people, not for robots. And so the way in which a job description is written is what is a unit of work that a human can do. Um, and we're not writing jobs for agents yet. We haven't created a, I'd like to hire an agent to do X, Y, or Z. And one of the reasons we aren't doing that is we actually don't know what they are capable of. Like, most people don't really understand what they can do and they can't do. And so it becomes a really hard thing. It's like, okay, well, Can an AI agent call my customers? Like, you probably don't want it to do that. Like, could it do that? Like, yeah, sure, it could probably make phone calls and it could like— but that is both capability and judgment together, right? But to be able to write that job description, you can't just say, hey, Kyle, why don't you send me over your SDR job description? I'll just use that. I'll make some tweaks and post it because all of the capabilities are sort of brand new and every 6 months or so they make some big leap. And so this is why it's important for you to know what your people are doing and at least the capabilities of those tools, because it's your job to extract the tasks out of your team that are perfect for robots and automate those. And so you need to be able to understand at the individual person level, what are you doing? What are you spending your time on? And the nice thing is these two things blend together, because imagine a world where, Kyle, you say, Okay, I've spent time with you. I kind of understand what you're doing. And I also know Cloud Code. So you could say, hey, Jordan, why don't you record your screen all day and just send me the recording? Well, you could use Cloud Code to even help you figure out what you can automate. So this is why these two things go together is like now because you have lived in the weeds and you sort of understand what they're doing and you played with Cloud Code to help say, how can I automate some of what the team is doing? Um, then you could also say, okay, well, because I know these two things, this is a virtuous loop. You can actually use the tools to get better at helping your team improve, and that just like skyrockets productivity, um, because you're systematically pulling out both the tasks that take up the most amount of time and that this master agent is capable of doing. And knowing the intersection of those two things is not easy.

And every time— one of the other mental models that I've been sharing with people is like, once you spend the time to extract that task, and I feel like to figure out what an agent can do, you want to go down to like the smallest unit possible of that work. Like, what is the prime number? Cannot be divided by anything else. And and have an agent do that thing. So it's not like, oh, do all of the, do all of the data manipulation for a BDR. That's too like broad, but you can build one agent to do their pre-call prep or like the, the lead enrichment, then this, a different one for scoring, then a different one to prepare their opener and then their hypothesis of need. And it can, you can break these things down. But once you build that thing, it's like you have unlimited effort and unlimited supply of that job now. And so if you can figure out a way to break down and be like, okay, you know, like one of the jobs that my BDRs do is they look through this report and they figure out who fits this sort of criteria and then they like pull them into another thing and do it. Okay, well now you can have an agent do that same job for you and you could get unlimited supply of that thing. And so do it across every single record in your entire TAM. And I find like that's the, you have to go down to this really small scale to then like scale that task now infinitely. And you just go piece by piece through your revenue function and you continue to knock down those barriers. We ran a pilot last week. I won't give away too much on this because it's a pretty nice competitive advantage, but we ran a pilot last week to better prepare reps to make calls and reduce the amount of time in between calls. And this pilot group made 85% more calls and produced 85% more opportunities than their baseline because there was like a couple tasks and we've got this like super sick GTM AI lead and a really good BizOps team and the these two guys, two of those guys were here sitting like right next to the, to the XDRs, the four XDRs who were in this pilot group. And, and it's just like, okay, we, there's this task that BDRs were doing, and now we've like figured out a way to fully do that with an agent. And now we've scaled that infinitely across every single, across every single customer that's in our, that's in our database basically. And so I feel like that is the different, that is like the different mentality that you need to have as a, as a CRO in this world, because the companies who can figure that out, like, okay, this was 2 weeks of work. And we're rolling this to every BDR next week, as soon as we've like finished the enablement stuff. There's a potential that like our, our volume per BDR goes up like 50 to 70 to 80%, like with 3 weeks of work. It's just like the, the outcomes are crazy. And I think the CRO just needs to understand how to build the team around him and point people to the opportunities and understand the art of the possible.

Yeah, yeah. It's not just the art of the possible. I think the other things is really defining the most leverage atomic task. And that's really, really key, which is that because your job is actually— you're like, so what I do is I go here and then I determine if they're a good fit. You're like, well, what does that mean? And it's like, well, a good fit means X, Y, or Z. And it's like, okay, well, when you said Z, what do you do? It's like, okay, well, I do X, Y, or Z.

And, you know, I look them up on this website and then I like do this other thing. Yeah.

