B (3:12)

rules are always changing?

C (3:14)

Yeah. Like every change you make changes the rules. And it's like, that just sounds really miserable to do. So we went with the other, which was everything is connected.

A (3:23)

Okay, so I. I love how you guys used AI to produce a game jam. Rules for your game jam A. I think that you're going. I think it's pretty clever. Now, I want to. I want to explain a little bit between Matthews. Matthew's my oldest son, and my youngest son is David, who's on the phone. And there's 13 years between the two of you. Right. David is a freshman at byu, Idaho. Matthew is a seasoned professional out there. Even though he's not a coder, he knows how to code, and he's a product manager. AKUB is a hardware engineer. He's electrical engineer, but he does a lot in AI. He's working on his master's in artificial intelligence now, but he's got three or four years of work experience as well. So we got a gamut of of experience in this. And I think the most fascinating thing is how the scores ended up. Poor college student. How much, David, how much AI did you actually use? How much did you spend on this game jam?

D (4:32)

I spent exactly $20 with the base subscription of Claude Code.

A (4:39)

All right, so, David, you spent the least amount. That's what it amounts to. Second amount so spent was Jacob, if I remember right. J. Right, Jacob.

B (4:49)

Yep. I used the $39 a month GitHub Co Pilot Pro plus, which gave me access to all the premium models and 1500 requests for the month, which is 100 with Opus.

C (5:02)

Right.

A (5:04)

All right, Matthew. Matthew, you're the product manager. So you needed the most help from AI or you thought you did. So how much did you spend, Matthew?

C (5:14)

So I already had the Claude Max 100 subscription. I didn't feel like shelling out for another one just to limit myself. I just kind of accepted I would lean into AI and take the score hit. On top of that, I spent about 12 bucks on nano Banana.

A (5:30)

All right, so did all of you use different AIs, or were you all just stuck with the one? Jacob Cloud, copilot David, Claude Matthews, Claude and Nana Banana. Did you guys all stick with basically one platform or did you use multiple platforms?

B (5:49)

I use the GitHub copilot and that gives me an array of models to pick from. But I primarily just used Opus and then I also used a little bit of Nano Banana to generate some sprites for my game, but that was just all on the free tier with the Nano Banana.

A (6:06)

Gotcha. And you, David?

D (6:10)

I mainly use Sonnet for everything, but I did use Nano Banana to generate just basic sprites. And then for variations I went into Photoshop and did that myself.

A (6:24)

Ooh, some. Some hand. Some hand jam in there. Right.

D (6:29)

It was changing buttons to change the hue of one color.

A (6:33)

Oh, there you go. All right. And Matthew, only two different tools or did you use a variety of different. And what models did you use?

C (6:40)

So I was in Claude code. I used Opus for almost everything along with the team of agents functionality that they had recently rolled out. I wanted to try it. I figured with a tight time constraint might be a good tool. It was probably a little wasteful, but I think it let me get a lot of scope. I did use Sonnet a little bit and that was for the nanobanana MCP calls. I probably could have used some Nano Banana free, but I wanted to programmatically call it because I had limited time and I had like 200 images to generate in my game. So I had Claude prompting nanobanana there. Finally, I used a little bit of elevenlabs free just to generate a few voice clips for my game, which I think added a lot of personality.

A (7:26)

As pred. So use a lot of different AI. I would expect that from a seasoned product manager. Right. You're used to bossing around teams of engineers to do things. So the agentic aspects or the agents that you had out there was pretty, pretty normal for you. Right? I mean, this was something that you. You already had a skill to do. Would you say that's right?

C (7:50)

I would say so. One of the things I did in preparing for this was I just to get in a sense of what I could do with the different cost tiers out there. I have Google AI Pro, which I didn't use on this game Jam because I didn't want to add another 20 to that. So I was considering it because Gemini is pretty smart. But what Gemini is not great at is agentic tool calls and finishing the job. What I did was I asked Claude OPUS to come up with a game design that would be something that was achievable but ambitious for a game jam. Right. I took its game design and I used. I compared the following models. I compared opus. I compared GLM5 because they say the Chinese models are nearly as good and they're a fraction of the price and it's because they're training off of Claude outputs. But yeah. And then I compared the Gemini Pro or 3 Pro and then Gemini 3 Flash. The reason I include Flash in there is that weirdly enough on the benchmarks, Flash Out Benches Pro for code, what I found was that Opus was leagues beyond everything else. I also benched GPT 5.3 Codex and where I'd ranked them. So Opus was far better. It got the artistic vision and added some details that made it actually fun to play. Like it produced a fun, balanced game. The next, I would say, is a toss up between GLM5 and Codex. Actually.

