Ep #159: The AI Tools I Actually Use as a Fractional CMO

Transcript

In this episode, I’m just gonna give you a play-by-play of the AI tools that I currently use as a fractional CMO, and how I think you should be thinking about AI tools. This is gonna be a little tactical, but I just wanna share what’s working right now with you. So let’s dive in.

 Hey, it’s Casey. Welcome back. Let’s just dive into the AI tools that I use right now. I’m on my desktop computer as I record this video, as I record this audio, and if I look, I have, I don’t know, I’m gonna guess 12 Claude Code windows open. And I would say that by and large- I’m really hard-pressed to think of any reason not to use Claude Code with Fable on the $200 a month plan.

To me, it feels like the cheapest thing I pay for that gives me the biggest output. Nothing in my life is as inexpensive as $200 and gives me such value on the other side. That being said, as I record this here on September 8th, just after Labor Day, I think about the weekend and how I spent the whole Labor Day weekend like a nerd playing with the new ChatGPT Astra.

And Astra was able to do some stuff that I was pretty impressed with. I watched the video when it was released on Thursday, just a couple days ago, and it said, “Now we’re AGI. , We reached 99.9% of the AGI benchmark.” We’ll talk about that in a minute. And they said one of the things that you can do is you can build a full circuit board with ChatGPT Astra.

It’s like, oh, interesting, ’cause I was working on a project and I kinda daydreamed about building a circuit board. I actually reached out to someone on Upwork to build a circuit board for me. What I found was that Astra was actually completely capable of doing a very complicated task, which is open up the software called KiCad and build a full circuit board for me, and it did it, and I think it worked.

I think it works. I ran it through all the tests I can run it through before purchasing, the build of it, and it really does seem to work. That’s pretty incredible talent, and it makes me think that Astra is as good or better than Claude Fable 5.1. And if you think about it, like maybe there are small differences between the two.

Maybe Claude could have done it, but I never had the reason to think that Claude could do it, so therefore I never tried it. But on the other side one thing I did notice about Astra was like it was a little bit more succinct when it spoke to me. When I would ask it questions, it wasn’t as verbose as Claude was.

And Claude Fable and certainly Claude Opus 5, they can just be like too chatty, just say too much, it’s kinda like chill, just get to the point. So as I see things side by side, I’m a big Claude Code user, but I think I would be completely happy using Astra right now. So instead of Fable, I’d use Astra.

Instead of Opus, I would use Sol. And if you’re unfamiliar with these models, I think of it like this. I think of like a smart high school student. That’s like your lowest level model. And then you move up from there to a college graduate. Then you move up there to a $300, $300,000 a year software engineer.

Then you move from there to like a top-tier software engineer or something like that, researcher, whatever you need. These are the different levels of kind of intelligence that these AI models have, and your job is to choose the right one for the right job. And honestly, if you’re doing stuff like me, which is like- Researching, building, compiling.

I don’t know why you just don’t use the biggest models. I spend hours every day working with these tools for personal things for business things. And what I find is that I don’t liquidate my credits on $200 a month using Fable. I might run out of Fable credits maybe once a month but then they reset in a couple hours, so it’s not even a big deal.

I find Opus to be great as well. So what I’m saying is that side by side, $200 a month for ChatGPT’s, whatever, $200 a month plan or Claude’s I think are both great purchases. I don’t know why you would need both. I’m sure very technical people like to have both, and there’s a taste that they have between the two and they’re like, “Oh, this one’s better for this thing than that.”

But for me, I’m pretty much just a Claude guy, and I use Opus and the top model Fable kind of exclusively. Next is what’s my interface with it? I think the interface matters. So the interface that I prefer is Claude Code. I think one reason is because I’ve always wanted to be Angelina Jolie from the 1990s movie Hackers, right?

