I built a game where every frame of footage is AI video, and you can actually play it 🎮
16 shots from Wan 3.0 → wired together like animation states → aim, dodge, counter, boss phase 2
No game engine. No 3D. Rendered frame by frame in 4K.
Built & designed by K28 🧵
4
4
244
I created this 3D world for $0 in direct model charges using OpenCode, DeepSeek V.1 Flash, and a custom design system with skills I built for AI agents.
The result is Seven Worlds, One Street: a neighbourhood where different places share a visual language while keeping their own character.
The experiment was about more than the final video.
I wanted to see what happens when a capable, lower-cost model is given a system to work with: visual principles, reusable scene-building tools, examples and skills that give the agent direction through the design and production process.
Instead of asking the model to invent every part of the workflow from scratch, I gave it a foundation—and room to make decisions within it.
The result reached a level of finish I’m excited about, without using the frontier-model credits I might otherwise have spent on the production. That isn’t a benchmark or a claim that the models are equivalent; it’s one experiment in how the surrounding workflow can affect what a model produces.
Building the design system and skills took work. But that investment creates something I can reuse and improve across projects.
I think more creative and production workflows need this kind of alternative: systems that make quality more accessible while helping people choose where expensive models are most valuable.
The goal isn’t to limit what an agent can do. It’s to give it enough direction to use its capabilities well.
#CreativeTechnology #AIProductDesign #OpenCode #CreativeCoding #HumanAIInteraction
1
171
What if you could explore a codebase like a city?
I used D4, a design system I built, to turn the Flask repository into a playable 3D world. A knowledge graph guides the city’s layout, while different visual styles change its character.
An experiment in AI-guided design and human–AI collaboration.
48
What if the next iPhone had NO SCREEN?
I built a $10 AI-native design engine and challenged it to prototype exactly that.
Meet TACTUS.
No touchscreen. AI-generated interactions. Gesture-driven controls.
Not just sci-fi renders—working interactive prototypes.
The future of HCI?
99
Replying to @BreezeOg1
The final output took around 30 mins to render the output. But no qcs were required. As the first output only I got the perfect result. So it was worth the wait
5
I built an entire interactive 3D architectural world for $0.
No Blender. No expensive AI subscriptions.
Just DeepSeek V4.1 Flash + OpenCode + a custom design system and agent skill I built.
The real breakthrough? Teaching AI agents my design language, not just prompting them.
Watch this. 👇
2
2
63
This isn’t a 3D LoRA explainer video.
It’s an interactive web app built by an AI agent I taught my design system.
GLM-5.3-Flash + Tavily. 3D pipelines, architecture, metrics, interactions.
Experiment cost: $0.
What if AI learned your design taste—not just your prompts?
1
91
I built my own AI-native design engine for just ~$10.
Then challenged it to prototype Tactus: an AI-powered device with NO touchscreen.
Not sci-fi renders. Real interactive components, research-backed systems & working logic.
Is this where human × AI interaction is heading? 👀
6
82
Replying to @nicopez
I will say both are the crucial step but even shot planning goes in the skill step. So the full workflow efficiency is what matters the most
3
Replying to @nicopez
The skill setup is the challenging task, as the full pipeline depends on it. Need a quite a few rounds of testing. But once done. U have a agentic studio that supports mass production for quality content
7
Replying to @AlexshevPm
Right Alex. Editing is the most difficult part to manage, as good output also depends on how well the final editing was handled. And a repeatable workflow can actually make the process much smoother. Which can also be an advantage over commercial tools. Specially over the ones which doesn’t allow such skills creation on their platform
1
19
Replying to @LongBlingAmt
The anime workflow was easier to setup as compared to the full agentic workflow
7
I built my own AI filmmaking studio.
Then created a 2-minute cinematic anime music video for just $10.67. 🎬
No expensive AI subscriptions.
Claude Code + custom agent skills + Wan 3.0 + WaveSpeed API.
91% usable clips.
The future of AI filmmaking isn’t just better models.
It’s better workflows. 🚀
10
7
238
