How to Acclimate Your Family to AI

Announcer (00:04): Welcome to Raw Data with Rob Collie, real talk about AI and data for business impact. And now, CEO and founder of P3 Adaptive, your host, Rob Collie.

Rob Collie (00:18): All right, Justin. Well, welcome back. One guest in a row and now back to you and I. That was a really good show.

Justin Mannhardt (00:26): Oh, man. Juan?

Rob Collie (00:28): Yeah. I haven't finished my re-listen to it. It's just so dense with good information so I want to re-listen to that podcast the whole way through and take notes. Not just things he was saying, but the things we were agreeing on.

Justin Mannhardt (00:42): Yeah. A lot of great lessons both as a leader, how to approach it, and technically some of the things they did, but not uber technical in our discussion, but just conceptual. I loved it. Hats off to those guys.

Rob Collie (00:56): Dear listener, if you didn't listen to last week's episode with Juan Garcia, I highly recommend it.

Justin Mannhardt (01:02): Bookmark it. Bookmark it indeed. My son last night took his girlfriend on a Claude Code Cate.

(01:11): Claude Code date. Is there an app for this?

Rob Collie (01:17): Yeah, it's called Claude Code. My son is about to graduate with, drum roll, a computer science degree. What is the job that is going to be hardest to get right now? It's basically entry level developer.

Justin Mannhardt (01:29): Oh, those charts are just crashing and burning.

Rob Collie (01:32): Yeah, it's really bad, right? Claude Code is a better entry level developer. It's kind of like an intermediate. It's kind of better on every dimension than hiring an actual human who's going to do things manually, right? As a father, I've been concerned, and I had two problems. I had the concern for my son, what his career's going to look like, et cetera. We're out here just running super, super, super fast out of necessity and learning a bunch of stuff and he's still kind of in the school mode, and so I know that he's not under the same sort of real world pressure yet. Then I had another problem and it's one of the things ... I love it when two problems sort of align with each other and become each other's not solution, but they mitigate each other. I had no idea what to get this kid for Christmas. So what did I give him? I gave him a couple of months of Claude Max.

Justin Mannhardt (02:24): Wow. You're the best. That was a Christmas present.

Rob Collie (02:28): Along the way, I've discovered that the way to stay in touch with my son ... He lives near you. He lives in Minnesota. The way to stay in touch with them is to use Discord. He lives in Discord. I mean, he's a gamer. So the first Claude Code project that he took on was of course the World of Warcraft add-on. The second one he took on was a World of Warcraft add-on. And then the third one I think is a World of Warcraft add-on/website. But hey, these are okay. We needed him ...

Justin Mannhardt (02:55): Love it.

Rob Collie (02:55): To have traction to start to feel what this is like. This is what the future looks like. We need to bring him along with that. People listening right now are like, "Okay. Well, how did the date go?" I don't know. He decided that last night for their date, instead of Netflix and Chill, it was going to be Claude Code and Chill. And he put some Easter eggs in the website for her.

Justin Mannhardt (03:21): That's cool.

Rob Collie (03:22): I take this as a sign of successful gift and successful open eyes of son to how the world is changing. The poor guy, when I was talking to him, he was really the right level of reactive to this. I was describing to him like, "I've got these apps and these platforms I've been building where there's tens of thousands of lines of Python and I still don't know any Python." And he did just speak it out loud. He's like, "Man, I'm kind of really bummed to hear that because I've gotten really good at Python." And I'm like, "Yeah, I know, man. I know, but now you can be even better." I think he's kind of through that first shock and sadness phase and he's into the excitement and gaining momentum phase. And so as a dad, this is exactly ... I mean, it's very rare that I've ever done something so right as a parent.

Justin Mannhardt (04:11): Well, the fact that he even went to the extent of tailoring his site or app to kind of be cute and caring and interesting. It's like, "Okay, man. Game recognized game."

Rob Collie (04:25): That's right. Anyway, so Claude Code dating is now a thing. I'm sure this isn't the first time this has happened.