I just wanna be in a terminal window ’cause I think it looks cool. The other part of me is maybe a little bit more pragmatic, and here’s what I think. I think that the terminal is the closest to bare metal that I get. There’s no kind of separation between me and the model. It feels like I’m right there.

When I’m on the Claude app, when I’m on claude.ai, it just doesn’t feel as powerful. And I think that The best model that’s available to anyone right now is Astra or Fable, and the best way to interface with them is as directly plugged into them as possible. Now, outside of like governmental work or maybe universities that might have some early access to models, my view is that using Claude Code through the terminal with the latest, biggest model I’m literally on the cutting edge of compute that’s available, and I’m paying pennies on the dollar of what it costs.

I recorded this in an episode years months ago, and I said that we’re in the $6 Uber ride era, and we are. Do you remember when Uber rides were six bucks, but somehow the person that drove you made 20 bucks on the drive? It’s because some VC was dumping a bunch of money in to keep the prices low so that we use Ubers and Lyfts instead of calling cabs, and now the prices are incredibly expensive.

I just got a trip home from the airport and it was 50 bucks. I feel like it used to be 25 and I think that might be inflation, but it also might be just the shift in pricing because VCs are trying to make money instead of just blindly invest in these ride-sharing companies.

I think that’s where we are right now with these AI models. I think my $200 a month, I’m getting thousands of dollars a month of outcome. If I hired someone for this work, I would be getting, I don’t know, four, minimum $4,000 a month. I’d have to pay a person full-time to do this, but like realistically, I think that this is probably worth $10,000 a month in, in labor that I’m getting.

And I’m able to do things that I never would do otherwise. I’m able to chase curiosities and have a sense of fun. So my take is that you should use, I don’t care if you use Claude Work, if that’s what… Excuse me ChatGPT Work or Claude Code. I want you in the terminal if you can, if you feel comfortable, and if you don’t feel comfortable, could you take a minute and just learn about it and get comfortable?

I really think it’s the better place to go. This is like then able to move my mouse and keyboard and do things for me What are the things that I want it to do for me? We had an issue recently on the CMOX website. This was a weird one. Someone couldn’t visit the website, and I just double-checked everything.

It’s everything looks fine. It’s not a problem. Then another person said, “I can’t go to the website.” It’s like, what happened? What happened was Xfinity, the ISP, had decided that our website, something was wrong about it, and they didn’t wanna send their users to our website. So they gave us a SSL rejection error, but our SSL was fine.

So it took a while, but the AI was able to figure this out and send me to the page to submit the form to ask for a recrawl of the website, which I did, and it fixed everything. That’s a level of, Like value that I struggled to get from people. Like to bring– Like I have a friend in my life, and I would call him and be like, “Hey, Bill, w-what do you think I should do here?

I’m stuck.” And he’d be like, “Oh, one time I heard that this was a thing,” right? But like no other way would I have ever been able to figure out that problem. It’s not like the DNS was wrong. It’s not like when I visited the website it was wrong, and when I did a ping to it. I tried all the basic stuff and it worked.

I never thought that Xfinity, which, Xfinity and Spectrum, which I think amounted to 30% of all website visitors, were getting this SSL error. I never thought that could happen. Little did I know that was actually the problem. The AI figured it out, and it didn’t figure it out just by running an AI search.

It figured it out by moving the mouse and keyboard and testing things and going to different websites and testing things. That’s fundamentally different than how a lot of people use AI. Now, I say this stuff and I think you see me as a bit of a nerd which is fine. But it’s a problem that had to get solved.

And had I not been able to solve it, a third of website visitors would be rejected from going to the website. That’s a terrible outcome. So the only real solution then is to hire a developer for it. But just to be able to ask the AI to go and look at it, and it can suss it out and figure it out, that’s value.

That’s how I see like our work as CMOs. Our work as CMOs is not to be CTOs, but sometimes you don’t have a CTO. So you can go pay for one for some consulting time or you can phone a friend. But as a marketer, you’re like Facebook’s saying that we got 100 website clicks, and that means we should have 98 website visitors, and we have 70.

