When we started exploring how to create games with AI in March 2025, the question was never:
"Can AI write a game?"
It was:
"Can one person with taste ship something people genuinely want to play — not just more AI slop?"
Today, we're open-sourcing our answer: two Sharky Skills for Claude Code and Codex, giving creators free 3D asset credits, real multiplayer infrastructure, and one-command publishing.
We're sharing them because we want to find the best AI-native game creators — and build a community for people who believe AI can bring something valuable and original into the world, instead of squeezing their work into feeds and platforms never built for them.
The video attached is what this looks like in practice: a multiplayer 3D horror escape-room story game, only codex + sharky skills. @thsottiaux@ChatGPTapp
It's an early cut — we're looking for horror-narrative creators to co-develop and sharpen it with us. If dark storytelling is your thing, DMs are open.
Try the skills. Build one small game. Publish it. Send us the link.
The era of boutique AI games starts with creators — and grows through community.
You bring the idea. We bring the skill and studio support: curation, player testing, streamer collaborations, and dedicated development resources.
Each quarter, we'll select the strongest games and help bring them to Steam or the App Store.
Come build with us 👇
In 2024, i handed a profitable company back to my ex-partner , bought out our investors, and started over from zero in AI.
Most people thought we were crazy. Two years later, here's where that stubbornness led — and an invitation.
we're mostly from ex-Tencent and overseas ugc platfrom series-founder . We've been obsessed since ChatGPT dropped in 2022. In 2023 we were the annoying friends telling every founder and investor we knew: this is the platform shift.
So in 2024 we burned the boats.
2025: bootstrapped, no shortcuts. We review every ai consumer side — AI comics, AI short dramas, AI companions, pro-consumer agents, interactive content, AI-native TikTok formats.
Everyone gave us the same advice: pick a story that's easy to fund. Show revenue fast. Show DAU.
We just couldn't make ourselves do it.
By late 2025 we placed our real bet: coding + LLMs. Build the products that demand the most intelligence and creativity, and build a team that evolves as fast as the models do.
Here's the belief that kept us going:
The last era was won by network effects — more users, more data, more ads.
The next era, the one AGI is pulling us into, will be won by an intelligence flywheel: smarter products → richer experiences and creativity for users → greater value captured → even smarter products。
So even when our user count was embarrassingly small, we didn't downgrade the ambition or pivot to something easier. We gritted our teeth and kept going.
By 2026 we were convinced that games are the purest test of consumer AI coding. A game is code + systems + art + story + physics + multiplayer + taste, all at once. So we built a UGC platform for AI-made multiplayer games.
Then reality taught us two things:
Foundation models improve faster than any proprietary agent layer we could bolt on top.
Easier creation ≠ better games. Most tools still can't get you real mechanics, coherent art direction, actual replay value.
So last month we did something that stung a little: we took the game-generation agent we'd spent months polishing, and opened it up as a Skill anyone can use.
It runs in Claude Code and Codex — and as of now, we fully support Grok Build too. Go show it some love.
We're not betting on one model. We're betting on creators. Whatever tool you love, bring it. More great tools = better games for everyone.
Our job focus on testing the best models and share the best users case ,skill ,taste and belief.
Our belief is simple: greater intelligence should be turned into more wonderful products for people
Honestly, the community is the part we care about most.
This year we talked to a lot of game creators. Some hate AI. Many use it quietly and are afraid to say so — because on itch, Steam, Roblox or ArtStation, admitting it can get you dogpiled. Some genuinely fear it will take their work. And the flood of AI slop makes every one of those fears feel justified.
That's exactly why we think creators need a new home.
X has already become that home for AI creators in general — genuinely, thank you X and @elonmusk for that. It's the one place where people building with AI can trade notes in the open. We want to do the same for one corner of this ecosystem: the AI-native folks making games.
Not a place that applauds anything AI touches. A place that filters hard for taste, originality, and people who actually want to ship something worth playing. A place where you can openly compare notes — which models are best for code, 3D assets, game UI, art direction, storytelling — and playtest wild ideas in front of real players, fast.
We're not building a content factory.
We're looking for the small group of AI-native creators who want to make games that used to be impossible for one person or a tiny team.
And we don't think those games will look like the old ones. The traditional game market probably doesn't have much room left. But the new, gamified one — playable AI-native experiences, whole categories nobody has named yet — is still wide open. That's where the interesting stuff will emerge.
We push the frontier of intelligence to redraw the frontier of creativity together.
