Each typed element in the design describes detailed aspects of the game. This is the primary "Context" element. Just select an element to see it's schema/data.
"Baron is an arcade biplane game set in a WWI-inspired, cartoon world. The player builds their aircraft in the hangar, takes off from an airfield, flies over a procedurally generated landscape under procedurally generated clouds, and lands again. There is no combat. Pillars: (1) Flying feels great: quick to learn, satisfying to master. (2) A readable, cheerful cartoon look: toon shading, ink outlines, bold colours. (3) Every flight is different: a seed generates terrain, clouds and airfields. (4) Instant play: opens in a browser tab with no install."
Try pasting the raw diagram into an agent, here is the prompt...
You won't get the identical game but all the specified details will be there in the end result. If you swap out some key dependencies... Unreal Engine vs three.js, you will get some wild result. Have fun and post your results. :)
kqr 9 hours ago [-]
I'm surprised how well this handled. I usually have large complaints on the flight dynamics in projects like these, and for this (in expert mode) they were not that large. Of course, it's no X-Plane, but I'm trying to keep in mind what sort of project it is. In a good blend of arcade and simulation, it should be a toy that is easy to get into, but someone who has simulator experience should also feel at home and not run into counter-intuitive dynamics.
I suppose I am a little surprised the biplane was able to loop on its own power from level flight.
Interaction-wise, I would have liked
- to be able to invert the Y axis on the mouse control! Nuts that this alternative is not included.
- a vertical speed gauge all the time, not just when the simulator thinks I'm about to land.
- go-arounds to be handled more gracefully, rather than as a very long landing.
- a trim control on the mouse wheel to set the airspeed. (And then the mouse control could keep returning to centre, at least in expert mode.)
- a toggle for opting out of metric units, and instead using more customary knots for speed, feet for altitude, feet per minute for vertical speed etc.
- pre-tuned navigation radios (or even just VFR charts) as an optional replacement for arrows in the sky.
With these changes (and possibly more I think of later) this could be a really fun way to get children into flight simulation, since it's so action packed.
diverted247 4 hours ago [-]
The goal was a light, fun, open, introductory flight sim. My kids wanted to try a loop, so I added flaps and added 25% more power to biplane. Just enough.
I liked this idea, but I would have hoped it would let me fold a plane and it would simulate its flight. When I try, it seems like I'm just picking from the preset 24 designs. Am I missing something?
rossng 7 hours ago [-]
I have experimented with custom folding already. Technically speaking it will work just fine, nothing about the physics is prebaked.
However I quickly realised it's a non-trivial UX problem and needs a bit more effort to make something worthwhile. It will come later!
bita_nidir 6 hours ago [-]
I'm watching!
rukuu001 11 hours ago [-]
Nice work :)
suprjami 5 hours ago [-]
That's very cute.
Current LLMs are great for making little throwaway HTML toys like this. I do it too.
theragra 29 minutes ago [-]
I made a whole Novuss game in 20 mins (Baltic billiards, a type of pool game with flat pawns instead of balls)
It inspired me so much, I intend to create a real game and maybe even release it on steam. Weird, but there are no 3d implementations yet, afaik.
blaze33 4 hours ago [-]
First WebGL game I tried where my joystick works out of the box, very nice!
I knew it was possible (Gamepad API [1]) but never bothered to code it for my WebGL experiment (that was before ai) although I did buy again the same joystick I had as a kid, the sidewinder precision pro [2], which is certainly the oldest usb device I still use today (released ~30 years ago).
As a flight simulation fan since playing Sopwith years ago, I really liked this!
Good flight feel, nice graphics, nice typography!
Feature suggestions:
+ Missions to deliver letters to random points on map.
+ Damage modelling so you get more points the more pieces remain attached to the plane on landing.
+ Trim control on arrow keys (up arrow for elevator down, down for elevator up, left for rolling left, right for rolling right).
+ Option to have obstacles like bad weather, birds, kites, other planes, including ones armed with paintballs or other projectiles that cause localized damage on your plane or obscure your view from behind the plane with paint or smoke.
+ Ability to move camera around the plane with mouse and zoom in and out with scroll wheel.
+ Better sound overall, especially less repetitive sounds for engines.