That’s a massive drop-off. What’s happening? Can’t solve the problem unless a CTO steps in and identifies it. So again, that’s where I think this kind of bare metal connection to the AI of me to a terminal window to Claude Code is just so powerful, and I wouldn’t want it to be separated by using virtual machines through claude.ai or chatgpt.com.

I want to do it on my computer so that it can manipulate my mouse and my keyboard to move things around and search on my behalf. Some other things that I use. I think ChatGPT for fully, like one-shotted images is awesome, especially the new ChatGPT images 2.0, I think is what it’s called. The ability to create high-quality images is nuts.

It’s really nuts. It’s so nice that capability has presented itself at the same time where we’re in this world of Andromeda with Meta, where we have to have rapid creation of new creatives to test. So we need new creatives, and now we have a really great tool for it. And I think ChatGPT is awesome for images, so much so that I think I’ll always pay 20 bucks a month to ChatGPT for images because I always want the ability to generate an image.

I have an idea. I want to try something. I want to see what something would look like. I… ChatGPT is good at that, and maybe Midjourney’s great at it too, but for the 20 bucks a month for ChatGPT, I’m pleased with it. The problem with images though is that they’re one-shotted. It’s like They just appear.

I like Claude Design. So here’s an example. I had ChatGPT generate a brand image for me for just like a fun side project, and it did a great job. I really liked it, and I had it regenerate a couple times, then edited it down for me. But now I have a JPEG that’s of a certain dimension, and it’s like it’s what I want, but it’s not in the file types that I want.

I want an SVG. I want a vector image that I can do things with. So I took it into Claude Design. Now, Claude Design, if you haven’t used it it’s an awesome tool. Claude Design is more reminiscent of Figma, which is like layers of text and shapes and images that you can manipulate. It’s more Photoshop and ChatGPT is just like the image appears.

So with ChatGPT, you can’t edit things without potentially changing the whole image ’cause it has to re-render the whole thing. But in Claude Design, I can go in and I can physically move things around. I think the result is that Claude Design is awesome at the cost of being boring. It’s like a little antiseptic.

It’s a little too clean, and that’s okay for some stuff. But I think starting your idea in ChatGPT and then taking it to Claude, it’s a great way to do it, to finalize it and make actual like vector graphics, PDFs, things that you would take to print. I would do that from Claude Design. Claude Design at the time of recording is atrocious on my phone.

So it is like big screen activity, but I think it does a great job at it. And then there’s folks that say “What about Cowork? Can I use Claude Cowork?” And my answer is “Why would you?” It feels to me like a half measure, and for some folks, they’re just not comfortable in terminal. But Claude Cowork to me just doesn’t feel like the answer.

It feels like a neutered version of the thing that I’m after. And for so many that’s fine. But for me, I want to be on the cutting edge of technology. If I’m gonna use the tool, give me the best version of the tool. I don’t want Haiku, I don’t want Sonnet, right? I just want Fable 5.1 and I wanna burn all my credits.

And I want to use the best compute with the least amount of space between me and the actual like compute. I don’t want these levels of abstraction like UIs and stuff like that. I want it simple and straightforward. I feel like I’m way more powerful that way. So Cowork, not needed. Perplexity. I was super excited when Perplexity came out and I used it for some initial research, and now like it’s fine.

It still is like good for research, but the web browser that they had, I don’t know. I have no interest in that. I wanna use the web browser that I use. I just don’t feel like Perplexity is doing anything for me, so I don’t use it. Sometimes if I want a second opinion, I’ll run through Perplexity. But honestly, if I really want a second opinion, I’ll just ask Claude in a separate Claude code window to abandon all of its preconceived notions and start from scratch and look and see if this was done correctly.

