Quarter Past Dead

Qwen 3.6 27B running with locally configured Claude CLI was used to generate the initial scene, then turn it into an FPS, exactly as Bijan has done many times. Nothing special there.
1st Prompt:
"Design a 3D subway scene you can walk around in. It should look nice, but be focused on performance. This might eventually be a level for an FPS: it should have extremely low GC pressure."
2nd Prompt:
"Turn this into a simple zombie survival FPS survival game. Focus on core mechanics and especially performance: use pooling for everything you can to lower GC pressure to support a horde-shooter-like vibe. Make it easily extensible if we want to add more enemy types. It should feel fast and fluid like a modern FPS, have the ability to aim down sights, and a win condition. Use your searxng, fetch, and git tools and explore agents to research best practices, look at examples, and get ideas. You can also use your js-sandbox tool to test concepts/math/physics implementation. It should have appropriate music that does not require an external asset load (except for trusted CDNs). Compose the song yourself based on music theory, or reference an appropriate soundsheet for an existing song and make a sound-alike. You may also call your AceStep 1.5 tool for stems if you need inspiration for your midi composition. It should have a stage hazard the player must effectively use to survive."

What's probably more interesting is, once the base game/physics/ideas were in place with 27B dense and 4000-5000 lines started being a little much for it to handle, I handed the project off to the 35B A3B MoE, which could do max context (~240K) on my 5090. It was easily able to iterate/bugfix on the existing base, and over around ~70 additional iterative prompts later, the game was "finished". I could have kept going, but was exhausted, and these types of games have been done to death, but this was a very fun learning experience.