D (9:24)

Codex.

C (9:25)

Codex produced something that was exactly to the spec with no artistic flair whatsoever. No design chops, everything was text, no icons or just visual interest on the page to explain what was happening. But the mechanics were 100%. GLM5 came up with a decent UI, but there were some mechanical issues in a couple prompts it probably would have been fixable, but it had a decent like sense of artistic flair there. And then the Gemini models left a lot of things unimplemented. In some cases were just straight up broken using anti gravity, which kind of surprised me, honestly.

A (10:03)

So. But you used agents to do some of the work once you got started, is that right?

C (10:09)

On the game itself, I didn't write a single line of code myself.

A (10:14)

No line of code written yourself. Let's switch over to David. David, you, you prepped for the game jam by coming up with a bunch of agents that you put into play and you gave them different personalities and things like that. Right? Is that what helped you?

C (10:28)

Yes.

D (10:29)

Yeah. I primarily had four different agents, all with different tasks and all with different tools that they could use. The main ones were an overseer, an architect, a mechanic, and then an integrator. The overseer, it would test the code every single time. It would test the code every single time and it would run diagnostics. That was its main job. And then it would, it would send, it would pretty much delegate the task to the specific agents, depending on what fixes were needed. And the architects built all of the skeleton stuff of what needed to be done. Like just the bare basics of like the basic user interface, how things worked. And then the mechanic would come in and it would do all the nitty gritty stuff in there. And then lastly the integrator, because I didn't have enough time to get all the art, I did that the end. The integrator would lastly come in and implement all of the correct assets when I had them.

A (11:55)

So you use. Use four different agents, all using the same model or were they using.

D (12:01)

All of them ran on Sonnet?

A (12:03)

All of them ran on Sonnet. All right, what about your approach, Jacob? Did you use agents? Were you. Were you doing all of the. The calls to AI and tools, or did you use an agentic framework at all?

B (12:18)

I kind of wish I had gone a little more agentic. I started, I decided what I want my game to be, and I decided what engine I want to use. So I picked phaser 3. I consulted AI and was trying to figure out what engine would require the least amount of manual user interface kind of tools and found out phaser 3 runs completely in the browser, completely with JavaScript code. So I'm like, perfect. AI can touch everything within that. And then I just pulled up the co pilot and said, write me all the structure I will need for a 3D platformer game on phaser 3. And it took maybe 10 minutes. And I told specifically make subdirectories for like the assets for the maps for the rules, and then create a markdown file in each one of those with, you know, the basic rules it's supposed to follow in each of those subdirectories. So within one or two prompts I had a basic 2D platformer. And then from there it was then adding in assets for my puzzles, such as, you know, elevators or doors, explaining the rules to it. And then finally I gave it a map. I had to draw out the 3D or like the 2D platformer maps myself. And I designed the puzzles myself. And I took a picture, showed it to the AI and explained how that you're supposed to solve it. So it would determine how high ledges had to be, you know, accounting for how high the character can jump. You decide how, you know, why the doors have to open so the character can squeeze through it. Did all that reasoning. I just designed the puzzle myself and once I had one, I just kept writing up puzzles and kept showing it to it and it was able to fill up the rest.

A (14:12)

So you actually drew out the puzzles on like a sheet of paper?

B (14:16)

I did, yeah. Let's see. Kind of like this here. Just get them out.

A (14:23)

And so it just took that and was able to take something from the physical world and create something digitally from that.

B (14:30)

Exactly. I was really hoping it would be intelligent enough to make its own maps. I thought once it knows the puzzle pieces, maybe it can make Its own, but it just, it made a bunch of random garbage. So unfortunately that was the one bottleneck in my game where it required a lot of human work on my end.

A (14:57)

All right, so what about you, Matthew? Matthew? You did a, a life simulator or a credit card points simulator. It was actually pretty slick, I have to admit, but it was like real to life. So how much intervention did you need? Did you have to give it structure like Jacob did, like, you know, file layout structure and how much hand holding did you have to actually do?

