How I Use AI in My Creative Work

Wait, so you have Claude actually editing and exporting videos and they're good? I created something that did that, but the coolest part of that was. Tell me about rock butts! I really like the mental model of this because. Okay, Olena, in this episode of the podcast, I'm really excited because we're talking all about the AI tools that we're using in the creator economy, and I have been itching for like months to make this episode. I think you are the perfect person to bounce AI ideas off of, because in my opinion, you have the best AI-related newsletter in the entire world. I read it every Tuesday, it's called "Me Plus Machine", it's great. But I wanted to start this whole conversation by sharing how I'm using AI to come up with new video ideas, because when it comes to the creator economy, the idea is the thing that makes everything else worth it. Like, it's the thing that has the highest leverage of growth for your YouTube channel, or for your social media profiles, right? So, for me, there's two tools that I use. Well, firstly, I like to work in Clod, because I'm familiar with it, it's been a long time. I know there's an argument right now that OpenAI is getting to become the better one, but I work in Clod, and I use MCPs to connect to basically everything that I do. And the two MCPs that I use to come up with video ideas are VidIQ and Sandcastles. And the way that I do it is I usually go into a Clod window, and I ask VidIQ, I say use the VidIQ MCP, and find me videos that are relevant to my channel, which I represent a lot of different authors and thought leaders behind me. So, I kind of point out the specific channel and what their niche is and what their specialty is. And I say, find me 10 ideas. It's got to be less than two years old. It's got to have at least a million views. And it's got to be at least an outlier score of 1.7x or something along those lines. And then, what I think is cool, this is something you taught me, is instead of delivering the results kind of like line by line as text, I ask it to deliver it as an HTML file in the shape of YouTube's interface. Because when I think it comes to like evaluating packaging, it's really, really important to see it as a thumbnail and see it as a title with the view count and the age of the video. And what I also do is I like to ask for like three different videos that do supporting evidence for that hero video. And so, now I've got like 30 or 40 different videos that help me do a scan of my niche and understand like what's winning, what's losing. And then, for short form, I basically do the same thing except the tool that I use as sandcastles. And sandcastles has come from Kane Calloway and it is emerged as like probably my favorite AI tool right now for the creator economy. With sandcastles, they have little prompts that you can run in your cloud that automatically analyze Instagram videos for you. So you can like grab the URL and you can put it into cloud and say analyze this. How many views it got, the outlier score, the age of it, all these different kinds of things. But that's what I'm doing right now. And it's moving mountains for me. I'm finding videos that I wouldn't be able to find organically. My favorite ones are when I find some video that's like actually spoken in Hindi. And obviously I can't speak Hindi. And you know, there's a lot of like Indian crossover that I just would not have seen. So you're in cloud and then you have cloud connected to the MCPs for sandcastles and vidIQ. And then like, how do you, what's the next step? Where do you take that? Well, I usually bank a bunch of ideas. And then it goes into like the scripting process. And actually use AI pretty heavily for scripting. And so the way that I have it laid out is let's say I find a video with vidIQ. And I want to create something like that. I'll have a good conversation with vidIQ about why that video won. And then I would find my client, let's say Mark Manson for example. And I would try and find ways to make it work with Mark. You know, maybe the title stays the same, but the content is like 80 to 90 percent different. And that's kind of like my starting point. So then it's time to write. And I'll talk to Claude and I'll say, okay, it's time to write. So Claude, when you open a new window, always reads the Claude Markdown file, right? And inside my Claude Markdown file is, it's actually your me plus machine anti-AI list. It's like all the AI tendencies, like it's not X, it's Y, you know, the use of M dashes, the word quietly appearing in the script a thousand times. So it has all of those in there. But when I ask you to write a script, it knows to read layer one, which is the Claude Markdown file. But then it reads layer two, which is the platform rules. So if I'm writing for a YouTube video, I've got a list of rules that I've picked up just from my experience being, you know, a strategist for five years on this platform. I've watched every Colin and Samir video, I've watched every Patty Galloway interview, every Mario Jews interview those kinds of things that I've made a list. If I'm writing a YouTube video, it knows to read my YouTube rulebook. If I'm writing a video that short form, you know, my clients, we opt to my short form for Instagram, there's an Instagram rulebook. And it's got all the different nuances on how to win on Instagram versus how to win on YouTube and those kinds of things. Then it reads the voice clone for the client. So I have one of these for myself if I'm writing on my, my, my