🔥🫰
Discord: https://t.co/ZCOgvgVDUm
https://t.co/KXMrGMUywq
#sharkgame #vibejam #grok
Looks great, keep working on it. Same happened to me 5 days ago. I was commenting to a game about a feature many people wanting for years and some imbecile told me I've no idea what I'm talking about. Now I'll show them what I'm talking about in couple of weeks :)
I'm a 35+ years experienced developer not a game developer but a developer now AI and free open source tools will make my dream.
This will change the gaming sector a lot, I'm waiting for great games and great ideas come to life now. Gates are down.
Take a guess: Was this 3D airforce game built with Codex, or Fable ? ✈️
After watching countless fighter jet rankings, aircraft breakdowns, and air-combat videos, I wanted to turn that experience into an actual playable 3D game.
today I used Sharky Skill + Three.js together with one AI coding tool to build: Skyforge Arena game.
What surprised me most was not that the model could generate a fighter jet. It actually attempted to reconstruct the aircraft’s individual components, internal structure, weapon systems, and physical relationships.
i can inspect different parts of the aircraft and explore its firepower. When you enter BATTLE mode, switch into the cockpit, track enemy aircraft, and engage in aerial combat.
The game also attempts to simulate aircraft positioning, flight attitude, automatic target tracking, radar and laser detection, and the physical feeling of a high-speed air-combat environment.
Aircraft and military enthusiasts might really enjoy this one and plz give me idea to prompt better version.
Play it here: 👉 https://t.co/Dj4pfypGb1
Build your own multiplayer 3D browser game with Codex or Claude Code:
Multiplayer and browser publishing Skill: https://t.co/H4EIKL1KGx
3D asset library: https://t.co/Jw9Qdm1pNU
3D and AI game creators are welcome to join us: Discord: https://t.co/ZCOgvgVDUm
Website: https://t.co/KXMrGMUywq
#codex #threejs #fable #vibejam #sharkygame #webgame
share a little experiment these past two days: gave Codex (GPT-5.6 Sol) and Kimi K3 (@Kimi_Moonshot) the same game design doc, the same prompt, and the same two agent skills — and had each build a multiplayer 3D web game.
Verdict first: Codex trusts the spec. Kimi trusts its own eyes.
and if you are entj ,plz leave kimi ,it will make you feel anxiety.
Codex is like a senior engineer who's been freelancing for years: fast, stable, delivery means done. In its world, acceptance criteria all green = finished. No overthinking, no over-testing, no performing. The cost was you need to know what you want ,especially the visual taste part.
Kimi is an obsessive QA freak crossed with a visual-first frontend artist: nothing counts as done until "I've seen it with my own eyes." It treats screenshots as ground truth and will happily burn tens of thousands of extra tokens to verify frame by frame and take so many times ,you need to be patient . Its visual instincts are clearly a cut above — I half suspect it was fed a pile of proprietary frontend training data. The cost is just as obvious, long runs are unstable. It blacks out mid-task and needs you need to patient.
But the process was way more fun than the conclusion — let me unpack 👇
Setup took one step. I pre-wrote a design doc for a horror cruise-ship escape game, fed the same prompt to both, and installed the same two skills:
· shark-game-assets — free 3D GLB generation via MCP; characters come rigged with idle/walk animations
· sharky-online — turns a single HTML file into a real online game (accounts, rooms, server-ordered sync bus); verify your email once, then every build auto-uploads to the server with a shareable link
Then I just sat back and played the client.
Codex barely gave me a chance to speak. Read the SKILL.md, went straight to work, delivered in 1h25m: a solo mode with an Echo NPC that holds mechanisms for you, 2–4 player co-op with an ending vote, 25 server-side rules — not one failed. Straight to production. My total feedback: more character variety, less stiff animations. That was it.
Kimi's run was way bumpier — and way more entertaining.
It opened with parallel subagents: one burned 53 tool calls just generating assets, another ran 80 on QA. Then the rigging service happened to be down, so it auto-degraded to static models — and my killer chased people in a permanent T-pose, arms wide open, like it wanted to give you a hug. The local preview flat-out died (a PATH issue + an ESM require bug + a zombie process, the full combo). Buttons wouldn't click; after three rounds of me going "still can't click," it tore the whole thing down and went zero-click: if the sync bus doesn't connect in 6 seconds, the game auto-starts in solo mode. You never touch the mouse.
But the fixing is what won me over: regenerated rigged models; compressed the GLBs from 4.37MB to 2.06MB with gltf-transform; attached the fire axe precisely to the killer's hand bone; then spun up Playwright, screenshotted idle/walk/strafe frames, and inspected them — and caught, by eye, that its own facing calibration was flipped 90°. A model debugging itself with screenshots. First time I've ever seen that.