Yeah. Or like I, the first thing I do is I go and look to see the 3 nearest customers. Right. And that's a very easy thing for AI to do. It's a hard thing if you say, well, what I do is when I'm talking to the customer, I get a sense of their tone. and then I determine how, if their tone is such that I should ask for their credit card. It's like, okay, well, AI is not going to be able to do that. So it's like, don't have AI do that. Like, tone analysis is like something that, you know, the tools don't have the capability to do, but everything else around that, um, and you have to be able to divide these things into, like, as, as deterministic choices as possible. And the agent will do really, really well with that. And it can take non-deterministic data, it can take the website or whatever, right? Um, but that's, I think the CRO of the future is really going to be able to need to say, not only do I understand the sheer number of tasks that are happening in my team, I have categorized those tasks based on time suck and agent capability. Because I know agent capability. Yeah, value. Exactly. Yeah. So time suck is the inverse of value here, which is like, just like, I'm assuming you're removing time from people. People. But if you can do that, then you can say, I know the tool can do that, and you have a team around you that help. And the training piece you just kind of glossed over, but it's like you have— you're gonna have to start using the output of these tools. And if you don't have an understanding what the, the actual end user of these things, you can't improve it. And this is like a death knell for a system because the second— this happens all the time, right? It's like when, when your ground, your, your frontline team starts to distrust the data, like you've lost. Right? And if they're not part of the process, and so they are the, the CRO of the future will be able to build these loops that understand both agent capability and also the, the leverage that they can get from their team.

Yeah, that I think is a great call to action for CROs. Because this is a question I get all the time. It's like, well, like, dude, like, I can't, I don't want to go as deep as you, or like, I don't have the time. I hear that all the time. And And I don't have a great answer to be like, well, this is like the minimum effective dosage. And I think the minimum effective dosage and now, you know, we talked about, uh, name drop or like, uh, name dropping the word repo. This is the product management equivalent. You have to develop some taste and, uh, you have to develop enough taste and intuition for what models can and can't do. That you can properly understand how to orchestrate resources and get an idea. It's like, okay, like an agent could probably do that really well. It's repeatable. There's a pattern. I have examples to feed it. It's like, it doesn't need to talk to 10 tools. It needs to talk to 2 or 3 tools. Like that's probably a good, that's probably a good candidate. And I do think like my, my, my stance on this has definitely evolved. I think you just need expertise now. You know, like I see, I've seen how rapidly we have advanced since having dedicated applied AI people in our go-to-market organization. And you need somebody on the team that can do what you do, which is like translate all of these things to actual solutions. Because I think like now things are so technical. Um, that without it you can really flounder unless you didn't.

Yeah, I'm just— I'm with you here. And what I— what I'll tell my clients is that my job is to fire myself. I'm trying to get the repo going for you. I'm trying to get the folder with all your context. Um, we're gonna do it 1 or 2 or 3 times, then you're gonna start doing it. And I have, um, I've got a client and, um, this person is amazing and she's just like So on top of it, she's like, Jordan, I wrote the 47 steps that we have to take to launch a campaign. I was like, God bless you. And I just like dropped that into Claude. And I said, every time we launch a campaign, we should do these 47 steps. And Claude's like, got it. I'll commit that to memory.

Yeah. And, and that becomes a skill or that just goes into your markdown.

This is just a markdown file. It's like, it's just like, so it's like, hey, so every time Claude boots up, I say I'm trying to run a campaign. It's like, okay, great. I got to pull in all of this intelligence, um, uh, on how to do this, all the, the checkboxes, et cetera, and go run them. So essentially becomes a plan. Now, in this case, a human made it. And by the way, now this person is in Cloud Code and she, and she'll come to me and, you know, I don't know, 2 weeks ago she came to me, said, can you make this change? I said, nope, but you can make this change. And so like, you have access to repo. Whenever I work, I commit it to the cloud. You pull it down from GitHub. And if you get some feedback from leadership, you just say, change this copy. It writes code, it updates all of those things. And to do this in Clay, you'd have to be like, change the prompt, da da da. But Claude can not only do it, it knows what you're trying to do so it can go check it, right? And so you ask kind of, you know, it's funny that you say, well, Sierra's like, I don't want to get that deep. I'm like, I'm sorry, you, if you want a job, like, like that's how I feel about it is like, If you would like a job, you must understand these tools because all of your organization could be rewritten with these tools. And that doesn't mean it could be changed, but, but certainly the whole playbook could be rewritten. And if you're not able to deliver an 80% improvement in a quarter, like there's someone that's willing to take your seat.

