Latest / AI DTC WTF / Use Claude and Codex without getting locked into one platform

Transcript

Use Claude and Codex without getting locked into one platform · AI DTC WTF

• 0:03
  Andrew Foxwell: If you work in d to c and you use AI and you're wondering what the f is going on every week, this is your podcast, the AI DTC WTF podcast, your home for tactical tips, strategies, and ideas that you can implement right now in your AI workflows to make your brand or agency more money. Welcome to another episode of AI DTC WTF. And great episode today because we have, you know, Thomas, who's at the beach in Croatia joining on this podcast. Legendary move. And we also have Dustin who is locked in a corporate office at his house.
• 0:38
  No. Was kidding. Who is... It is at his place in in New York. And Dustin, glad to have you on the podcast.
• 0:46
  Dustin Thompson: Yeah. Thanks for
• 0:46
  having me. Excited to chat about some AI today.
• 0:51
  Andrew Foxwell: Yeah. Well, the thing is we're gonna get really tactical today. We're gonna talk really deeply about a whole bunch of stuff. The thing I will mention at the start of this podcast is Dustin is a member of Foxwell Founders and is the person that has been building and retooling
• 1:05
  Thomas Moes: a
• 1:05
  Andrew Foxwell: lot of the processes we have at Foxwell Founders and Foxwell Digital. And so, you know, there's no one better if you need somebody to help, by the way, on building out AI systems processes and thinking about it. And Dustin's been an absolutely incredible partner for us. So I'll just say that from the outset as a word of transparency. If you see me fawning over Dustin, it's because I wanna keep him in my corner.
• 1:27
  But let's go. Thomas has a question right away. So I'm gonna have... Shoot things over to him, and we'll go from there.
• 1:33
  Thomas Moes: Yes. So the last six months, I've seen most agency owners, most advertisers are going deep into Claude and talking about making good, well, schedule, tasks, doing good skills, etcetera, etcetera. But the last few weeks, Codex, ChatGTP for work, the Mac app, a lot of things have happened. So I see a lot of people are transitioning over to that kind of ecosystem instead. And I think that's that's kind of a a problem that we have to jump between different platforms.
• 2:07
  And my question to you, Dustin, what I would like you to guide us through today is how to think about structuring and making sure that you're platform agnostic in terms of your workflows. And and how should we think about these things?
• 2:22
  Dustin Thompson: Yeah. That's a great question. I would say in the last six months, even like since Codex launched from OpenAI, they've been really, I think, outpacing Claude. They're since they dropped, you know, Sora and doing a browser, now they're dropping. Their focus is really on competing with Anthropic at this level of building software that's usable for everyday people, agencies, corporations.
• 2:52
  So if you haven't used Codex yet, and you have only been in the cloud ecosystem, I would say it's definitely worth trying. But to your point of getting locked in, I 100% agree, I use both all day long, back and forth with no issues. And it's really just depending on what one it's managing my subscription usage, so I can spread that out. But also, they just have different features and functionality. So like, in this past week, Claude introduced the ability to record a skill or record work on your computer, and it'll build a skill for you.
• 3:26
  I think Codex had that a month or two months ago. So you can see that the competition, they're really neck and neck right now. And it is worth exploring your options, I would say, being platform agnostic comes down to how you're actually using Claude. So in my case, every skill that I have is shared between Claude and Codex. And in that, also, like, if I use Grok build, for example, I still have access to all my skills.
• 3:57
  So I'm using for coding work, I will use Claude MD, but it it basically is just a SIM link to my agent's MD file. So they're exactly the same. I change it once, and both Claude and Codex would know how I wanna operate and what skills I have available.
• 4:14
  Andrew Foxwell: This episode of AI DTC WTF is brought to you by Runneth from Motion. And, you know, there's so many interesting things about Runneth that I absolutely love. The best marketers I know right now are not sitting there going, how do I get AI to answer my questions? They're thinking in systems. It's, I have taste, I have judgment, I know my customer better than anyone and they're putting that into Runnith and having it work on it in a really sophisticated way.