own behalf. And the way that I train it is I just get it to look at all my emails, like anything that I've personally written and it kind of builds its own voice clone. With someone like one of my clients, you know, they have probably a newsletter that you can read a thousand articles of and the more data, the better, right? Like when it comes to training AI, I find that I get bored of training it after like inputting five or six sources, but it's better if you input like a thousand sources. How are you getting that many? Yeah, I mean, I'm ripping through credits like crazy. And in the long run, it's worth that money because it's going to create something that helps me identify something that's excellent. And then how are you like storing it? Is it all in Markdown files? So I did it one time. I had a big long conversation and I like fed it a bunch of different channels from a bunch of different niches and was like, here's like a thousand Instagram reels and identify the ones that have overperformed and the ones that have underperformed and then make a big assessment of the entire pool. So what I would probably do is like get as many categories as possible. So like what is the niche? What is the length? I think a big one is format. So like we're talking about carousels, right? And it's just like, it's not just like all carousels. You want to compare the carousels that are similar to each other. The closer they are to each other and then you can see the differences between them, then that's when you can like start to make like better observations out of them. This category does better than that category. And this category does worse than that category. The ultimate output of that would just be like a file for this format. And then you just have like all these files that you can mix and match. So you can be like, I want to make a top bottom carousel in this niche. Then it can learn like the best practices from everything in this category and everything in this category. Yeah. So you're saying if you want to make like a yap style reel, train it on yap style reels. If you want to make a talking head at the desk kind of reel, like separate them at the time when you're teaching it so that it knows the difference between the two. Yeah. The more tags and categories that it has, the more it can just like compare groups of them together. Have you gotten to the point where you're like building your own tools? Like I mentioned using sandcastles and vidIQ. The next layer would be like not meeting subscriptions and just building your own tools. I love tinkering with all these different tools. I mean, my favorite is just like vanilla clod or vanilla codex and just like. Really? Yeah, like I have it remember things as I write as I talk to it. So when I was learning about using AI to edit videos, for example, I kind of created this whole tool for like, you know, taking a long form podcast and then like cutting it up into like little pieces, adding captions and styling and stuff. So like I created something that did that. But the coolest part of that was not actually the outcome. It was just that like now my clod knew how to edit videos. So in conversation, I can just be like, here's 10 iPhone videos that someone took like, make something and it just like has all the tools like installed to do that. Wait, so you have clod actually editing and exporting videos and they're good? They're like 70% good. Wow. Okay, wait a second. Okay, so explain to me how you did that. So I used revotion, clod plus remotion and I'm like actually not that good at AI. I just tell it what I want. I think like that's the secret to AI. It's just be like, know what you want and then like, I do think you have to be good at something to like recognize good versus not good. We just go back and forth a lot and then like as it's outputting something, I'll be like, okay, well, I noticed like, okay, in this video, the head is moving. So you have to track the head. It's like, okay, you're like cut off on the the person. Like, don't do that anymore. So it's like, okay, how do we do that? And it looks up some tools to track your head or something or. And are you doing this in a single conversation over the course of like multiple days? Or are you like branching it out? Like, what's that look like? It depends. This was probably multiple conversations. So clod has like a context window and I think the maximum is like a million. But if you fill the whole million, it's like, it can't remember that much. So I've heard like the key is maybe like 100 to 200,000 token context. So I try to keep conversations like less than half or I go beyond all the time. But a lot of times you spend the whole beginning of the conversation giving it context, right? And then it like comes in and like makes the same mistakes that you just told it about last time. So the way I solve for that is I always start with like a project. And then I have a decision log. And I tell it like as we go through the conversation, make like add notes to the decision log, whatever we decide something. So that next time I can like end the conversation and then like start at the same like context. So when I notice I'm reaching the 200, 300, 500, whatever context, I'll be like, okay, we're going to end this conversation please write a prompt that I can hand off to the next chat. And it'll leave me a markdown file, and I just copy and paste that into the next one, and it can read everything and keep going. - Interesting. And so that's how you built this editing tool. You asked