2 weeks.

Yeah. Well, I mean, but I'm, I'm setting a reasonable bar so people don't get— but you know, how long have you been doing this?

It's an example. Yeah.

Yeah, you've been doing this for a long time. And by the way, your 2 weeks maybe a year ago was a quarter, was 2 quarters. And that's not just a tool capability, that is your understanding of the tool capabilities, understanding training, and greasing the wheels. So not only are you getting better and improving, you're getting, you're getting better at improving and proving, which is like a crazy thing to think about.

Yeah, yeah, yeah. And, and Like, I think that it's the competitive pressure that's there because if you are competing with a company that is building an actual like AI native GTM that is, that has agents doing all of this work that humans used to do and it's just like so much more efficient, you're going to lose. And that lesson is tough. But okay. So I want to move to like some more practical stuff to teach people about like, all right, You, Kyle and Jordan, you have officially convinced me I'm going to go like set up Cloud Code and I'm going to try to build something. It doesn't matter what it is. And I would encourage people to pick something like really small and predefined and only connects to like one system and just like get some practice reps and learn how to like deal with the terminal because you can't like, you can't even like copy the, you can't even like highlight what you're typing. This is like a, it's just like a clunky, goofy, it's a goofy thing.

Yeah.

It's like, it's a goofy interface. But I think the two, so we've talked a lot about, okay, Claude is like writing code and executing things and it's building files and building skills, which are like things that are going to happen over and over in a defined way. But just explain like, to be successful, two of the most important inputs are context and integration. So access to other things. And so maybe just explain what context is. And why it matters so much. And then we can talk about like how to build useful context.

Yeah. And the context is really deeply tied into the thing that you are trying to do. And so in my case, all I'm selling is the ability to, to invent a campaign in your head and by the end of the day have everything ready for that campaign, all the copy and all the leads.

Zoom out from that to like one level higher of abstraction so that it's like not as campaign-centric. We can refer back to it, but like, because people will want to use Claude to help them with like forecasting and deal analysis. Like there's a bunch of other things. So let's just like maybe set a broader context foundation.

Well, yeah. So in that case, what you would do, the generic steps that you would take are like, what is context first?

Just explain that.

Yeah, basically context and any given problem that you're trying to solve are very closely related. And this is everything. If you sat someone down and said your job is to do deal analysis and you said this is absolutely everything you need to know to be able to do that right, that's like context for that task. And so, and this is really much harder than you think, because if I ask you a question like, like, how do you be a good parent? Like, and how do you feel? That's like that question you just couldn't— I could ask you until you're blue in the face, that you would always remember new things about that. You're like, oh yeah, yeah, don't let them drink spoiled milk. I forgot, I didn't, I didn't specify that, right? Like, that's a much harder thing to do. But if you said, um, like, what's the best way to, um, uh, you know, to read a revenue book or something, right? That's like, this is like a lot easier. It's like, well, first you skim the end, you go backwards, and blah blah. And that's something that you could legitimately provide an agent context on because the sheer amount of information that you need to be good at that, that well-defined thing is low. And so this is what you're trying to do. You're trying to find a unit of work that isn't— doesn't take like all of everything that you are to do.

So yeah, then what you're trying to do is let's use like close one analysis. Okay, perfect. Close one analysis. Well, You can't just like have the calls and be able to do it.

Yeah. So, so, and you're going to have to fill in some of my context here because I don't do this often. But what I would do in this case is pull all of the transcripts of closed, lost, closed, won, and pull two CSVs. And this is why it's like important to structure this context. So Claude has— it doesn't have to go do it for you. It's like this is a CSV of all of the accounts that are closed, lost. This is a CSV of all the accounts that are closed, won, an Excel file, if you will. And then here are all of the transcripts just for closed lost and just for closed won. And there's some join key between them so that Claude can connect it up, some Salesforce ID or whatever. Right. This is very helpful. And you've done— and by the way, don't get crazy about this. You can have Claude do just that for you. But it's important to think about that and say, okay, great. Now I have closed won, closed lost. I have the transcripts. And, and those transcripts also have rich information. So it has the, the name of the person that said it, their title there. Yeah. And, and like a good thing, for example, this is why building context is important. One thing I loaded all the transcripts for a client in and I have a, I have a Gong prompt that if you're trying to get stuff out of Gong, I can give you that will do all this in Cloud Code and Clay. And but it grabbed all of the names and I said, go look at all of the names and all the transcripts and identify my people versus the customers. So if you see them more than 3 or 4 times, label them as owner.com staff. And so that's, that's a thing that you would do to build context, right? So now the model knows, okay, well if I see Kyle Norton and Kyle Norton talks about like blah, blah, blah, that's not a customer. So remove his, you can include it, but like remove the analysis about what he said because he's on my team.