It's also the one that froze for 10 straight minutes — twice — and got killed by the system: "no accepted raw events for 600000ms — cancelled."
Each time I had to manually tell it to continue. My guess: the single turn was too heavy, the context too long, and the tokens just couldn't keep burning.
One last easter egg: B model keep shipping character that glides along the floor, arms raised, sliding forward — that's not walking, that's low-altitude cruising. I'm not saying which one. Guess in the comments😂
🎮 Both games are live, free, solo or co-op:
【TRIANGLE ESCAPE】
👉 https://t.co/hSoSySIVLG
【Triangle Loop】
👉 https://t.co/FUlaG0vqtv
🛠️ Assets → build → test → publish → auto-hosting: one pipeline, and out comes a shareable multiplayer web game. Both skills are open source:
· Multiplayer + publishing: https://t.co/ofUuSjjZHu
· Free 3D assets:
https://t.co/TlsI4BHjY3
Can't wait to see more 3d multiplayer games with our skill 🔥
#codex #sharkygame #kimi #threejs #vibejam #multiplayer #3d
I challenged the new model to build a short 3D multiplayer escape game and it started extremely fast just 20-25 minuted to generated a mvp,but painful either, here was my process
The overall scene and multiplayer logic came together quickly using Three.js. But the real pain came during the 3D asset stage.I used an external 3D model generator + custom rigging/skinning workflow.
That part was genuinely painful — models kept turning static, facing the wrong direction, losing all animations, or behaving completely broken after import. I had to go back and fix the same issues over and over again.
After tons of iteration on details, story beats, interactions, and fixing these rigging nightmares… we finally got a playable version.
Still a lot fun and i have tried like 20 times and hadn't successed on board the ark. If you wanna get the ticket ,you can try here.
【Doomsday Ark: The Last Ten Seconds】
https://t.co/9xK05Q6wDC
you want to try our workflow yourself:
• 3D Asset Generation Skill (free): https://t.co/Jw9Qdm1pNU
• Multiplayer Hosting Skill: https://t.co/3itWJ5Rpdl
question: Does this feel more like it was made with Fable or Kimi K3? Drop your guess 🔥
👇
@threejs
#threejs #kimik3 #kimi #fable #vibejam #sharkygame #3dgames
🚀 Build real multiplayer web games — no backend, no Supabase, no hassle! Just use Claude, Cursor, or Codex + this skill
https://t.co/H4EIKL1KGx
You can turn your game idea into a real playable multiplayer web game in minutes. No server setup needed! #WebGame#HTML5Game #Multiplayer #ClaudeAI #IndieDev #claudecode #vibejam
When we started exploring how to create games with AI in March 2025, the question was never:
"Can AI write a game?"
It was:
"Can one person with taste ship something people genuinely want to play — not just more AI slop?"
Today, we're open-sourcing our answer: two Sharky Skills for Claude Code and Codex, giving creators free 3D asset credits, real multiplayer infrastructure, and one-command publishing.
We're sharing them because we want to find the best AI-native game creators — and build a community for people who believe AI can bring something valuable and original into the world, instead of squeezing their work into feeds and platforms never built for them.
The video attached is what this looks like in practice: a multiplayer 3D horror escape-room story game, only codex + sharky skills. @thsottiaux@ChatGPTapp
It's an early cut — we're looking for horror-narrative creators to co-develop and sharpen it with us. If dark storytelling is your thing, DMs are open.
Try the skills. Build one small game. Publish it. Send us the link.
The era of boutique AI games starts with creators — and grows through community.
You bring the idea. We bring the skill and studio support: curation, player testing, streamer collaborations, and dedicated development resources.
Each quarter, we'll select the strongest games and help bring them to Steam or the App Store.
Come build with us 👇
Grok 4.5 @grok vs. Kimi K3 @Kimi_Moonshot but I’m not revealing which is which yet. 👀
Inspired by Christopher Nolan’s upcoming film The Odyssey, I gave both models the exact same prompt:
Reconstruct the Trojan Horse and the ancient geography of Troy as a playable 3D interactive experience using @threejs
One took under 20 minutes.
The other took nearly 100 minutes.
The results are A and B above.
Which one is Grok 4.5, and which one is Kimi K3?
Drop your guess:
A = ? B = ? and which model took only 20 minutes, and which took nearly 100? I’ll reveal the answer soon. 🦈🎮
My first time to use the supergrok plan after this testing
#TheOdyssey #ChristopherNolan #Grok45 #KimiK3 #ThreeJS #AIGames