Claude Code Has a GameDev Problem transcript

So, I launched my demo. It's been a very interesting experience trying to figure out what's the right balance between managing a demo versus the whole YouTube thing. I really wanted to talk about I've noticed that there's been a lot of new developers who have been posting on YouTube talking about LLMs. I feel like there's been a lot of confusion because there have been vibe coders who try to oneshot a full game and on the surface it looks like a complete highquality game um until you look further into it.

I've watched a lot of videos talking about this and I think a lot of them kind of missed the point. And when you're using an LLM, what an LLM does when you ask it, hey, is this a good game design or is this is this fun? It just looks at reported data online or is training set. And I I can only speak for myself, that's not how I actually design a game. How I decide if something goes into my game is by simply playing the game and asking myself, did I have fun?

And LM do not do that. There has been zero progress on this. But they don't work like a human. If I was working with a human, when something breaks or something goes is going great, the person on the other end is going to give me real human emotional response. Game developers are passionate. Artists are passionate. LLMs are not passionate. Does my calculator have emotions? Right? Telling an LLM to, "Hey, go play this game and tell me how it made you feel is going to give you random, useless information because we're making games for other people, humans, to experience." And I have a very strong emotional response to the game that I'm making.

I care about it. The LLM could care less. And sometimes it is a little bit annoying because in LLM it it could trigger me. sometimes and I know I've heard some developers who get bothered by calling LLM tools but I think it's very important to understand that I use that terminology to remind myself of its limitations its blind spots have never been improved I think 4.5 was probably the turning point that brought a lot of attention from real developers I don't think the models have gotten much from my workflow.

You can put me back on 4.5, remove all this weird distillation warnings. My experience would be roughly the same. The the models do not just improve just because the model goes uh the version goes up. There are certain tasks that I'd rather just downgrade my cloud model down to Opus 4.6. Um I can say for certain that the LMS are becoming worse teachers. Learning is becoming much harder to do. But I can say for cloud code, I've noticed that if I ever ask the right question or ask it to, hey, that didn't make sense.

Can you dig deeper? I've been kicked out of the model for distillation warnings. If I'm talking to Fable, it just does this thing where it's like summarized. I'm like, what are you what are you doing? It's almost offensive in the domains that it's working, it literally can't do it. And for me, when it hides its work and summarizes it for me and does absolutely dumb things on the back end, it's extremely frustrating. There's a lot of things it's not better at.

There's a lot of things it's never improved on and it needs me to help guide it. But if it hides too much of his thinking or summarizes, it's doesn't matter if you've gotten smarter on the stuff that makes a fun demo reel or it makes a good oneshot prompt on Twitter. But if I'm trying to actually make a real game. So if I'm trying to not make AI slop and the models as they get smarter, they make it harder to squeeze the slop out.

It gets very frustrating. So I've had moments where I actually drop the model down into older versions to improve the quality of the output. I'm pretty sure there are certain areas where I can go to 3.7 a version of claude that was pretty terrible for a lot of things and still get the same performance or better performance. A model like this should never think for 40 minutes and not give me any feedback and just dump its output.

I am trying to gain design and I'm trying to steer it. If it doesn't improve, I mean, obviously I'm going to go to local like dumber models that are run locally. Doesn't matter if you can snap your finger and get instant perfect code, instant perfect art. If your bottleneck is game design, you're going to be sped up, but it's very minimal. So something that would take you four, five years, maybe now takes you three, maybe four years.

There are real problems with LLMs when it comes out to games, making making good games that I have not seen any noticeable improvement on. Now, I may be wrong on this, and I know that there are people that watch my videos that know way more about this stuff than I do, and I'm really curious what you think. Is what I've been vibe noticing, is that actually true? But I do think that one thing to keep in mind is that the oneshot prompts that you see are oftentimes done by expert VIP coders.

And I say expert, I just mean people who've been doing it for a long time. Just because you went on to Claude Code or Astra or Fable or whatever the latest model is, and you did the same prompt as someone else, just because the output you got is not the same quality, it might not have anything to do with your prompt. I've noticed, this is just me, you know, vibe guessing here. that the models are usually the best versions of themselves the first like one or two maybe three days of releasing a new model.

It's not consistent. You just don't know what you're going to get any day. The quality is so inconsistent. I just feel like a lot of these LLMs, the way they're designed, I'd much rather have a four like a Opus 4.5 quality model run locally that doesn't have all these gimmicky weird stuff that happens on whatever server they're running. The thing that I learned, the biggest surprise that I've had using LLMs is that the area where the LLM is the weakest ended up being the area that I became the strongest having had no experience prior.

When I started vibe coding, I was like, "Okay, I'll make the art and then the the code will be handled by the LM." I didn't really think about game design too much. At this point, I've been struggling with game design for nearly two years. In order to progress to this point where I have a demo out and it's positively received, game design is something that I had no idea that I was going to learn so much about and I had zero help in.

The bottleneck naturally became the area where I had to focus because there was no other way to overcome it. There is areas that you're going to learn a lot. And if you're doing gamedev and you're a brand new game developer, get ready because you're going to learn a lot about good game design. If you're successful on making a good game art, you can buy assets for you can maybe vive code some of it. For coding, basic generic coding, which most games are, can be viodated.

I'm not doing anything novel when it comes down to code. You probably aren't, and that's fine. the game design that is 100% bottleneck. You're probably not going to get any speed boost. If anything, you're going to be slowed down by using LLMs. It's going to take longer because the LLM is going to make some suggestions that are terrible, and you're going to play test your game. It's going to be poorly received because of it, and then rinse and repeat until the game is good.

I had to redo my game one full time. If you look at the old videos, it used to be a card game where you played it on the characters. It's no longer that. You don't assign cards onto characters. It's just an auto battler and that was because of play testing and getting absolutely destroyed in the play test reviews. Now, if you go on the demo, obviously I got all that stuff resolved before I launched the demo, but it was not easy.

Game development is changing, but it is still very difficult in ways that LLM cannot help with. I've been really trying my best to make a good game as an actual vibe coder. there is still so much involvement that to hear people talking about, oh, you don't need to take time. You can just pump out a game and it's good. One, it it kind of makes me feel bad cuz I'm like, what am I doing here? I've been using this with intention for almost two years now.

Not quite two years, but close. Either I'm doing something wrong. As somebody who sees the potential of the technology very early on, I I love it a lot. It just doesn't align to what I've experienced. This idea that I can just oneshot something in two weeks and it's going to be good and wellreceived on Steam is crazy to me. I I can't even imagine showing you my prompts. How much does a year and a half of full-time prompting get you every single day pretty much?

And even then, my game's not done. If LLM's ever solve the game design problem, we have way bigger issues. Okay? LLMs have never made any improvements on game design. If you're watching this and you think otherwise, please let me know. Gamedev is different and extremely still very hard.