C (15:24)

So truth be told, I kind of had this idea of running around in my head for a while because I'm familiar with the subject matter. I just played the points game for a while and I just got off of booking a complicated trip with it. So it was fresh top of mind. What I did was I had started with Claude, actually, the desktop app, and I used dictation a lot. I just rambled at it. All the ideas I had in my head in like a message or two, it said, I want to design a game around this and these are the constraints. I've got 24 hours. It needs to run in the browser because if you're going to do a game jam that's judged by your family, it needs to run on their computer. That's something we kind of ran into. And I ended up with like a thousand line game design document, which I don't recommend. It was way too long, but it was enough. Once I had gone back and forth with Claude, I read through the document repeatedly, keeping like an eye on the details and just made sure I had the details right. And this isn't really planned mode yet. I then took that document, put it in the root of a project in CLAUDE code and told it. I enabled the Team of Agents feature, which is kind of interesting. It's a token hog, but I'll kind of get there. And I explained I want to implement this game and that's kind of where the structure of it came from. I let the AI specify where to put everything. The constraints were that it needed to run on the web and it did that. It went with Vite and I believe React components. It went with Gosh, what is the name of. Found a framework for state management that allowed for the game to be resumable and save state easily, which was important. So what ended up happening was I set it off and it produced something that was almost functional. It builds a lot of the subsystems around the different complex components that you've got banks, you've got the different hotels and Airlines and all the point systems that they use. It's actually a really complicated game. You can have a job and you can get laid off at semi random. And it depends on the role you have. The role you have also depends on which cities you can go to. And each city has different airline hubs.

A (17:42)

It was much more complex than what you probably could have done without AI. There's no way you could have done this game in 24 hours without AI is what I'm hearing.

C (17:50)

And even if I had the models in my head, because I knew about a lot of these concepts and I explained it to the AI just from personal experience, being able to relate it to a real world model of something. And throughout development I used real hotels and real point systems and I just changed the names later. That was actually extremely effective because it didn't even have to search the web to get a general gist of what I was talking about. Like this is general knowledge I was modeling.

A (18:17)

Well, so, so there's one big tip, right? If you're modeling stuff it already knows about, it doesn't have to go and, and find that stuff out on the web to consume it. So David, let's switch over to your game a little bit. Your game was. Explain your game a little bit. And some of the things you ran into using, using the AI.

D (18:37)

Okay, so similar to Matthew, it was a simulator. Mine was a bit more of a sandbox. It was a world sim civilization simulator where civilizations would grow and die naturally. However, you can provide aid and all these different things depending on what they request. It's why I named the game Playing God. It's mainly just a ton of probability rules of what constitutes the growth of a civilization, what constitutes them. Gathering resources. And once they get big enough, they can riot against each other, break into multiple.

A (19:24)

So did you come up with these algorithms or did you ask AI to come up with these algorithms?

D (19:29)

I came up with all the percentages and then I pretty much wrote down a detailed MD file of exactly what I want the game to be. Exactly. It's why I used all of the agents, because I wanted it to end, be executed exactly how I said.

A (19:49)

So I see a common, I see a common thread between these. You guys all used your own. You could, you, you didn't just give it one idea and let it run. You had to interact with it. You used markdown language for the primary way of, of manipulating and controlling the AI. Does that sound right, Jacob? Does that?

C (20:12)

Yeah.

A (20:12)

Okay. All right, so big question is in this 36 hours, it ended up being about 36 hours or 30 hours, I guess. What did you learn the most on using generative AI? Because, I mean, this is pretty intense. 30 hours of interacting with an AI, doing work with it. A lot of learnings could come from this. So, Jacob, what would you say would be the biggest learning that you found from doing this?

B (20:49)

Sorry about that. Yeah. What I've found is that AI alone could not make me a good game.

D (20:56)

Right.

B (20:57)

I. I had to be the architect of that. It made me a very basic platformer, but it didn't understand the puzzles or the rules. I had to create those. It didn't have any direction with the art. It just created blocks for you to jump on. I had to tell it what I wanted the vision to be, and after multiple iterations, it was able to do it. So when people say that AI is here to replace us, to take our jobs, I think we just need to look at from the approach that this is to augment us. This is going to be a tool to allow us to make, you know, to execute our vision.

A (21:33)

So I like how you used augment, because, you know, I'm writing the book AI Augmented.

C (21:38)

Yeah, right.

A (21:39)

Becoming AI Augmented. So did you truly feel like, I mean, you guys play games, but none of you guys are game developers?

B (21:48)

Nope. No. I made a couple Python games back in college, but part of assignments, so. Yeah, was well outside my technical ability. I. I do quick scripting and programming, but that's for automation stuff. As electrical engineer, never have I programmed anything on this kind of scope before.

A (22:09)

Well, and then Matthew, you're a product manager. Program manager. Product manager. And your coding skills are. Back in college. You haven't really done a lot of coding since then, but you do have that product management bent. So when you look at Matthew's product, it is much more productized. For a better word, it's cleaner. It, you know, looks nicer than the others. David, you're a mechanical engineer freshman. Yours look like a mechanical engineer. I just say it. And Jacobs look like an electrical engineer wrote it. Because that's who you guys are, right? I had a. Matthew, what did you. What's that?