Okay.

And so this is why you need to break this up into like You need to just say everything that's in your head, and you can say it to Claude too. And so you say—

and another example, like, is like, what's your sales methodology? So you've got all this unstructured, all this unstructured stuff. It's like, my sales model methodology is Challenger, Force Management, but I don't like these parts of Force. I actually use these parts of GAP instead, and I use Medic as whatever, and blah blah blah. And you just like, okay, that's another important piece of context. If you— if you— it's like if you were training a person to do it. You have an intern, the intern has infinite amount of hours, and they're going to do this task. They're going to read every single transcript and they're going to categorize things. Like, how would you teach this intern who doesn't know anything about your business enough to go do this task? Well, they need to understand your, your ICP and your positioning. Okay, well, you have documents for those, so just like upload those as—

yeah.

As files into this folder.

And, and the nice thing too is you can reverse, you can reverse this so you have to do less mental work, which you can say is, um, and ideally what you do is you just record yourself doing this. Like, you know, get on, get on your own Zoom with just you and record yourself and say everything out loud. I'm making this choice for this reason. I'm making this choice for this reason. Here is why X, Y, or Z. Um, and then you can say like, here, I did this manually for 20 rows. You go run it and then test against these 20 and ask me questions. So you can have Claude try to close your own internal. You're like, I forgot. Well, that doesn't apply because, um, Jessica is an early stage SDR, and so she's ramping, and so we wouldn't do blah, blah, blah, whatever, right? And so, so that's really important. And then you can just work just on testing that context over and over again. And then you can say, okay, great. Now reliably this thing will get, you know, 50 out of 50 right. And now it's ready. And that's the kind of thing. And by the way, you go to sleep and you wake up in the morning and Claude can pick up right where you left off. And that, that's not true. And it doesn't matter how much because it's— remember, it's managing the context. It's just trying to organize the context around the problem that you're trying to solve. And that's really, really, really key because The, the more you can compact that context just to the most useful things, both compact and divide it so it's like only calling the context that it needs, it gets really, really good at doing what you want.

Yeah. And one of the skills that I, um, saw on X and then put into my Claude instance is, um, this idea of like recursive learning. Like when you learn a thing, then like write that back into my like core principles. Like when, when you, when we, when you learn something from me about like a structured way I do things, write that into, I've got like a principles markdown file. So there's like owner, there's my decision-making principles. And then there's a bunch of owner principles in there as well. Our company values, some of our operating methodology. And one of the things to know about context is that you can build it up over time that then you can use across a bunch of other tools. So like you, you, you build, and I don't, this is outside of my depth. Like, should this, a lot of this context list lives in this like master Cloud Markdown, but some of it lives in like these other Markdown files in different folders. And I think mine is actually a bit of a mess, but, um, It is, but like, it doesn't matter. That's the crazy thing sometimes is Claude just figures it out, but you can over time just like keep building onto the context, into the context of your library, of your folders. And so then eventually it can understand, like, I'm trying to build like something to help with decision-making at scale. So I don't want to go to as many meetings. I'm in too many meetings and a bunch of the meetings, like 80% of the meetings, I agree with the decisions being made. I, and I probably didn't need to be there, but 20% of the time I'm like, well, man, I'm glad I'm here because this would have gone in a direction that's probably not good. That's my job as an executive is to like make sure that that doesn't happen. But if I can go to 80% fewer meetings because there's, I have some other tool. And so what I've been doing is I have a bunch of decision-making heuristics that I've written down. And then I had Claude, I was like, hey, I want to build out. So I want to build on top of this document. This is going to be our decision-making principles doc. Go read. And so I MCP'd into Notion where all these call transcripts are. And I'm like, go read all of these call transcripts and tell me about decision-making principles I am using that I have not yet documented. Documented. And it caught a bunch of really interesting stuff. And these are now the principles that it uses. And it works pretty well, but it actually, like, it, like, over-relies on the principles in this document, uh, compared to, like, just general good, like, decision-making.