Justin Mannhardt (04:31): We've been married for a while, you and I, so it's like we don't think about using Claude Code as an excuse to go on a date.

Rob Collie (04:37): That's right.

Justin Mannhardt (04:38): It's usually like, "Hey, let's go grab beers with friends." That's the move. Coincidentally, we had some fun with AI in our household recently. Your son is getting ready to graduate college. I've got two sons. They're nine and seven, just innocent little dudes. So I do spend time thinking about what their lives are going to be like when they're adults. The way technology is changing, you wonder like, "How back in my day we had this thing called an iPhone." It's going to be so different.

Rob Collie (05:15): There's no way to have any idea what that looks like. No. That's way beyond the event horizon of what we can predict.

Justin Mannhardt (05:24): And so I've tried to find fun and creative ways to just expose them to what AI is because I can't imagine a future where AI isn't part of it. One theoretical outcome, but it's like, "Ah, no."

Rob Collie (05:38): I don't understand anyone who would suggest that it's going away. If it's a bubble, that's a financial bubble that's part of the companies that are behind all of this. The hardware, the LLMs, they survive any sort of bankruptcy event, right? They still sit there and someone buys them for pennies on the dollar and off we go again, right?

Justin Mannhardt (05:56): Totally. So we've done some things over the past 12 to 18 months where we would go sit in the basement and we would pull up Dall-E or something like that and be like, "Hey, Connor. What do you want to draw a picture of?" "Well, I want a T-Rex wearing rollerblades with a bazooka in a city getting attacked by beavers." And like, "Okay, sure." So they would have fun doing that and we would go around and we'd save them and they'd print them off at my mom and dad's house and show them off. But it started to inspire them with this idea this technology just allows you to have ideas and then to have so little friction between making that idea real.

(06:32): And so the other day, we were having dinner and it was actually my oldest son, Nick. He asked me. He said, "Hey, can we play with AI tonight?" This all just kind of happened on the fly and I said, "Well, sure. What if we worked with AI to write a story together?" And it's sort of a wandering story about the story. There were some neat takeaways, both as a father and also a business leader.

Rob Collie (06:54): I mean, I can imagine this being something that kind of doesn't go well because if you haven't thought about it in advance, if it's happening in the moment ...

Justin Mannhardt (07:03): Yeah.

Rob Collie (07:04): I mean, just even the Dall-E stuff or generating images, if it's having trouble drawing that bazooka or not putting him in a city like what you meant, the kid is going to get like, "Oh, this thing sucks." It's so easy to lose traction. They're not going to be like, "Oh, the four-fifths of the picture that it got right were so amazing." They're just going to be like, "This thing doesn't work." They're kind of right in that regard, right? So I could see this not going well, but it sounds like it did.

Justin Mannhardt (07:32): It did. It ended up going well, but you're still right about this point. I think that's an important part of the story and the experience. Just to set the scene, so we went downstairs and I screencast my laptop up to the TV. And we're using Claude Code for this.

Rob Collie (07:44): Oh. So they're seeing command line. They're seeing terminals.

Justin Mannhardt (07:48): Oh, yeah. We're in terminal.

Rob Collie (07:50): They're like, "What is this? What is this primitive experience?"

Justin Mannhardt (07:54): And so the prompt, I gave Claude Code a prompt that was something to the effect of, "Hey, ask each family member one question at a time and kind of go around the horn and ask each of us an open-ended question, a multiple choice question and a yes/no question to gather information about what this story is about."

Rob Collie (08:14): This is like an AI Mad Lib.

Justin Mannhardt (08:16): And so right off the bat, what does Claude Code do? Just face plants right away, but in an interesting way. And it's probably because I was in Claude Code. I explained, "I want you to ask us these questions and then you're going to write us a story." First thing it does, it shows me the multi-question, multiple choice dialogue you see in the terminal. It says, "Great. Where do you want me to build your web app?"

Rob Collie (08:39): Wow.