Using the Ctrl key for a common control (here: decelerate) in a WebGL game is bad design. This game also makes use of W, which is asking for accidents.
aflag 19 hours ago [-]
I think the interesting to know would be how much it cost to build and how much prompting was necessary. Overall it doesn't look like anything that I wouldn't be able to get by just prompting claude in claude code
diverted247 4 hours ago [-]
The initial build was $10 of Opus 5.5 tokens from the design.dai specification.
This is a one or close to one shot Claude code; it even looks like all the 100s of games I have been vibing while on the subway.
Some of them are incredibly good by the way; I have one that’s better than the commercial ones of that genre in the AppStore.
simonjgreen 3 hours ago [-]
Love it! I'm especially impressed how well this works on mobile. My feedback:
- Don't auto correct roll back to level (or make it a toggle)
- Ability to change units (eg f vs m)
- Toggle to disable the auto-rudder
- Charts and some for of radio nav as an option rather than the path through the sky
I like the choice of licensing, might have a poke around as this is a really cool little project :)
pampas 7 hours ago [-]
It's really fun but please please add a legend of the keyboard shortcuts I can see while playing.
em3rgent0rdr 13 hours ago [-]
CTRL-S (slow engine while going up) causes my browser to try to save the webpage, fyi. Cute graphics and feel.
KeybInterrupt 11 hours ago [-]
That is exactly what killed me ^^ ...I was trying to land the Plane and well I could not save it...
lazybean 3 hours ago [-]
Looks cool!
I'll need to try again, I failed to take off.
I notice the back flap is going into the road.
piloto_ciego 13 hours ago [-]
This was fun. I would like the ability to trim and retract the gear on the bigger machines other than that 10/10.
dvh 19 hours ago [-]
Make rotating phone dialog skipable.
niwrad 12 hours ago [-]
This is fun. However, I couldn't figure out how to bring the throttle back down.
rukuu001 11 hours ago [-]
CTRL
swiftcoder 9 hours ago [-]
Or you can use the number keys to select a throttle point directly
sheepscreek 13 hours ago [-]
I don't know if OP founded dai.studio, but if they did - they would do well to post the DAI/AI graph that generated this game. This will allow people to get a taste for their tool without needing them to invest their time (lower the entry barrier).
I'd love if this generated terrain based on google maps and a position
wahern 7 hours ago [-]
This reminds me of Corncob 3D, the DOS shareware game. All it lacks are aliens to shoot at.
modinfo 8 hours ago [-]
Interesting! make fun! also nice to have /?debug with settings :D
1 hours ago [-]
chiengineer3 4 hours ago [-]
Chitown26.com
footoverhand 13 hours ago [-]
Very fun, thanks for posting! Daughter really interested in it :-)
diverted247 4 hours ago [-]
does she have a feature request?
nodesocket 18 hours ago [-]
Ok this is actually pretty awesome. Avid Flight Simulator player. While this clearly is not that, it's surprisingly fun to play with a hint of realistic flight.
leetrout 19 hours ago [-]
Like an updated corncob 3d! Source available anywhere?
note: iOS in particular is really slow to start up these.
Maybe it's all AI slop. None of them are fun games, they're just tech demos.
The interesting things are they are all small. All assets are built at startup. They aren't using any libraries. That suggests there is lots of low-hanging fruit for quick to start web-games.
Telling the LLM to build from scratch and follow AAA practices frees the LLM to use whatever is appropriate, rather than whatever some library dictates. The LLM was also good at optimizing, at least to an extent. It would run experiments, turning features on/off to find where the speed was going. It would also do A/B screenshots to see how much changing techniques effected the output.
Each demo from prompt to first output took 2-3 hours and then a few prompts in my free time to clean up the most outstanding issues.
One thing that helped a bunch was telling the LLM to model cars using NURBS instead of polygons. Another was telling it to have a separate modeling agent, and a separate judging agent.
None of the demos are perfect. They got to a point where I was happy that they showed want I wanted to see. I wasn't planning to take them to perfection. The gen-city-auto demo has the most issues. While it made a giant city, there are lots of places where a road makes no sense. I could prompt through that but all I really wanted to see was a GTA like city in a small amount of code and for that it proved the concept is valid. Given ways for a designer to declare what they wanted where I could see the concept generate much more "designed" cities.
suddenlybananas 2 hours ago [-]
The aquarium one seems to crash firefox for me.
sgt 10 hours ago [-]
> Maybe it's all AI slop. None of them are fun games, they're just tech demos.