it how to do it. It kind of gave you some answers. And then you just kind of looked at its output, gave it feedback, and you did that what, five, six times, and now it's producing something that's actually almost passable. - Dozens of times. - Dozens of times, okay. - Like hours of conversation. - How do you figure that out? Are you explicitly telling it, hey, remember this part? - In my Cloud file, I have instructions to self-improve. It does log a lot of things on its own, but at the end, after I've handed off, like after it's created that document, I'll always be like, we're closing out this chat, please make sure everything is documented or something. And I'll ask it multiple times, 'cause it'll go through and document everything. And then I'll be like, okay, anything else you wanna do, and it'll be like, wait, one more thing, and it'll do something. - Is there ever any fear of all of the files that it's pulling, contradicting with each other, or becoming bloated? - Every week at the end of the week, I have like a, I'm looking at it. It's called weekly vault cleanup. So it'll go through my, everything, and just clean things up. So I have my vault on the side, and almost everything in there is like, written by clod and managed by clod, now clod and codex, because I just kinda switched to using both instead of just clod. - And by vault, you just mean like a folder. - Yeah, it's called vault, because I was originally using this tool called Obsidian, which lets you like read Markdown files, but I realized I don't really look at Obsidian anymore. I just let clod, like I used to go and edit it, and now I just let clod do it thing. What I've noticed sometimes is like, especially if it's something I'm actively working on, if I go in and edit it, clod will get confused and like edit it back. So if I, so I tell it like this is what I want instead, then it'll like remember it, 'cause I don't know how many times I've been like writing my newsletter, for example, and I'll go in and like make edits, and it'll come in and just like change it back, I'm like no. - So everything takes place in the, in the clod window, and then you kind of like out, like you, you tell it to do the job that you want. - Yeah, or if I am gonna make edits, I'll be like I made edits. - I see, okay, got you. So any other kind of like sanitization, clean up stuff that you're like really aware of when you're communicating with your AI? - Like managing the context is a pretty big one, just like what information is it looking at? What information does it have? So especially when we're doing the script writing, right? I have so many different files. It's like brand files, voice files, platform files, content type files, things like that. So you don't want it to be reading all of them all the time. I noticed like, I ran into problems with giving it too much because then it starts like skipping files. It's like, you know, it spends its whole context window on reading instead of the job it's supposed to do. - I find that it will lie to me, straight up lie to me, very often, and I am losing trust, and I'll ask it to like read 15 files, and it will be like, okay, I did. And I was like, okay, well, what happened in file number six? And it's like, oh, well, I actually didn't read that one. And I was like, wait, you just told me you read that one. I don't have problems with it like lying necessarily. Like, if I tell it something, I feel like it does it, but then I've noticed if it's like too big of a job, it'll skip steps. So for example, in my scripting process, I used to be like, follow this process, we're writing an article, whatever. But now I'm like, okay, this is the process, start with step one, and then breaking it down helps a lot. And you don't even feed it step two until step one is done. - So I have like a process doc that explains like, all the steps and then which folder each one is in. I'll ask for step one, which is like idea validation. Or in most cases, I've already like kind of validated the idea, and it would be like research. - And are you using Whisper Flow? Or like, are you dictating? - I love this thing. - We didn't even bring it up. But like dictating instead of typing, oh my gosh, game changer, like it's basically all that I do anymore. And I like Whisper to it when I'm trying to be quiet and I talk to it, and it's just so much more efficient. I love it. - I whisper, I know how this is. (laughing) - You take the Whisper Park very seriously of Whisper Flow. One thing that I found AI to be super, super helpful for is connecting to my team tools. So like Slack and Notion primarily. So anytime that, you know, I need to add something or modify something in Notion, I'm actually not really doing the work in Notion. And if I am, I'm using Notion AI's tool, which is also excellent. But I'm usually using Claude, and I'm telling Claude, like, okay, we just wrote a script. Let's add it into the scripting pipeline. I have a special formatting that it knows how to do, and then it drops it in place. It assigns the right person to review the script, or if it's a video editor or something like that. And it's so helpful. And similarly with Slack, like if I'm gone for two or three days out of office or something, it gives me this like really concise summary. It updates me, which is really nice. And it can also like fire files and conversations to different people. - It's like a super game changer to have Claude connected to Slack and Notion because I can find anything so fast. I also have a weekly report where Slack, or