• 4:39
  It's putting it in there and Runnith is amplifying what they can do with that particular data. It's not a chat bot, right? It's not just Claude, it's not just ChatGPT. That's the thing I love about Runneth is that it pushes it further and it has your ad data, your competitor data, your creative strategy, it has your product data and it helps you understand all of these things together so you can genuinely train it on how you think and it becomes useful from an operational level, saving you time, giving you context, and giving you important information so that you can make the most absolute best decisions as a marketer today to help make your business run more efficiently. Motion can do so much.
• 5:21
  Runneth can do so much on top of it. So I encourage you to give it a shot if you haven't. Runneth.com and go and see what you can build.
• 5:28
  Thomas Moes: So a lot to there, and and I think that's that's super interesting. So so first of all, how... For for someone who isn't kind of familiar with Simulinks and making sure that everything talks together, What does that mean? How how is that structured? Where would you get started if you're...
• 5:49
  Let's say, I'm locked into the cloud, co work, your cloud code ecosystem. Everything is just natively built in there. How should I think about and get things out and create these symlinks and and and and really kind of structure the whole setup?
• 6:05
  Dustin Thompson: Yeah. The great thing that Codex has done is they have already taken care of a lot of that for you. So if you're an everyday co worker user, when you log in to Codex for the first time, it's going to have an onboarding flow that says, hey, do you wanna import your memories from another AI provider, aka Claude? You can import all those memories. When you open up a new project inside of Codex, it's also gonna say, hey.
• 6:28
  We found these skills and memories. Would you like to bring those over? You just click yes, and you're done. So from a non technical perspective, like, they've made it very easy to move to Codex and start using it. But I think everyday usage, it's really about where are your skills stored, making sure that you have a central location where your skills are stored.
• 6:52
  So if you are going back and forth, you know where to point Claude or Codex or whatever you're using to the right skills. And Yes. Skills are basically just text files. So
• 7:03
  Thomas Moes: Yes. Okay. So, again, I'm just digging digging a little bit deeper here. So so having one folder or one place to store your your your skills. And I see...
• 7:14
  I personally, I use GitHub. I've heard people use Dropbox. And kinda what are your recommendations? And and, you know, as you said, skills are just text files, you store them somewhere. But but which kind of setup is the optimal setup you think for for someone like us?
• 7:33
  Dustin Thompson: My preference is going to be GitHub, and I know that could be considered more technical. But even if you are not technical, I would just ask Codex or Claude to say, hey. I'm trying to achieve this goal of having my skills on on GitHub. Help me do that. And Claude or Codex or any agent is very good at using GitHub, and it's gonna be able to walk you through the process of how to put your your skills on GitHub, how to make sure that they are either private or public, depending on what type of skill it is, and how to access them moving forward from inside your harness of Claude, Codex, or whatever you're using.
• 8:09
  Thomas Moes: And and one of the things I really love about, like, GitHub, it is kind of a versioning part. So if we know some junk, there's a way you kind of you you made the phone change. Basically, you can kind of revert. Can you talk a little bit about that history, logging, revision thing on the term?
• 8:27
  Dustin Thompson: Yeah. I would say the versioning is, like, the number one reason why it's worth having skills stored on GitHub versus just Dropbox or a look your local drive. It really... Once you get something that's working, you can commit that and push it to GitHub. And that is now the working addition of the skill.
• 8:45
  And you now have more freedom to go and try to tweak the skill, make version two, make version 10, whatever you're doing. And in the back of your head, you have... You know you have that working skill always available to revert to if you need to. And again, like, if you're not technical, it's not a problem because you're just... You can, in plain English, just explain to Claude that, hey.
• 9:08
  I think I messed up this skill. Can we revert back to this previous version of the skill? And it's gonna do that for you in two minutes.
• 9:15
  Andrew Foxwell: Yeah. One thing I was wondering is is there's a lot of folks that have been getting on sort of software. I'm getting I'm getting pitched now on a lot of people wanting to come on a podcast talking about software that brings all this stuff together for their team. Mhmm. And so we're talking about, you know, you being platform agnostic and, you know, I know that you've shared a situation where it sort of like decides which skill to use where.
• 9:44
  So can you talk about that and then talk about how you get your team on this to make sure that the team is is sharing skills properly? Because I know there's tools like HQ and stuff, but I've heard mixed reviews on that from some folks as well. Like, it's too confusing. I don't get it. It's it's got a really large learning curve.