Justin Mannhardt (08:40): It thought I wanted to build an application. But right away, Nick was like, "What is it doing?" I was like, "Great lesson. Sometimes AI is not going to understand what you want it to do or it's going to make assumptions based on what's going on," and so we need to reexplain what we were doing. Eventually, it took a couple prompts to re-explain.

Rob Collie (09:05): We're just writing a story.

Justin Mannhardt (09:08): Just writing a story.

Rob Collie (09:09): Okay.

Justin Mannhardt (09:10): So then it did the thing. It asked the questions and it went around. It definitely tripped up with the open-ended question. It couldn't help but just ask multiple choice questions this whole time.

Rob Collie (09:23): Okay. Claude Code behind the scenes has its own system prompts that you can't see that are telling it exactly how to behave and it's sort of like training it to be the perfect co-programmer. And the perfect co-story writer isn't ... It would be a different set of built-in instructions that you and I can't modify.

Justin Mannhardt (09:42): So we got through the questions and then the last little thing we sprinkled on there says, "And as you're writing this story, insert at each appropriate moment in the story a prompt for an image generator so we can create pictures to go with our story."

Rob Collie (09:59): So in that prompt for image generation, did it put any seeds or anything so that the pictures would look consistent with each other?

Justin Mannhardt (10:06): Yes, it did.

Rob Collie (10:08): Wow., How clever. Claude Code, you are so clever. One of the dangers in this, or potential obstacles or pitfalls, is that in the course of doing all this, it might start to lose contextual memory of and clarity about its assignment. When you were talking about, "No, no, no, no, no, no, no, Claude. I'm not trying to build a website," did you end up putting anything into the Claude.md file for the project that helped it remember? For example, I've been working on this thing. I met with a former podcast guest, Dave Gainer. I had breakfast with him recently. And he was telling me about this manual, I think it's manual for him, workflow that he does where he puts the same prompt to all three or four of the most current LLMs, gets their answers, and then cross-pollinates to get a better and better answer.

(10:59): And I'm sitting there thinking, "It sounds like he's copying and pasting." So I've been working on a new agent called [inaudible 00:11:08] that does exactly that. This sounds like a really simple thing, right? Enter the prompt and it'll federate it out to the big three LLMs and then allow you to interact a little bit with each one's response and then say YOLO, cross-pollinate. And it turns out to be a little bit more difficult. This wasn't just a one-shot sit down, right? I've been having to fight this thing. And one of the things that this thing keeps doing is whenever it runs into a problem with one of the LLMs, it just says, "You know what? We're just going to downgrade the LLM. We're going to go to one of the lesser LLMs. We're going to go from Gemini 3 Pro to Gemini 2 Flash," prioritizing success over quality.

(11:49): I had to eventually just put it in Claude.MD, "Do not suggest downgrading LLMs," right? This thing isn't going to work. Dave's not going to use it if the trade-off is I've got to use crappier LLMs to make it work. I want him to see something. And maybe he's already aware of all this so this might just be a total dud, right? He'd be like, "Well, Rob. I've already got one of these." But anyway, that wandering of mission, how many times I had to tell it manually, "No, no. Stop, stop. Stop downgrading LLMs on me," I could imagine something similar happening in your story writing. You really want an app, right? How about a website?

Justin Mannhardt (12:26): Yeah, there was definitely some of that redirect that was happening, which I think is a good lesson for the kids, but also just a reminder. If you go in there, we have this expectation where we can just make a request and we're going to get what we want right away. Not to take anything away from the power of everything, but there was still that lesson of iteration in there. So we got our story, wrote us a nice 4,000-word Markdown file broken into six parts. And so then we popped over to Nano Banana and we gave it all the image prompts, gave us the pictures, and it just banged them all out in sequence.

Rob Collie (13:03): I'll stop you right there. I don't know what Nano Banana is. Somehow, I've missed this.

Justin Mannhardt (13:08): Nano Banana Pro is Gemini's image generator. It's pretty new, created a lot of hype. It's also really good.

Rob Collie (13:16): You would use this over OpenAI Sora.

Justin Mannhardt (13:19): I never really got too far into the image things. I've used GPT with Dall-E, I guess I can say., Back in the day now.