But yeah, I mean, this is just another way, another way that you might even think about this. And this is why it's really important to use the tools, because you might say Hey, here are 10 meetings where I disagreed vehemently with the, with the thing at hand. And by the way, all the context for those decisions came from these other meetings or whatever, right?

So you can start to mention documents.

Yeah. Yeah. So it's like you can sort of map that out, which is like, look, if it's any of these topics, always escalate to me. These topics are too complex. You're not going to have the context, right? If it's any of these topics, like, here are my tried and true, you know, frameworks that I use. But what I want you to do is, like, these are the 100 meetings that I had where in, in 80 of these meetings I said I have no opinion. 20 of these, create a test case against these, against my heuristic frameworks. Go make, go make independent Kyle choices and then go see if they matched up with real Kyle choices and then tell me what's wrong. And it's like, well, here's why X, Y, or Z. And you can say, okay, great, this is how I close the gap now. That's a, and the beautiful thing about this type of task is that you might realize, Kyle, that this is foolhardy because you're like, I want to outsource my judgment. And you're like, well, that's a big thing to outsource, but, but it might be possible to say, I just want to know with some confidence if someone submits a meeting, like, do I need to attend this meeting? And it might even be send an email that's like, if we are going to talk about any of these topics, I'll join. If it's not about these topics, like, I don't need to be there.

Yeah. My goal is actually not to outsource my judgment, but it's to not go to those meetings, get us, have my, like, it's called Kai, my Slack handle's Kai$, and so this is K-A-I$ is my AI chief of staff that I'm building. And so Kai's job is one of the skills is to is to take the transcript, summarize it into the context, the decisions that were made, and like who— what were the other options for that decisions and that decision, and like what did we come to, and then sort of the action items. And so one job of one of those skills is to pull out all of the action items that I should be aware of. So if there's an action item assigned to somebody on my team, I need to have that in one, and it writes it into this Notion page of stuff I'm monitoring. And it's like, okay, Jordan, you're on my team. You said you were going to do this thing and I just need to be able to see that you need to do that thing and be like, hey Jordan, you said you would do that and it's not moving. And then the other thing—

This is exactly how this would go, by the way. I would commit to do something and be like, you said you would do that thing and I'm using AI to check. On you. I was like, don't use my tools against me.

Yeah. And then eventually Kai would just nudge you and be like, hey, Jordan said he would do this thing.

And the most demeaning, because I built—

so there's a Slack bot now in, uh, in our Slack, which is the Kydala Assign Slack bot, which is, which is basically my way to reach— take my AI chief of staff and reach it into, into Slack to like do things. So like read messages, respond. And so what I want is I want what were the decisions made so I can read. So instead of going to 5 hours of meetings, I can skim 5 hours of meeting summaries with decisions in 20 minutes and be like, cool, cool, cool, cool, cool. Ooh, like that one's no bueno. Like I want to go talk to that person and like explain, you know, a different viewpoint because I find that you know, my boss and I will agree on 80, 90% of stuff, but then the stuff we disagree on, like, we want to be in that meeting to, to like share the other person's viewpoint to like, you know, run it to ground. Or me and my like VP of route ops, we agree on like most stuff, but like, we— he wants to check me when I'm going rogue and doing like, you know, my thing. And I want to make sure that I've supplied him with the right insight. And so there's all these people that like don't need to be in the same meetings all the time, but do need this ambient awareness of the things that they care about and have opinions on. And so like my vision for this thing is I get a summary of all the meetings, all the action items and important decisions, and then it tags me in the ones that are most likely for me to want to pay attention to. And then I'm getting like massive scale on my judgment.

And this is also why understanding context is really helpful, because you, you talked about it, you've talked about something that if you think about the way in which you would deploy what you're doing to the team, you are building something yourself to be able to, um, remove, to be able to like deploy as much of your, your own internal context as possible. And the only thing that you want to do is like where the circles don't overlap, I just want to work in that un-overlapping part, right?

Yeah, exactly.