• 10:04
  And so I'm wondering what you think is the best solution now for those two things.
• 10:09
  Dustin Thompson: Yeah. Like for tools like that, I mean, I haven't tested them to a great degree, I'm a team of one. So it's always hard to measure. But from my perspective, if you can get everyone to the point where they're comfortable using GitHub to get their skills. And Claude and Codex has this built into their team's plans, where you can basically have your private GitHub account synced to your team's account, and you can make available all the skills that your team should be using.
• 10:40
  So that's already built into their systems. And that makes it a lot easier for people to log in to into co work, go look at their skills that are available to them and just click install. They basically click a button, and now they have all those skills. And from your organization admin side, you can go and be like, okay, let's take a look at, like, who is using the skills and what skills are actually being used by the team. So that's kind of a way to look at, are these skills actually useful to my team?
• 11:08
  And if they're not, what do we have to do to improve that? Is the skill broken or if we're taught the process to actually take advantage of this skill? So from my standpoint, like, tools like HQ are are great. I think there is a bit of a learning curve. But you...
• 11:24
  Just out of the box using a free setup on GitHub, I think you can achieve quite a bit in terms of sharing skills and different processes with your team.
• 11:33
  Thomas Moes: I agree. I think that's probably the smoothest way to do it. And and you're not... Again, you're not locked in. Right?
• 11:38
  So I think that makes perfect sense. And and even just as you described, kind of... You just have to, in plain language, explain what you want to do and what's the ideal situation. And I have now... For over 30 people are kinda using my internal skills library in in GitHub, and and then it it works quite well.
• 12:02
  I just wanted to kind of have the... A follow-up question because we were talking about kind of the right way to store your skills. But you mentioned something called Simlinks, which is bit of a technical term. But I think it's really important in terms of how this works through platforms. Can you explain a little bit about Simlinks and what that is and what what people should think about using Simlinks?
• 12:24
  Dustin Thompson: Yeah. It's basically where you're going to have one canonical version of a scale either stored... Usually stored on your computer or a network drive. And the Symblink is just a way to point Claude or Codex at the actual version of the file. So you can just change the file one time.
• 12:45
  You don't have to keep two copies of a file. The Symlink is just pointing your AI tool to the correct version of the file, whether that's stored locally or on OneDrive or wherever you store it. So it's like a very clean way. So you're not actually keeping a copy and paste job of the same file. You have one canonical file, and that's all you need.
• 13:05
  Thomas Moes: Makes sense. Okay. So that's the skills part. So so what about schedules? I...
• 13:09
  I've I've kind of... I'm trying to move away from club co work again. So I had 96 schedules or something going on over there. I moved most of them over to my open claw. But how should they...
• 13:25
  How should people think about scheduling? And and we talked a little bit about this in the start, but where would you recommend focusing on using formerly known as Codex? Because now ChatTP is kinda trying to kill that name a little bit. But from where would you use ChatTP for scheduled work, and when would you use Cloud for scheduled work?
• 13:52
  Andrew Foxwell: If you're listening to this podcast, you've probably heard about the Foxwell Founders community. It's a community of the best digital marketers in the world brought together in Slack and a bunch of other tools that help us communicate, support, and take care of one another. It's got 600 members from 30 countries, and we spend an average of 500,000,000 a month just on Meta. So if you're advertising on Meta, if you're taking control of your own ads, if you're doing stuff on TikTok, Pinterest, Google, you're doing things with CRO, you're doing landing page testing, you need help with pixels, I mean, on and on and on, there's not a better place than the Foxwell Founders membership. Apply today at foxwell founders dot com and let us know why you think you'd
• 14:31
  Dustin Thompson: be a great member. To be honest, I don't really think there's a, like, a type of task, a scheduled task where I would say, oh, Codex is gonna be better at this or chatty PT for work. I'm gonna call it Codex, though. I don't care. But, like, it really comes down to, like like, when I'm working in both these systems, I'm I'm usually having different types of projects in Codex versus Claude.
• 14:54
  Just... And there is no reason. They're both at capability wise, would say the same. So if you are going to have scheduled tasks running, I would say it's just like, is this project already something I'm working on in Claude? Or is this something I'm working on in Codex?