C (22:57)

Sorry. It's like Jacobs is fascinating from a technical perspective because, like, it's just brute forcing it. Like, all those pixels are really just JavaScript elements. Like, he just had the LLM just try and it felt like it was kind of beating up against the data model a bit. Would you say that?

B (23:14)

Yeah. I used nanovannana for two sprites. The rest I just asked throw, like, I put in Some dumpsters put in, you know, some city looking stuff, some lampposts. And it actually programmatically wrote me JS files for each of the sprites in my game. Likewise, I said, make me some music for my game, make it sound retro. And it brute force wrote me MIDI files from scratch. It didn't use any kind of audio generator model, it just wrote me some MIDI files. And actually it worked out like this. They sounded pretty good.

A (23:50)

So very, very much electrical engineer, right? I mean.

B (23:53)

Yep, yep.

A (23:55)

And then I. David, go ahead.

D (23:58)

I tried something similar myself. I did chatgpt for this and it created a script that made a wave file. It was all in Python. It just made a Python script that created a wav file. And like, without any knowledge of musical theory, it did an okay job.

A (24:26)

There you go. But just okay, not polished, right? I mean, so this, this is where, you know, having some subject matter expertise would really come into play. Would you guys agree?

C (24:41)

I think so. One of the things I really tried to focus on was making sure that I was aligning my game design with what LLMs do. Well, see, my game, at the root of it, it's data objects and text, and it's the kind of data that it can just generate. Like it knows about all the hotels in these areas. And even if it's not totally precise, it doesn't have to be. All these subsystems are just functions affecting data objects that are located in other parts of the app, like different modules. So I feel like I designed mine around AI's strengths, which is why it hit the level of complexity it did, which allowed me to focus more of my time on just playtesting it. Every time I would just fire it up, playtest it, look at what's wrong, have a task list. On the side of these are all the things that need to be changed. And it just became this kind of iterative thing. I didn't specify how the algorithms for, like, satisfaction should work immediately because I kind of trusted that maybe Claude would come with something interesting. And it did. And then I played it a few more times and oh, the character's losing after a month of inactivity. Well, that's not working. And it was just a very. It felt like I was working with a thinking partner is kind of how I often refer to it. I'm bouncing ideas off of it and it's bouncing ideas back.

A (26:01)

Did you find the same thing, David, or was yours more prescriptive, like, hey, you're going to go do these things?

D (26:09)

See, the thing is that with Opus tends to do that a lot more because I've been working with Claude a bit more and it tends to try and be more creative with. However, I had, like, detailed things of what I wanted to do and I had the agents, so it never really tried to push back or iterate on things itself.

A (26:33)

All right. Because you were, well, you were more prescriptive. You knew exactly what you wanted out of it. So you liked a month or instruct. You were like the master orchestrator and you were telling the first cello there, just do what I told you to do. I don't want your opinion.

C (26:49)

Where is.

A (26:49)

Is that how it works, JACOB PLAYS CELLO, by the way. So is that how it works, Jacob?

B (26:56)

Pretty much. Yep.

A (26:57)

Yep. All right, Matthew, you were going to say something.

C (27:00)

Well, I guess the. The difference in approach there, and there's a little bit of opus versus Sonnet, but it's more of a broad, like, how do you look at the problem? I knew about the Team of Agents feature and I had explicitly enabled and I told to use it, but I let Claude choose how to divide the work up.

A (27:22)

Two different. Two very different approaches there. What about you, Jacob? Where would you sit in between those two approaches? Or did you have a completely different approach to that?

B (27:31)

I think the challenge with my game was the spatial reasoning, which AI is. It's a big thing that's lacking. And still I occasionally try to get to do 3D CAD modeling for me, and it just cannot.

A (27:45)

It can't handle it.

B (27:46)

So with that bottleneck being map design, I was only able to make five levels. And what I did is I had the basic game down, all the mechanics were pretty much there. And then it was just iterative. I'd play through it. I'm like, huh? I was carrying this block and I went up in ELVIR with it and the block fell down. So then I would tell the AI what happened. I said, go fix it. And I'd be like, this level's really bland. You know, make this platform, you know, the rooftop of a building and make this, you know, add some more art into this. And it was so funny. I had some of the best models available to me. I had a month's worth of premium requests to use in 36 hours. I had up to, like, four requests going at once sometimes on different parts of my game. And I just felt like, still, like, I wasn't able to get as much in as I wanted with such a tight time constraint.