But the other thing that you're doing in doing that is that you're understanding the edge of what's capable. And then what you're doing is you're, you have the ability to then go inside your organization and start to saying, I want everyone's job to look like this. If the job is either that you're doing something that you don't need to— in your case, it's meetings and judgment, but in other cases, it might be actually like literal work, you know, not meetings. Like, someone has to click on something, or it's like, okay, great. So you know how to do the hardest piece of this. And that means that if you understand all the tasks of your team, you can start to say, I'm going to enable you to either— it might be at your job, remove your thing that only exists at your job. So this problem that you have is like maybe it exists in executive leadership, but it's not— it doesn't exist all the way down, right? Like, yeah, but they have their own version of that. And then you as a CRO need to determine, do I enable my SDR to learn cloud code? And that answer is like, maybe no, but there are some things that we need to implement at the company level to make their jobs easier and to remove the task. But there are, are going to be things that they do in their daily lives that they probably should have their own tools to do for whatever reason. And so that has compound interest because you teach the organization, you do it, you teach the organization how to do it. And most importantly, from an organizational perspective, you can remove all those things that could be orchestrated at a much better level by building your own tools. You now essentially have an engineering department for any problem, which is a wild way to think about it.

Yeah, it is just a different— it is just like a different way to see the world. And, and I think— I don't know if you'd be able to really understand it without getting hands-on the tools and sort of experiencing that one like little win. And, you know, like, I've, I've coded a couple things I tried like Replit and Bolt. I don't actually think I built anything in Lovable. And it just like never got, and it was never, never enough.

Yeah, I know.

Yeah.

I spent, I spent 2 hours and it's like, oh, we can't do that. And I was like, son of a bitch. Yeah. Why would you not tell me that at the beginning?

It was really exciting and I had a lot of fun and I felt like I learned, but like I never shipped a thing that was actually useful.

Yeah.

Yeah. But then. In Cloud Code, especially over the break, I had it like checking my email and like responding to emails. It can finish things.

Cloud Code can finish things.

Yeah, it can actually like take it and do the thing and you're like, wow, okay, like what else can this, like what, and my brain was just like, I don't, we could do that, we could do this.

And so, yes. It's the best way to think about this is it is an alien intelligence that no one has built a great translator for yet. and so you have to build your own. So, and the only way you can do it, and, and the, the alien will have as many conversations with you as possible, and it will try to explain things to you as possible. And so it's not like another, you know, I always sort of joke, um, it's like you could learn Apollo, like one could learn Apollo, like, but this thing is infinite. There's no, there's no end to it. It's not like one day I wake up like, fuck, I'm, I've done. I know, I know, I now know all that I know about AI and I understand everything that it can do and I understand how that relates 100% to me. And so this is why it's like fundamentally different than other technology shifts, because not only is it sort of infinite and also growing at a faster pace, which is like a little scary, but also you need to be able to figure out how you can talk to the thing because It's not just a person. Like, it's not— you can't just talk to it like you would talk to a person. It works in a different way than that. And if you don't do that, well, then what you're saying is that, like, this thing that will swallow the world, it will swallow my job, it will swallow your job, um, uh, because, you know, it's, it's coming for us. Um, you will be much, much, much better with it, but not if you don't just constantly dance with it.

Yeah. Yeah. And I don't know if it's hyperbole to say that like most of our jobs are going to be like building and managing agents. That feels like AI hype.

Well, I mean, I will just say this, like I have designed my company around what AI can do autonomously. So I'm doing the opposite. Which is like usually a company gives you a job and it's like, this is your job, here's everything that your job entails. And I'm like, that sucks. Like, I would rather define my job by what AI can do exceptionally well. And this is what allows me to have leverage in my job, is that I can package it around taking the future to you and include in that package only the things that I know that AI can do exceptionally well. And not the things that it can't. And so like I told a customer the other day, we built a whole, you know, we took 12 prompts and we did all this enrichment and I was like, ship this manually to Clay because it's just going to be a better workflow for you. You'll be able to edit it, you'll be able to push to your systems. You have observability. It's a better tool for this. Claude Code was a much better tool to do the original, write all the prompts to test them, etc. But now the best place for you to deploy them is in Clay. And so, like, yeah, just like through the UX so you can see it. And because I don't want cloud code because I basically have shipped off my understanding about what it's doing, which is a wild thing. It's just like, I don't know, it's like, so an engineer asked me, so like, what framework are you using? I was like, what do you mean by that? Like, like, like, are you talking about like a window? Like, I'm using a double pane framework is what I'm doing for Windows.

Framework. It's called Whisper Flow. Yeah, that's my framework. I yap and it does.