• 15:09
  And run the task from there. And if you do want to switch over because of the access, the tools have, you can basically just start a new chat in Codex or Claude Cloud CoWork and say, hey, I have this scheduled task over on Codex. Can you move that over here? And it will do that very happily.
• 15:26
  Thomas Moes: Yeah. That makes that's perfect sense.
• 15:28
  Andrew Foxwell: Yeah. I think another one that I have is there's, you know, a lot of things that you've been exposed to in different organizations now since you sort of really started leaning into this new business, Dustin. And I'm wondering, you know, what are the common things that you see in agencies or marketers that they're trying to replicate or they're trying to figure out to utilize AI for? I mean, I think ad creative is one that people talk about a lot, but I think you've been in the in the in the weeds and nitty gritty a little bit more of this. And I think a lot, you know, wondering what you see people are trying to accomplish and how you're sort of advising them to get that accomplished.
• 16:05
  Dustin Thompson: Yeah. I mean, I think, like, the ad creative and that type of thing is always like what we see online as something to shoot for. But it's arguably one of the more complicated things to try to automate because there's a lot of moving pieces there. I think, like, reporting is number one, like that is always something that you should look at automating as much as you can. Pulling the data, verifying the data, and then putting together deliverables for that reporting.
• 16:34
  And then other like quick wins would be really just sort of any sort of business process that you are currently manually working on. It's not as sexy as generating creative. But I think overall, it's that's the type of work that is very suitable for scheduled task and AI in general, and it can take a lot off your plate as a agency owner or a business owner.
• 16:58
  Andrew Foxwell: Yeah. I think it's interesting. Like, you know, a lot of people are coming up and I think there's... You know, we're... We just did the the survey for Foxville Condors.
• 17:05
  People are going through and and sharing feedback that there's a lot of overwhelm. Right?
• 17:09
  Thomas Moes: I think
• 17:09
  Andrew Foxwell: in that...
• 17:10
  Thomas Moes: I think
• 17:10
  Andrew Foxwell: that comes I think that's comes in in today's age with what we're doing. You know, I think some of the some of the biggest executions that I've seen that have really been people... Helping people feel less overwhelmed as like daily summary. This is what you need to do today. This is how Yeah.
• 17:27
  These are the things that you still have outstanding and going through that. Do you recommend when people are starting to even just getting getting into the the building of some of these things, just talking to the models and saying, this is what I'm doing. And can you help me figure out a way
• 17:41
  Dustin Thompson: to make this better? A 100%. Like, ChatGPT just released desktop... Voice mode for desktop last night. That's why I'm tired because I didn't get access until 10:30 last night.
• 17:54
  So... But, yeah, like, either using voice mode, using WhisperFlow, or just, you know, typing work still. But just doing kind of a brain dump of, hey, this is my situation. Here's what I do. Here are the tasks I'm doing.
• 18:10
  And having and giving the the AI some some overview of like the goal of what you're trying to achieve. I want to be more productive. I want to be more profitable. I want to generate more creative for my clients. Whatever it is, taking a big block of text like that and starting a conversation and going back and forth for thirty minutes, even an hour, you're going to get a lot of value.
• 18:33
  And throughout that conversation, it's gonna really be able to guide you to be like, here's some options we have to build, and here's the recommended order to actually build those in. So, yeah, a 100%. Just talk talk to your AI more.
• 18:46
  Thomas Moes: That's something I've seen also with the people I'm working with is that they're they're almost afraid of of asking all the double questions. And if you just ask it, it'll it'll help you kinda get started. And one other thing I just realized with a lot of people from my team is that ends up that AI is telling them what to do. And I say, you can always ask the AI to do it for you, and usually, it actually can. Right?
• 19:13
  So they were like, what? They can use the browser? What can they can use? They can organize my files, etcetera. So so so remembering who's the boss, talk to the AI, but also ask the AI to actually do the stuff for you that they recommend is also a big hack.
• 19:27
  I realized that people are not kind of curious enough or hungry enough for get them to actually get the job done.
• 19:35
  Dustin Thompson: Yeah. That's a great point. Like, I think, like, one of the mistakes people are making is if you're working on something, just go ask Claude or Codex. Like, here's what I'm doing. Can you do this for me?