And that tends to work very well. So I don’t need a different model, I just need the same model in a new instance to check itself. And then Copilot I’d be satisfied if I never touched the Microsoft stack. I don’t know about you, but like Microsoft Teams, I was just on a Teams meeting, took me two minutes to get connected, and then my speakers were wrong, and I hate Teams, so I don’t use Copilot.

If you live in the Microsoft ecosystem, maybe it’s great. These models probably have a lot in common, but I don’t stay up to date on Copilot ’cause I don’t care. Same with Gemini. The only time I’ve ever used Gemini in Gmail is when I accidentally click the button, and I know I’m not alone on that one.

So if you like Gemini and you like Google NotebookLM, you think that’s a useful thing, great, use it. But by and large, I just feel like ChatGPT and Claude are going to be kind of the front runners, I think. I don’t know. It… I think we will be surprised with what Google’s able to do and what Microsoft’s able to do.

But as far as true LLMs, I do Claude and Chat. So some thoughts here for you. When you’re playing with these tools and you’re doing stuff I use the $200 a month. I don’t use tokens. Like I have no token utilization. I’m always afraid of it. I’m afraid that I’m gonna use something that’s gonna generate, that, that’s gonna need tokens, and I’m gonna wake up in the morning, I’m gonna have a $10,000 bill.

I just am like, it’s just like a worry. So I don’t use anything with tokens, and it’s never been an issue. I don’t know if I’m doing something different than other people, but I’ve never needed to buy tokens generally speaking. I have a project where we do use tokens, but that’s a single specific project.

It’s well scoped and the tokens are limited so that we can’t have overspend issues. But generally speaking, when you develop things, don’t use tokens. Don’t use the API, ’cause when you use the API, you’re gonna pay the token cost and that can be a runaway cost.

The next is if you’re doing stuff with AI tools, don’t build stuff that you have to maintain. Your job should not be to maintain. Maybe your job is to come up with an MVP or an idea, but it’s not to maintain it over time. So when you think about building with AI, build stuff so that once it’s built.

It’s like software. Build software if you want. Custom software just for you to spin up a document or to build a report or something like that. Like you can do– You can use AI to create software that you push a button and it gives you the outcome that you want. And you can do that without utilizing tokens.

It can just be with like Python scripts. So there’s ways for you to use AI to create appliances that you push a button and a thing happens. It’s really nice. That’s a really good use for us CMOs. So let’s talk about a use for that. With your clients, you might want to summarize certain data. You could have data pipes that like pull the data down and it does something, and then maybe your LLM boots up once a week and it crunches through that, and it gives you a report.

And that report tells you like where decisions need to be made. That’s a pretty cool thing to be able to do. That’s not an expensive thing to build. I would use Claude Code for that So your job here is not to use every AI tool. And honestly, I think it’s to not use AI tools for the sake of AI tools. I think you have to choose what you wanna do and then get it done as efficiently as possible.

And if that is using AI, then so be it. Or if it’s using somebody else or asking a team member for help, so be it. I was thinking about asking a team member. Like I do a standup with my team in the morning and I say, “Hey can someone do this for me? I need this done by this time.” It’s like an hour-long task.

Inevitably, some other team will be like, “Happy to do it. I’ve got these things happening and then I’ll get done, at four o’clock today.” “Cool.” So I had the call at nine o’clock, I talked about it. By five o’clock I get it back to me. Great, right? Took the person one hour. I ended up waiting six, seven hours for that to be delivered to me.

But it’s fine, right? Like that, that’s fine. Now with AI tools, an AI tool that with computer use that would take four hours to do the thing that a human would do in one hour, I could get in four hours instead of waiting seven or eight. And that’s a fundamental shift for how I’m thinking about AI.

Can I just ask the AI to do it? And if I can, can I get a result faster than a human on my team has capacity to get me the answer? And right now is it’s an inefficient process for the AI to do it, yet with its inefficiency it still delivers me a great result. Man, what do I care if I ask it to do a bunch of stuff and then go to sleep and then the next morning I wake up and it’s done?