Yeah, yeah. What I told him was, I was like, you check out the repo, you tell me what thing I'm using. But that's, that's the thing though, is that you can, if you have that discernment about what the tool is good and is not good at, you can say, because I'm just not— I don't trust the tool enough to like vibe code and like write to your CRM. Like I'm not, I'm not, I'm not a gambling man. You know, it's like, oh, I'm sorry. You know, I guess, I guess what, like Jason Lemkin like deleted his whole production database because like someone gave him the wrong access token. And it's like, don't do that. Like don't, if you're an engineer, only give me read access. And so these are the types of things where it's like, okay, well, because I don't, in your case, I don't have a whole team that can go validate everything it's doing here. It's like if I develop a net new campaign, brand new, new prospects, I can exclude— I don't have to touch any of your own systems. I'm just structuring data from the world. And then it's so much easier to take that and then inject that output into your workflows. And AI can be amazing at that because I don't have to worry about all the complexities of, you know, like sending an email as me, which, you know, I don't, I don't have AI do the things that you're doing. But that's because I have the luxury of designing my job around what I can do really, really, really well without worrying about what it can destroy.

Yeah. Hence the like artifact of your work is just like a monster spreadsheet.

Yeah. It's like, here's a spreadsheet, like here's the whole campaign and here's how you can create monster spreadsheets like really, really, really quickly. And it's connected to databases and APIs, etc. But at no point am I like, right? Oh, I'll just like write all— change all your customer records. So I have— my business is about deploy, not destroy. And so the way that, the way that I have done that is to avoid even selling things where destroy can, can, can be involved.

That's smart. I want to touch one more topic quickly. So, um, where do people start? So like, I want to start, I want to like learn and sort of like build something for myself. What would you say are like the next steps for a CRO from right now?

Sure. So the first thing I would do is confine your ambitions. Like, just start really, really small. Like, and And that might be something as simple as like, here is a spreadsheet of all of my customers. Can you go search the web on them and tell me what about them that I don't know? I mean, it doesn't matter, right? Just something really, really, really small and just do it in Cloud Code. And so you should— you can— if you Google Cloud Code Bash install script, that's the exact Google result. You can copy and paste that into the terminal. And there's a little thing that said you have to copy and paste echo. It's like just select echo and paste that back in and then type the word Claude, C-L-A-U-D-E. And those are the only 3 steps. And then you're in Claude Code. Um, and, and what happens when you, you paste that first bash command? It goes and talks to the web and installs something locally and it does all the other— and by the way, it might open—

it's setting up Claude Code in your local file. Else.

Yes, exactly. And by the way, it might ask— you might see an installer pop up that is like Xcode Tools, and you'll hit yes to that. And it's like the other tools that you need, Python libraries and that kind of stuff. And then actually the last thing is like create a new folder and type the cd, as in it's change directory, one, you know, cd, and then space, and then drag and drop that new folder into the terminal and then hit Enter and then type Claude. And from that moment out, Claude can do everything. So you can ask it to do things. You can say, here's what I would want to do. Research your own capabilities and tell me how you could do this. And you can say, like, here's— like, what I've done is I've recorded my full day and just talked about it, and I put the whole recording in the transcript locally. Can you help me tell me, like, what Claude could do for me?

And when you say record your whole day, it's not— it can't watch your screen into it. Like, you have to narrate your whole day. You have to narrate, like, just like, now I'm going to the Salesforce report. Yeah, just literally start a Zoom in the background, but it needs voiceover because it's just the transcript.

Yeah, yeah. I mean, it can— Claude Code could write things. Gemini can do this too, actually. It can write things to extract the screens and and to take the audio and it will call Whisper, it'll call OpenAI's transcript thing. And so Cloud Code could do all of that, but this is why it's good to have an understanding of what the tools are capable of because it's like, just use Zoom plus the transcript. Like if Zoom gets the transcript, like you just have saved yourself hours and hours and hours of work. And then just drop that transcript in and say, help me understand what you're capable of and what are the things during the course of my day you could help me with. So like, this is recursive intelligence, right? You can ask it what it can do and it can give you ideas. And you're going to find that the first ideas are probably big and unmanageable and you'll try them and they'll fail and that'll be okay because you're not in danger. Like, you know, don't ask it to delete all the files on your computer. Like, don't do that. Just like be reasonable, right? Like, and as long as you're sort of reasonable and you say, hey, I want to automate this spreadsheet task, like, can you help me? And then you play with it and that's how you learn. And you'll get some paper cuts along the way, but you generally won't have a gushing neck wound.

Yeah, let's hope so.

Yeah, yeah, 100%. I'm not reliable if you have a gushing neck wound.