Sure, probably, but as long as you're having fun, and you get to experiment at a velocity (due to vibe coding) that is probably unprecedented.
I also think there may be some inherent problems with just supporting games in WebGL/wasm etc, because you're still within the constraints of a browser. You still have to pass through a JS glue layer for every WebGL call, for example, not to mention the data marshaling costs.
The cool thing is that you can reach literally anyone with a web browser, however.
briancarterio 17 hours ago [-]
[dead]
schuhwerk 18 hours ago [-]
[dead]
aislophater 16 hours ago [-]
Oh wow, ai slop that is barely the same quality as the first iphone apps. This site is garbage. Thanks to AI the talentless finally have a voice. Whoopie.
peddling-brink 15 hours ago [-]
Oh wow, a low quality comment denigrating the doers. Thank you for your service.
ceejayoz 4 hours ago [-]
Isn't the criticism that they aren't really a doer here?
suddenlybananas 2 hours ago [-]
At least the author posted their prompt here.
throwawaybackag 6 hours ago [-]
The comment tone isnt great but they have a point. I really dont think having an idea and getting Claude to implement it counts as being a "doer". The barrier to entry for doing something like this with an LLM is so low now. It seems obvious to me that these vibe coded projects are worth much less attention than a pre LLM equivalent that you know took the author a lot of time and effort. All these posts are basically now "I had an idea and wrote a prompt". It's great if folk have fun with this but it seems pointless sharing this stuff if anyone can do it in literally less than 10 minutes with a 20 USD sub.
theragra 22 minutes ago [-]
Not everyone. I found that if you want quality, you need to spend a lot of time ironing out issues.
Sure, prototype works in 20 minutes. Then, 10 nights later I still can't get my game to have really good ai opponent. Even with MCTS, genetic algorithm, a lot of prototypes and 30 cores for the game to use.
Source: https://github.com/dai-studio/ex-baron (BSD 2-Clause)
Each typed element in the design describes detailed aspects of the game. This is the primary "Context" element. Just select an element to see it's schema/data.
"Baron is an arcade biplane game set in a WWI-inspired, cartoon world. The player builds their aircraft in the hangar, takes off from an airfield, flies over a procedurally generated landscape under procedurally generated clouds, and lands again. There is no combat. Pillars: (1) Flying feels great: quick to learn, satisfying to master. (2) A readable, cheerful cartoon look: toon shading, ink outlines, bold colours. (3) Every flight is different: a seed generates terrain, clouds and airfields. (4) Instant play: opens in a browser tab with no install."
Try pasting the raw diagram into an agent, here is the prompt...
"Build https://dai.studio/ted/ex-baron/raw"
You won't get the identical game but all the specified details will be there in the end result. If you swap out some key dependencies... Unreal Engine vs three.js, you will get some wild result. Have fun and post your results. :)
I suppose I am a little surprised the biplane was able to loop on its own power from level flight.
Interaction-wise, I would have liked
- to be able to invert the Y axis on the mouse control! Nuts that this alternative is not included.
- a vertical speed gauge all the time, not just when the simulator thinks I'm about to land.
- go-arounds to be handled more gracefully, rather than as a very long landing.
- a trim control on the mouse wheel to set the airspeed. (And then the mouse control could keep returning to centre, at least in expert mode.)
- a toggle for opting out of metric units, and instead using more customary knots for speed, feet for altitude, feet per minute for vertical speed etc.
- pre-tuned navigation radios (or even just VFR charts) as an optional replacement for arrows in the sky.
With these changes (and possibly more I think of later) this could be a really fun way to get children into flight simulation, since it's so action packed.
Great list, I will add these in time.
https://github.com/dai-studio/ex-baron
License: BSD 2-Clause, feel free to submit PRs.
https://www.rossng.eu/paper-flight/
However I quickly realised it's a non-trivial UX problem and needs a bit more effort to make something worthwhile. It will come later!
Current LLMs are great for making little throwaway HTML toys like this. I do it too.
It inspired me so much, I intend to create a real game and maybe even release it on steam. Weird, but there are no 3d implementations yet, afaik.
I knew it was possible (Gamepad API [1]) but never bothered to code it for my WebGL experiment (that was before ai) although I did buy again the same joystick I had as a kid, the sidewinder precision pro [2], which is certainly the oldest usb device I still use today (released ~30 years ago).