Rob Collie (13:27): It was back in the day, yeah.

Justin Mannhardt (13:28): But it did a really good job. There was, I think, six or seven discrete images and the last one is where you can kind of feel like, "Oh, it got tired because the perspective is off," or, "Something is really small and it should be bigger." But it maintained a very consistent stylistic theme the whole way across.

Rob Collie (13:49): This is really interesting. So do you attribute the consistent style to Nano Banana understanding the assignment? You think it applied the style. Or was it something about the prompts themselves? Because now I'm thinking if the prompts came from Claude Code and then you handed them to Nano Banana, if there were some sort of numerical seed, it's not like Claude Code was going to know what system you were going to use. I'm starting to think that there's something just generally smarter that happened here.

Justin Mannhardt (14:19): Yeah. So Claude Code didn't know what I was going to use for the images. We'd even put in its output at one point something to the effect of, "And you could now go take this to," and list it off, three or four AI providers and none of them were Gemini or Nano Banana. But the prompts, I'm just looking at that chat now. Yeah, I think there's something special about Gemini where it realized it needed to maintain.

Rob Collie (14:46): It sounds like you gave all of the prompts to Gemini kind of in bulk all at once.

Justin Mannhardt (14:51): That's correct.

Rob Collie (14:52): So it had the opportunity to understand that these were all related and it also had the opportunity to get tired on the last one. Because if you had been doing them one at a time, there would've been no exhaustion, but it might also have looked like seven different illustrators. Almost certainly would have, actually, so that's kind of neat. Okay.

Justin Mannhardt (15:08): Yeah. And the kids were like, "Wow, that's so cool," and, "Look." Just to give some context on the story we wrote, it was about this young girl named Zara. She was an inventor and she has a magical wrench and there's a swarm of robots and floating sky islands.

Rob Collie (15:23): Oh, it sounds like a Mad Lib to me.

Justin Mannhardt (15:24): Yeah, just so good. So the kids see these pictures like, "That's the story we wrote. That was our idea," right? And so they're just feeling really cool about it. So this is when I think I texted you this. I was like, "We did something fun with the family." We ended up using lots of tools in this whole process and one of the tools we ended up having to use was Microsoft Word.

Rob Collie (15:45): Go on. It's just another one of those Nano Banana things that I haven't heard of. What is this newfangled AI tool?

Justin Mannhardt (15:53): So of course we get the story, we get the pictures, and then, "Okay, what do we want to do now?" "We want to listen to our story."

Rob Collie (16:02): Oh.

Justin Mannhardt (16:03): So I asked Claud Code. I was like, "Hey, what are the best AI text-to-voice tools that I could use?" and it rattled off some things like, "Oh, you could try Notebook LM or this." And so at this point, it was getting kind of late so Dad was getting a little tired.

Rob Collie (16:17): Yeah. We're not doing another coding project.

Justin Mannhardt (16:18): Okay, Notebook LM. Ah, that's just doing a podcast summary of the story, not just reading the story. So I'm like, "A-ha. Read Aloud in Word."

Rob Collie (16:30): Hot damn.

Justin Mannhardt (16:31): So I copy and pasted the whole thing into Microsoft Word and I press Read Aloud and I kick back and it goes, and you'll understand this in a second, "#Zara," the whatever, whatever.

Rob Collie (16:46): Double hashtag.

Justin Mannhardt (16:49): ##Chapter one.

Rob Collie (16:51): Oh, no.

Justin Mannhardt (16:51): That's what was written in Markdown, right? So Nick's like, "Why is it saying hashtag? This is stupid," right? Just like you were saying earlier. So I went and I find and replace all the hashtags. I'm like, "Okay, let's start again." So it starts reading, reads the first part, and then it goes, "Image prompt one: a sprawling sky filled with ..." because the image prompts are in line with the story.

Rob Collie (17:17): Nice, yep. Still in there, yep. Yep.