sorry, where Claude will go in, go through the entire like company Slack and like any changes that were in Notion, and it'll just like give me a report of like what is going on across the company every week. So it's like product changes, or like this team is doing that, or decisions that people across function will like have made. I like learn so much from that. And it presents it to me in a HTML file. - Oh, okay, so then what's it look like? - I ask it for like a, it's called my company pulse. And I just have like an overview tab of kind of like a summary of all of the departments, and then I have a tab for each department. It'll just summarize all the new things or changes that have happened. - Even in departments that you're not necessarily like super affiliated with. - Yeah, because we're media, right? So like we make content about new things, or we need to stay up to date on like what's happening, because like if there's a new product release and it's like we might have to do something with that later. - That's really clever. Even in the work that I do, there's departments that are like kind of different from the stuff that I spend a lot of time in. And it would be really helpful to have just like a weekly summary of what sponsors came in, or like what we're doing with paid ads on any given week. You know, those are things that I like touch semi-frequently, but you know, there's entire team members having these like conversations with each other every single day, and I definitely don't keep up with that. So that's super clever. - And HTML files are excellent. Like I love HTML files, and just like asking it to present the information in a way that's not just big blocks of text, 'cause like I cannot read five paragraphs as a response. Like I just get too bored. I've actually asked Claude, like it remembers for me. It's like short, fifth grade reading level, bullet points when applicable, 'cause like I just can't read it. But then HTML files are really cool, and they can be like interactive too, so you can kind of like gamify a little bit, which is cool. - There was like a period of time where Claude was having like super wordy, not falling instructions like all these things. So there was like this skill that was like like zombie, or it was caveman, and then it would be like bug one, bug two, and like five words for each. - Where do you go to find new skills to teach your AI or to stay up to date with how things have changed between, you know, Opus five and Opus 4.8, and all these different kinds of things. What are your resources? - My husband. (laughs) - 'Cause he works in the AI space, right? So he's like super deep. - Yeah, he's like super deep in frontiers of AI, he's like on the front line. He calls himself a tinker, 'cause it's like things that are not quite working yet, but he's like trying them out and experimenting with them. I think one of the most important things is like getting a feel for their personalities and how they respond. So it's like the more you use them, you're like, oh my gosh, every time I talk to Opus 5, it skips my instructions, it like thinks on its own. And it's just like, okay, like I need to change the way I talk to Opus 5 versus like now the reason I think a lot of people are switching to like Codex for like content creation is like Claude used to have a better writing personality and now Codex, I think the personality just like doesn't sound as AI anymore and it's not as wordy. Claude has become just like so wordy. And then I was just like, someone else told me this and I was like, I'm kind of wordy too. And then it was like, no, but you're like paying for the words with like the tokens, right? So it's like, oh yeah. - Yeah, I'm really seriously thinking about switching over, I'm pretty displeased with Claude and how it's been the last month or so. - What model are you using? - So I've actually defaulted back to like 4.6, Opus 4.6 or Opus 4.8 sometimes. - Yeah, I do too. - Five is untrustworthy, Fable is kind of lost its thing, although Fable was sick for like two weeks. - What I found for five was like, I was saying it was like, it ignores my instructions, but it's really good at thinking on its own. So if you just tell it like, this is my goal and then it'll come up with its plan and it carries it out pretty well. But I've kind of like switched to Codex and Grockbots. - Yeah, tell me about Grockbots. This is a brand new thing and you just wrote about it last week in your newsletter. - Yeah, I think it just came out like a week or two ago and I think I really like the mental model of this because when OpenClaw came out like six months ago, it was like all the rage, but it was just like it didn't quite work. Everyone always says like, Treat. A.I. is your employee, or your teammate, or something. It didn't quite work, and I tried to build out my little team of agents, but there was some bug that came in where every time I switched an agent, it would be like, "Oh, your usage maxed out," and then it would just stop working for two hours, and that was the end of my open-cloth era. So Grockbot came out, and it's really similar to that where you kind of just have a team of agents, and you just think of it like you have a team under you, so you talk to one, and it's like every time it has a new job, it'll spin up a new agent, and the difference is it persists. So it's an agent that is always there. So I have a fleet of agents, and I've named them all after Pokemon, so I have like. I have like, and so every time you want to do something, you