I'd say demos like these stand to profit the most from LLMs, if the goal is to make as much as possible in a few days: a barrage of quests are easy to generate, so are gear choices, and some skills for the initial 9 classes to pick from. A human would generally spend a lot of time here, thinking about whether the class/skill choices fit their world, what type of progression is fun and isn't. It's also where player testing would be important for a game to set good pacing and balance the difficulty.

Of course, the game itself is barely playable, it randomly stutters when I walk too far away from camp, the character controls are unintuitive, etc. A lot of this stuff could be chipped away by spending more time on the project and testing it yourself, getting a feel for what you want the game to be. That by itself should require a game to take more than a few days, if we expect others to play it and enjoy it. Something simple like movement controls could take many game iterations to iron out, and those aren't hard technical tasks.

Still, I can't entirely wrap my head around the fact that I live in a world where a machine can create this with minimal intervention by humans, and do a somewhat OK job at it, to the point where I'm willing to spend 10 minutes playing it.

Yea it's buggy and janky in places, but equally it's coherent from a 3d perspective and 2d UI one

I don't know how many people have been trying to get claude to vibe code games, but it's really not good at it, I've been fiddling about and trying to make it work for a few months now

Yes, you can make flappy bird or snake level stuff, but anything much bigger honestly just falls over after a while

And that's after lots of prompting and feedback

Now what I'm really interested in here is how much of this was "strongly steered" vs oh that's cool, let's do that, ie trying to "sculpt" and fit an artistic vision of some kind that the driver is envisioning vs liking what it outputs and just asking for more of the same

The distinction between using it as a tool vs just being excited to take whatever it gives you

The other thing I'm wondering is how much of this is strongly indicative of claude's capability with typescript, I personally don't use it, so that might be hampering me more than I realised

This may be an unpopular thing to point out, but to those that are saying it's blindly copying code, I'm pretty sure most of the Wow code on the internet wasn't written in typescript, so there's some transformation going on there, how meaningful it is I'm less that certain about

I'm beginning to feel like I just wasn't ambitious enough[0], I was thinking of this as a good opportunity to see if Fable was capable enough to teach an abstract skill like game design

Though I am happy that the code it's generated so far is actually quite tractable, ie it's been working hard to keep itself maintainable, which honestly is new from my perspective, not sure what other people's experiences have been, but I tend to find that agent's in general are just a bit too willing to increase LOC without enough in the way of features to justify that line count in my mind, it makes the juice just not worth the squeeze in my mind

-[0]: Here's what 45% gets you on a 5x plan (https://github.com/Folcon/inkstain-engine)

If you'd like to play the actual game with other SWEs I'd invite you to help me start a guild for SWEs: https://unplanned-downtime.com/

It's tale as old as time at this point: the LLM produced something sort of shaped like some other software, and did it in an impressively short amount of time, but it's basically impossible to bring a codebase made in this way up to production standards, or to maintain it in any reasonable way. Nobody's gonna pay money for this, or want to play it instead of Warcraft. Why should anybody care?

And here I am watching my 5-hour window disappear over a couple simple tasks in a CRUD app.

Repo: https://github.com/levy-street/world-of-claudecraft

hugged to death?

This is not how I'd design much of this. Does that matter? AI and whatever training data used seems to differ.

> nginx/1.24.0 (Ubuntu)

commenting so I remember to check again later when it's back.

I built this with Fable over a couple of days, on the side. It's a vanilla-WoW-flavoured micro-MMO in the browser: nine classic classes, three zones, a 5-player instanced dungeon, parties/duels/trades, and persistent characters. Free to play: https://worldofclaudecraft.com — and fully open source (MIT):

https://github.com/levy-street/world-of-claudecraft

Honestly the most mind-blowing part for me was how much it shipped that I never asked for. The level of polish and completeness coming out of the model genuinely surprised me — quest logs, threat metrics in the combat log, eating/drinking, spirit release on death.

We already have some contributors on GitHub!

Some people will see this and think "wow, in 5 more years I'll actually be able to make World of Warcraft". Some will see this and think "Wow, I can make World of Warcraft now with 1/100th the cost and engineers". Neither of these thoughts are right, though. The reality is more like "Wow, someone can make a game 10 times as good as World of Warcraft for the same[1] cost and number of engineers".

[1] - roughly $63 million, 5 years, 60 engineers

Pure vibe coding pretty much limits the max scope of your project to the capability of the agent, and if you're trying to make something real and maintainable, you'll always end up with something worse than the person who guides it at a lower level.

That being said, there are certainly going to be systemic/built in ways to have coding agents generate coherent architectures. Just it isn't really baked in right now.

Probably an ideal model with current tech is the main coding agent thread that implements things, and a secondary thread that constantly compacts/prunes and cleans up according to specific style rules.

Problem with doing both in same context via AGENTS.md is the LLM often won't adhere to all rules if it's overloaded with many competing instructions. But you can scope the code styling agent prompt tightly and get much higher focus and adherence.