[1] https://developer.mozilla.org/en-US/docs/Web/API/Gamepad_API... [2] https://en.wikipedia.org/wiki/Microsoft_SideWinder#Precision...
Good flight feel, nice graphics, nice typography!
Feature suggestions:
+ Missions to deliver letters to random points on map.
+ Damage modelling so you get more points the more pieces remain attached to the plane on landing.
+ Trim control on arrow keys (up arrow for elevator down, down for elevator up, left for rolling left, right for rolling right).
+ Option to have obstacles like bad weather, birds, kites, other planes, including ones armed with paintballs or other projectiles that cause localized damage on your plane or obscure your view from behind the plane with paint or smoke.
+ Ability to move camera around the plane with mouse and zoom in and out with scroll wheel.
+ Better sound overall, especially less repetitive sounds for engines.
+ “Roguelite” progression where you get to keep unlocked perks, with random missions and events (https://hypertexthero.com/roguelite-flight-sims/)
+ Mission and plane editors.
+ Controller support.
I’d buy this on Steam or Itch.
The source is here: https://github.com/dai-studio/ex-baron
BSD 2-clause, PRs welcome.
Some of them are incredibly good by the way; I have one that’s better than the commercial ones of that genre in the AppStore.
- Don't auto correct roll back to level (or make it a toggle)
- Ability to change units (eg f vs m)
- Toggle to disable the auto-rudder
- Charts and some for of radio nav as an option rather than the path through the sky
I like the choice of licensing, might have a poke around as this is a really cool little project :)
I'll need to try again, I failed to take off.
I notice the back flap is going into the road.
Source: https://github.com/dai-studio/ex-baron
See design.dai, plan.dai, progress.dai
Can you add rudder and elevator trim?
https://news.ycombinator.com/item?id=50503397
Explains how to fly, and links to the repo
> Press F to lower Flaps for Take Off
> Press SHIFT to Throttle Up
> At 50km/hr, Down arrow to take off.
> Arrow keys are control stick.
https://www.reddit.com/r/vibecoding/comments/1xwdvhf/im_a_hu...
Here's mine:
https://greggman.github.io/gen-city-auto/ (430k, 4 cites, tons of vehicles, trains, signs, water, bridges, rain, road details) - video for highlights: https://x.com/greggman/status/2117605120745881692
https://greggman.github.io/cyber-web-city/ (200k, 10 cities themes) - video for highlights: https://x.com/greggman/status/2109451367941378213
https://greggman.github.io/webgpu-driving/ (170k - 10 environments, several vehicles, tunnels) - video for highlights: https://x.com/greggman/status/2104533425478930568
https://greggman.github.io/webgpu-aquarium/ (102k - several fish, different ocean styles)
note: iOS in particular is really slow to start up these.
Maybe it's all AI slop. None of them are fun games, they're just tech demos.
The interesting things are they are all small. All assets are built at startup. They aren't using any libraries. That suggests there is lots of low-hanging fruit for quick to start web-games.
Telling the LLM to build from scratch and follow AAA practices frees the LLM to use whatever is appropriate, rather than whatever some library dictates. The LLM was also good at optimizing, at least to an extent. It would run experiments, turning features on/off to find where the speed was going. It would also do A/B screenshots to see how much changing techniques effected the output.
Each demo from prompt to first output took 2-3 hours and then a few prompts in my free time to clean up the most outstanding issues.
One thing that helped a bunch was telling the LLM to model cars using NURBS instead of polygons. Another was telling it to have a separate modeling agent, and a separate judging agent.
None of the demos are perfect. They got to a point where I was happy that they showed want I wanted to see. I wasn't planning to take them to perfection. The gen-city-auto demo has the most issues. While it made a giant city, there are lots of places where a road makes no sense. I could prompt through that but all I really wanted to see was a GTA like city in a small amount of code and for that it proved the concept is valid. Given ways for a designer to declare what they wanted where I could see the concept generate much more "designed" cities.
Sure, probably, but as long as you're having fun, and you get to experiment at a velocity (due to vibe coding) that is probably unprecedented.
I also think there may be some inherent problems with just supporting games in WebGL/wasm etc, because you're still within the constraints of a browser. You still have to pass through a JS glue layer for every WebGL call, for example, not to mention the data marshaling costs.
The cool thing is that you can reach literally anyone with a web browser, however.