• 19:46
  And usually, the answer will be like, yes, I can. Or yes, I can. But we need these done first. And like you said, usually sometimes the the models get lazy, and they'll be like, yeah, this is how you do that. And I'll just say, well, do it for me.
• 19:59
  And then I'll be like, oh, I should have thought of that here. I'll go do that. So that those are just like two really simple things that I think people just don't think of to do. Percent.
• 20:08
  Andrew Foxwell: Yeah. Where do you see, Dustin, like, going through this as time continues to go on? The tools are evolving so, you know, so rapidly. Where do you see this continuing to evolve? For for example, like, I think that people that have a better strategy of this is what we're trying to do.
• 20:31
  This is what's gonna make us more money. And the AI then helps them get there is better than people that are more led astray is when they're sort of letting a... Or they're going through and doing all these side quests with AI and all this kind stuff. What are the... How do you see this continuing to evolve as time goes on in terms of making businesses more money, being more productive, etcetera?
• 20:50
  Dustin Thompson: Yeah. I mean, I think it gets back to what you just said. Like, you need a strategy. Just winging it just because the models are smart is not necessarily going to make you more revenue or more profitable. You need a strategy of this is what we want to accomplish using AI and then being able to actually track if you are getting that...
• 21:10
  Those productivity gains out of that. If you just go like side quests and just fun projects, definitely worth doing. I do them all the time, just more from a learning perspective. But from a... For your business, really knowing what your goals are and having a strategy is vital, just like anything else.
• 21:28
  Just AI is just a new tool, really. It's not magic. So if you go in blind on how you're going to use it, you're just going to actually probably have more cost and the same productivity or even less because people are getting distracted and spending a lot of money on tokens that... And projects that will never actually help you or be used.
• 21:49
  Thomas Moes: It's gonna be good advice. I think that's that's very important points. Yeah. Cool. I'm I'm excited to see where this this is going.
• 21:57
  As you said, Andrew, it's accelerating a lot. It's been going faster and faster, and, also, the the models are just making bigger and bigger jumps. So it's gonna be interesting to see where we are just a few months from now.
• 22:11
  Dustin Thompson: Yeah. And I I think to, like, that point, I think it's important, like, just because a new model comes out doesn't necessarily mean it's gonna be better for what you're using AI for right now. Like, it will sound cool, and it will probably be more expensive. So you'll want to use it, but it's not necessarily going to, like, getting back to, like, that business strategy. Is it actually going to help you become more profitable or grow grow revenue?
• 22:33
  Maybe. But you need to look at that carefully before you just say, okay, Fable five's out. Let's just you run everything on Fable five because it's smarter.
• 22:41
  Andrew Foxwell: Yeah. It's interesting. Like, I mean, token managements can... You know, one part of it, but it's it's also just understanding, like, where the pros and cons are of what a model can do. How do you keep track of what the model is great at and what it's not?
• 22:58
  Yeah. I mean, obviously, like, you can look at notes, but are there third party sites that you look at that sort of verify, hey, this one does this better, this one does this better. Right? That, you know, to kinda keep up with this as as it goes because, you know, this is something that changes hourly, it feels like.
• 23:17
  Dustin Thompson: Yeah. Definitely. Yeah. It's overwhelming. That's why I said you don't have to be, you know, first to adopt a new model.
• 23:23
  Like, I do look at the benchmark sites, and there's a lot of them out there, but I always take their grades and their scoring with a grain of salt. Like, some models are really good at actually training the model to be good at benchmarks. So that's one thing to keep in mind. And also, just... The benchmarks, you don't know exactly what's going on.
• 23:42
  So like my way to test new models is basically to use them on a task where I know what the expected output should be. And then I can kind of get an idea of how this model is performing. I also love to just read its reasoning and thought trace while it's working to get an idea of like, the model's personality, for lack of a better term, to see like, okay, how is it thinking about this problem? And it's like, obviously, that takes time because you're sitting there actually looking at the model work. But it does give you a good idea of, like, what its thought process is.
• 24:15
  And then I look at the output, I'm like, okay, this is better than what I would have thought or what I've been getting recently from a different model. So that's how I kinda look at it, especially from anything that's creative or involves some level of taste. For for the coding models, I just have evaluations that I run on the same coding project. So every time a new model comes out, I'll just actually just run it through that same set of evals. And then I can just look at the score.