Like that’s m- magic. So I think that this the AI tools right now that I’m in- interested in are just like, just choose a platform. I don’t care. It doesn’t matter. You say, “I want to use Gemini or Copilot or Claude or ChatGPT,” they’re all acceptable. So much so that I would say that one is– they’re all A students in class and one’s the A+ student this week.

That’s it. And right now as I record this, it’s Astra, it’s the A+ student. Who knows? Maybe next week Gemini takes the space. I’m not gonna switch models because one’s an A+. I’m gonna focus on what’s easy for me and how do I keep without context switching all the time and just keep my life simple.

So I think that’s the approach for you. Just choose a model, choose a platform, get as close to bare metal as you can And then don’t build stuff for the sake of building stuff. Just solve the biggest problems that you have. That’s your goal, solve big problems, and then spend a little bit of your time playing, having fun, trying new stuff.

But by and large, your goal needs to be solve big problems for yourself and for your clients, and not just dork around and chase the next fastest model. It reminds me of every hobby. Are you in a hobby? If you’re in photography, you know that cameras come out with new updates, and man, are they cool.

You wanna sell all your stuff and jump from Nikon to Canon or Canon to Fuji or Fuji to Leica or Leica to Hasselblad. You wanna move because the next coolest iteration and invention or whatever came out on a camera. I remember I was buying a camera and man, it was like two weeks later, Canon came out, I think with their R6, and it’s a SLR camera, you put up to your eye, through the lens.

And your eyeball, as you looked around the field of view, that became the focal point. So if you’re a wedding photographer and you look at someone’s eye, the focus instantly goes to their eye. That’s amazing. That’s so cool. It feels worth leaving the Nikon system that I’m on to jump over to Canon. But I saw that, it’s okay, so then I move to that, and then what’s the next thing?

Nikon’s gonna come up with something, and then you’re gonna jump back and forth and back and forth, and you just lose money and time on it all. Your better bet is just to focus on what you got and mastering it. I think of some of the greatest photos ever, Henri Cartier-Bresson on like a, I guess in his later part of life, like a Leica M3.

That camera was built from the 1950s or ’60s and on. That’s a solid camera. It doesn’t need to be more complex than that or a Leica 3G or whatever. You don’t need to have the latest and greatest model to be successful. You need to solve the right problem. And I think that fundamentally, as long as you’re as close as possible to the raw compute, and you just accept that you have a new tax in your life that’s 200 bucks a month currently, best deal ever, for the, for like the Max Pro whatever plan for the AI platform that you wanna be on.

As long as you ask it smart questions and you don’t put yourself in a position where you have to maintain software, man, you can get so much done right now. I think of, like, all of this software that exists that’s AI one-hit wonders, it does one thing well. Skip it. Skip it.

Just stay with the big models. Work hard on them. Maybe you lose a little bit of efficiency, but you gain it back in lack of context switching, lack of new platforms to learn, and I think it’s gonna make you more money and your clients more money in the long run. All right. As a quick note, We’ve got kind of a binge guide for you.

This is something I wanted to put together. So as you think about becoming a fractional CMO you probably have a lot of questions like how to get your first client, or if you’re an employee, like how do you talk to people without outing that you wanna leave your full-time job, or how do you go from agency owner to a fractional CMO?

Maybe you wanna hear stories of folks. I like took the best of the best content that we have, and I put it together in one place. That’s cmox.co/start. So just head over there, cmox.co/start. It’s like all my best stuff right there ready for you to dive in, and we’re updating it regularly. Probably like once a month we’ll be updating it.

Dive in wherever it feels right for you, and keep going on learning this stuff at cmox.co/start. And when you’re ready, just book in a call with my team and we’ll see if you’ve got the experience and if you’re cool enough to be a fractional CMO with us at CMOX. Go to cmox.co/show and book in your call.

All right? Good to see you. Take care. Bye.