have to start a new chat. It's just like the same teammate every time. It's more intuitive, and it makes more sense. - What are some like content creator economy agents that you have that you use? - I have Poregon, for example. Poregon is head of my me plus machine newsletter. Poregon then went out and hired like, Poregon 2, actually, for research, and then it has like Tinkaton for like figuring out angles and outlines. - See, I only know the main Pokemon, so this is kind of blown my mind a little bit, but. - Yeah, so it's like the thumbnail one, so it like helps me with like, thumbnail ideation. It's not really thumbnails. It's like just images within the newsletter. - For the newsletter. - Within the newsletter, yeah. So it's just like, it went out and hired like a team, and each one has like a job, and the funniest thing is you can put them all in a group chat together, and they talk to each other. So I'm just like, "Hey, you guys are in charge of my newsletter." Like last time it didn't go so well. Like last time halfway through, I like had to abort and switch back to cloud, but this time I was like, "Okay, last time I didn't work so well, like what happened, like figure it out, we need to clean up this process so next week it's smoother." And I just saw them like in the chat talking to each other. - So wait a sec. So if you're writing like a newsletter article, you're coming in as kind of like the CEO of all these Pokemon, and you're saying, "This is what I'm looking for." And then they communicate with themselves, and they give you an output, that includes images. - Well, kind of. I've wrote them, which goes out and like, just tells me the news every day. So it's like, this is what's happening in like the AI sphere. And every day, it'll just give me like five to 10 points of like things that are happening. And I'll be like, "Oh, that's really interesting." And then like, I really should do this sooner, but probably at like Thursday, Friday, I'll start looking at, that was an interesting point, that was an interesting point. I'm gonna try this myself. And I just kind of like tried over the weekend, and then this is new. So this whole process is new, but I also had cloud doing this for me. So actually, cloud would do the research for me, and then send me a message in our gimbals Slack every day with like news of the day. But from there, I'll talk to Porygon, who is like, me plus machine project manager. So I actually just talked to Porygon, and just kind of like, tell it what I want. And then it talks to whichever bots are in charge of like that step of the process. So I'm not really talking to like the rest of the Pokemon. - And this is like out of the box, Grockbot. Like you sign up, you create the first initial agent for the thing that you want it to do. And then it does everything, and it does a pretty good job. - Pretty good job. It's, I would say it's pretty buggy, but not necessarily that like Grockbot is buggy, more that I don't know how to use it yet. So in the conversation that came up afterwards is like what happened, like why did everything go so wrong? What happened was I had my chief of staff bot, and I told my chief of staff, I need someone to manage my newsletter for me. And then when it spun up Porygon, I also told Porygon, I gave Porygon instructions, and then Porygon spun up like the team, and I gave like team members instructions, and they didn't like align on the instructions. So that's why they were all like out of sync doing things like that I didn't want them to do. I can't really control the context. Like you know how in Claude, I go like I'm really conscious about like it has this information, and then I'm like starting a new chat. And this one I don't really have that capability. I don't really know what it's doing. It's always there. So I'm a little nervous about that, but I think it's okay because each one has its own specific role. So I imagine like since the role is kind of always the same, and it's all hooked up to my notion, my personal notion. So it's like keeping its memory there. And then the second thing is like the models that it has. I don't get to choose which model it's using, and I just feel like the thinking isn't all there. One of my favorite things about Claude is how good its research is. And it was picking like Groc or something. I have no idea what it's using. But it wasn't a strong Groc thing. Yeah, it was okay, but then when I went to have Claude do it, it was just so much better. And I was like, okay, we're switching to Claude. This came out a week or so ago. So it's just kind of like finding the personality and limitations. And like the more you use it, the more you like know, okay, I'm never going to ask it this again. Okay, Alina, I've got one last question for you before we wrap this up. Who is an AI creator that you spend a lot of time consuming their content? He's not a creator. I just follow whatever my husband tells me to do. Is that an answer? It's just at Tyler Nishita. I would say like on X, it's not necessarily like creators. It's just like people that are like really good at what they do. Thank you so much for watching or listening to this episode. If you're new here, please consider subscribing or leaving us a review on the audio platforms. It helps us as a very new show. If you want to check out another episode, just look through our library. It's starting to build up. We're at like 40 episodes now. And we have a new one coming out every single Thursday, so thank you for being here. And I will see you then.