And, and something you said before is like, aim small, miss small. So like Take on something manageable. Take on something manageable. Don't wire it up to everything with rewrite access. You know, like, you know, because one of the other things we didn't touch on is the importance of integrations. And so like, this is great if you're bringing stuff into Cloud Code. What most of us want is for Cloud Code to go out into our digital world and do stuff. And so Generally what you're going to do is just like be like, hey Claude, I want to set up MCP access for my Notion. And it'll walk you through, like Claude will tell you how to go to notion.com or notion.so/integrations. And it'll tell you to click this thing and then take the API key and paste it into Claude Code. And then you'll like give it the right access and it'll just, it'll guide you through all these things. Okay, cool. Now, I want to set up— now I want you to have access to Salesforce, like, yeah, read access to Salesforce. Yeah, this starts to be like a little sketchy on the info side.

Just give us CSV. Yeah, yeah. Oh yeah, that's right too.

Yeah, like, probably don't do that, but it could, and it would be like, okay, go get the— give me your API token.

And you could also ask another question: is this good for AI? Is this wise for me to do? Can I get in trouble? Is there— well, okay.

And actually, did it end up in the neck wound?

Yeah, yeah, yeah, yeah. But you can ask Claude that. And actually, Claude has a feature called security review. So you can do like /security-review or something, and it will go check to make sure that you haven't done any like ridiculous security thing, and it will go fix the security. Like I did— I'm not going to say the thing I did because probably someone could figure it out, but like I did a silly thing and I was like, oh man. And so I had to go fix that silly thing. But I said, Claude, did I do a silly thing? And it's like, I can confirm that you did. You did a silly thing, like that was bad of you. And I was like, can you help me undo it? It's like, okay, great. Go do these things. And then I ended it.

And you can start with like read access, like just, you know, set up like my Notion has read, right? My Notion has like read-write. And because I can like pull in some documents and I'm like, I actually want to update this document with this other thing, like go do it. And it'll just like rewrite the Notion page. But that can get a little screwy. And so maybe you don't do that right away until you've like built some familiarity and, you know, you like get a feel for the process. So, I have no idea if people are going to think that this is like their favorite episode and they're like, this is awesome. I learned so much. Or if you're like, or they're going to be like, you and Jordan just meandered your way through 90 minutes of nonsense and none of it is helpful.

Yeah, I mean, I always joke that when I give talks on stage that there's two groups in the audience. Those people are like, I have no fucking clue what just happened and like, and I need to learn. And I have no fucking clue what just happened and I don't care to learn. Like, 100% of the audience is still confused and that's okay. Like, you just have to lean into it.

Like, and it can help and everything's on YouTube. You can start with, you can just search like Claude code setup for non-technical people and there's a bunch of stuff there.

Actually, like to even tell you this Meta thing, there's a Gemini button on every YouTube video and you could say, summarize this for me given this context and it'll go do that. Mm-hmm.

Yeah. That Gemini button is like an incredible piece of product. I think I use it all the time. Yeah. Well, this was awesome, man. I appreciate you doing it. And we didn't talk about Cowork as much as maybe we should have, but that could be the other— this is the last thought I'll leave people with is like, maybe a bunch of this is just available in Cowork and you want to start there instead of the Terminal thing. I have only played around with it for like 30 minutes because I couldn't get the workspace to set up on my flight yesterday. Unfortunately, I was losing my mind. But play with Workspace, try the terminal thing, dive into the YouTube. There's a bunch of cool stuff about people using Cloud Code to manage their personal life and just explore. And this will, the exploration, and it'll feel like you're wasting a bunch of time, but you will learn the primitives and sort of the first principles that you need to know how to deploy this in your organization. Because I think it is existential for all of us.

Yeah. And you can make little fun web apps. Like I built playbooks.blueprintgtm.com to like create go-to-market playbooks for folks. 100% vibe-coded. And so it runs everything in the cloud. It runs the whole process and it produces you like a go-to-market engineering playbook that takes like 20 or 30 minutes, but everything is agentically, everything, the website, the design, like Claude's design skill is really good. And so you just kind of start building things that get more and more impressive. Yeah, so do it.

Cool.

Thanks for having me.

All right, thanks. Thank you for listening to the Revenue Leadership Podcast. If you enjoyed it, don't forget to subscribe, and you can find a link in the show notes. And be sure to leave a 5-star review, share it with your network, and please join me next Wednesday for another great conversation.