A (28:42)

So that's. That's interesting that you said that, because listening to you talk, AI is not going to replace us anytime soon. Right. Because we still need to do a lot of work. It can't do everything for us, but it sure can get rid of some of the grunt work for us that, you know, and the thinking. I, I, to me, as I, as I teach my students at Vanderbilt, I'm trying to teach them to think at a higher level, which all you guys have been able to kind of pull off with this, to think at that architect and orchestrator level. That is where we need to still sit and control these AIs. Would you guys say that's completely true or did you find some areas where. No, you still needed to get down in, deep in, into some of the coding.

B (29:32)

From a technical standpoint, the only things I ever did was just maybe shift a platform over, and that was literally just changing it from like 250 to 300, you know, that was the most programming I did. What I found is like, it couldn't replace my imagination in the solution. You know, I feel like that's where I have a competitive advantage over AI. It can code all day long, but it doesn't, it doesn't know what to code unless I can tell it. And I feel like that's where the real power is with AI.

A (30:05)

Yeah. What about you, David?

D (30:08)

I found that sometimes it would try and implement something in a weird way. So as long as you're very clear with your instructions, it does fine. But you do sometimes need to be very ultra specific. But the thing is, I've had problems with Gemini or ChatGPT when it comes to trying to be ultra specific. And the biggest problem with those is that request timeout. And that's something I really liked about Claude, is that it would work on something until it's done.

A (30:45)

Interest. Interesting. All right, Matthew, finish us off. Matthew.

C (30:49)

I have to say that it's a matter of what you ask for. If you have a clearly defined end goal and you have a problem. I like to scope things out in terms of problems and then the deliverable, but I don't specify the solution. I do that on purpose because sometimes I feel like the creative in my game was actually the funnier bits for Claude's idea. Some of them kind of evolved by accident. Like, I had this idea that, oh, you need to be able to buy stuff to hit the spend on your credit cards. And I was originally thinking, well, maybe it'll be like an Instagram feed or a shop. And eventually, just through convergence and revision, it looks like an ad on the side panel. And that's actually really funny. But all the listings in there, I didn't even give it specificity around the listings other than some should make you feel better, some should boost your happiness meter and some should hurt then. And Claude wrote everything else and picked up a great little vibe. Getting supplies for game night with your friends is a very happy thing. Buying some influencer starter kit for $500 is a really bad thing. It came up with all of those ideas itself.

A (31:58)

Yeah, that's pretty clever. So I really love that you guys showed how you're using the tool from different perspectives. Even though you're all my sons. Right. And even though you were all raised basically the same, your perspectives on how to use AI are all very different from each other, which I think is pretty clever. So hopefully you know the audience enjoyed this. When is the next Game jam and are gonna. Are more gonna be invited? Are you gonna open this up to the rest of the other siblings that you have, all seven of them or. Or is that too dangerous to do something like that?

D (32:38)

I'd. I'd be worried that they would accidentally spend all of their money on AI because they don't know how to use it.

A (32:48)

Oh, Jacob. Or what do you say Jacob to that?

B (32:54)

I'd love to do this every year we do on President say. I'm like, let's just make this the pulse for presidential Game Jam. I think it has a nice ring to it. Honestly. I think our siblings could do it. I think it's just a matter of convincing them that they can really. If anyone has a vision for a game, they can do it. It's just a matter of helping them realize it's available to everyone now. Anyone can do this kind of stuff.

A (33:24)

All right. What do you think, Matthew?

C (33:26)

I would love to do it in a year as far as anyone can do it. You know, it's an interesting. So I. The usages that we talked about, like Jacob used about a week of his month's worth subscription for 40 bucks and he felt like he couldn't find enough work for it. I was slamming through my 5 hour usage window on my hundred dollar plan in like two and a half hours pretty consistently up until the end when I was just doing small tweaks and a lot of that was the agent stuff. But being able to take the problem and structure it in a way that AI can solve it. I do actually think that there is some skill there and I don't know how I agree to develop it other than just using it obsessively and learning that at least for now.

A (34:11)

Well, I mean, maybe they should check out my new book coming out in Q3 this year, sometime this summer, called Becoming AI Augmented. Thanks boys for coming on the show. I appreciate it. I love having you guys on the show. This has been kind of fun. Thanks for listening to Embracing Digital Transformation. If you enjoyed today's conversation, give us five stars on your favorite podcasting app or on YouTube. It really helps others discover the show. If you want to go deeper, join our exclusive community@patreon.com com embracingdigital where we share bonus content and you can always connect with other change makers like yourself. You can always find more resources at embracingdigital. Org. Until next time, keep Embracing the Digital Transformation.