Justin Mannhardt (17:19): So I said, "We're just going to deal with it." So anyway, it's like the kids really liked it and, "Oh, Dad. Let's clean up the image prompts tomorrow and we'll listen to it again." They're really proud of this thing at the end of the day, right?

Rob Collie (17:28): Aww, that's so cool.

Justin Mannhardt (17:28): Because they feel like they made it. And there's some interesting things I noticed that I don't think the kids would notice about the thematics in the story coming from things that have already been told. So the main character has this wrench that's the source of this magical power, and later in the story she realizes the wrench doesn't hold the power she does.

Rob Collie (17:51): Ooh.

Justin Mannhardt (17:51): So it's like you go to the story of Thor or something like that that's come before. And you see sort of those regression to the mean elements, but it was still original enough because of the Mad Lib vibe that got into it that made it fun. I got one last chapter in the story about the story. So the next morning, Nick was asking me all kinds of questions because we ended up using ... Oh, by the way, my wife works for the hospital system and the only AI they have is Copilot. And so she's watching it and she goes, "Oh, I see why you like this as a comparison."

Rob Collie (18:27): I know, right? I do run into people who are really, really, really stoked about Office Copilot though and I'm like, "Wow. Okay."

Justin Mannhardt (18:36): Compared to nothing maybe.

Rob Collie (18:37): Yeah.

Justin Mannhardt (18:38): Nick's like, "We used so many tools. We used Claude Code. We used Notebook LM. We used Gemini. At one point we used regular Claude for something," and he's like, "Wow, there's lots of AIs. You don't have AI. You have lots of AIs." And he says, "So what else could we do?" And he asked me, "We could write a sequel to the story, right? Could we make a movie out of this story?" And so we kept going with all these ideas so we created a GitHub repository of our story, as one does nowadays. In there, we've got the images that we created, the final story, the original prompt we used to go through the exercise. And then Nick helped me, we used Claude again for this, add a document in there called Creative Quests. And so one of the quests is, "Create an audio narration of the story. Turn this into a video. Make a video game based on this story. Contribute your own story."

(19:39): So I think we'll link the repo in the show notes if people want to fork it and play around. I wish I would've saved this part. I might be able to resume the chat with Claude and go get it. I was like, "Okay, but there's real lessons here beyond just, 'Oh, this was a fun exercise with my family.'" Things we've talked about before on the show that are good reminders like AI is not going to get it right on the first shot. Giving it good context and good instruction and managing the context window, all those things proved out to be true in this case. The idea that we have lots of different AI tools and some tools are better at things than other tools was a lesson. And so for the kids, and I think for anyone, just that idea that every single day that we make progress with this technology, the friction between an idea and making it real decreases.

(20:34): That was sort of an uplifting moment for me as much as you were saying about your son. Entry level jobs might be going away, but it's like, "Yeah, but there's also this new thing that's happening that I'm going to maintain my optimism about how that affects us in the future." We'll see.

Rob Collie (20:50): How long was this process with your family from the moment you sat down to the moment you're like, "Okay, time to go to bed"?

Justin Mannhardt (20:56): A little under two hours.

Rob Collie (20:57): Okay. That's a lot of ground.

Justin Mannhardt (21:00): Yeah.

Rob Collie (21:01): Yeah. How cool. Because again, going into it, you had to know ...

Justin Mannhardt (21:04): "This might not work."

Rob Collie (21:06): Yeah. But well done, Dad.

Justin Mannhardt (21:09): Yeah. To the point about your son on Python, Nick has on his little Chromebook. He has a learn Python for kids thing. And so he's like, "Do I need to learn this?"

Rob Collie (21:23): No, no.

Justin Mannhardt (21:25): No.

Rob Collie (21:26): Son, let me tell you the truth. No, you don't.

Justin Mannhardt (21:29): To parents out there: Exposing our kids to this stuff is really important.

Rob Collie (21:34): I have to go take the dog to the groomer and I tried to get Claude Code to do this for me and it just won't. It doesn't understand the assignment. I really like these two stories sort of happened at the same time. And introducing our kids, it's kind of what it comes down to. So until next time.