• 24:45
  Be like, okay. It did really well on these certain projects, but not as well here. So, yeah, long answer is it's still just a lot of work to stay up to date on models. And that's why I think it's, like, important for, like, agency owners to not be... They don't have to rush into every new model, kinda let that filter through the community.
• 25:03
  If you have something that's working, just to... It's not not a bad thing to stick with it.
• 25:07
  Thomas Moes: I think that's super good advice. And and talking about kind of overwhelmed and people feeling overwhelmed, and getting so tired of all these kind of AI influencers are always kind of prices maximizing or this is changing everything and everything is... So so so so just actually putting it to work on your projects and and your processes as as a super good advice. And and when... Yeah.
• 25:33
  That that's that's really smart.
• 25:35
  Dustin Thompson: Yeah. My favorite, like, phrase I've found on x is you have to be on the lookout for token dealers. So if Anthropic is pushing a model, are they a token dealer? Yes, they are. So you have to keep that in mind.
• 25:48
  Andrew Foxwell: Dustin, what else you what else you building now that you're pretty stoked about or things that you've done that have been, you know, really inspiring to you recently?
• 25:55
  Dustin Thompson: I mean, obviously, the Foxwell bot, and I'm always tinkering around with that and seeing how we can improve that. On my side, I've been just working a lot on some more advanced skills that just are helping me on my daily, you know, everyday tasks. So whether it's media buying or development, those are like just skills I've been leaning into and automating those skills as much as possible. I also like once a week now I kind of just do a there's a scheduled task on both Codex and Claude that does a complete does data gathering basically on my computers and my repos to look for new content ideas. So I can figure out, okay, this is something I was actually working on, I would have never remembered this.
• 26:39
  But now I can say, okay, that was actually kind of cool. And I should write about that or post about that that might help someone. So that's what I'm working on right now. Nothing super exciting, but trying to just kind of bring this, you know, the same things that we've done for you, Andrew, to other agencies and business owners, helping them use AI to actually be more efficient, instead of just burning tokens for fun.
• 27:05
  Andrew Foxwell: Yeah. I mean, I think the Fox will about, you know, the survey that we just did everybody, there's high percentage of people that said that this is a main part of what they're using Fox will for Fox will found us for is the connection to the to the knowledge is through that. And I think that's a I think that's a huge part of it. Right? So much of it is like that analysis, a quick answer that's got, you know, the context and the database.
• 27:28
  And we talked to recently on an episode to Jess at Fireteam and he was saying that he has a AI copywriter that's better than any human now. But he preempted by going through and saying that actually they have three hours of meetings to brainstorm on the copy be... With AI listening Mhmm. You know, before they go in. So the...
• 27:52
  It's just context. It's just like the amount of information that it has is so rich that... To pull from. And it can hear the nuance in the way that they're talking about it and things like that. So, you know, that's a different ballgame.
• 28:03
  Yeah. Obviously. Yeah. Right. And I think that's part of the reasons why the Foxwell bot and why the Foxwell NCP works so well and so helpful to folks because it's got years of the best media buyers in the world talking about stuff behind closed doors.
• 28:14
  Yeah. And gives you advice. Yeah. I mean,
• 28:17
  Dustin Thompson: that's just, like, generally great advice is context is key to what the outputs you're going to get. So if you don't provide it the right context, you'll probably get Internet average knowledge back. Yeah. So... Because that's what the model has to work on.
• 28:32
  And unless you're okay with being average, then provide more context, and you'll get a better output for sure. One other thing I did... I am working on actually that I forgot is my model orchestrator skill that I use for Claude code. Last weekend, I spent a lot of time testing the same concept in co work with mixed results. So hopefully have something out soon on that.
• 28:54
  But the use of sub agents in Claude co work works, but it... In most cases actually is less token efficient to do it that way than just running one single model.
• 29:06
  Andrew Foxwell: I love it. I love it. Well, Dustin, thanks for joining us. If anybody needs, you know, a great AI guy, we'll put your contact information in the show notes. Thomas, thank you for joining us from the beach in Croatia.
• 29:18
  And we are... Thank you for listening, and we'll be in touch. And let us know if there's anything else you wanna hear. And you're at foxholedigital.com, available to you. Until then, take care and keep it real.