GAME LAB / CASE 01 / DS4FLASH0731
DeepSeek V4 Flash 0731 led the build through DS4. Qwen 3.6 A35B joined as a vision subagent, turning visible failures into targeted fixes.
Starting with one Unity3D car asset, the agents created a racing game for the Godot engine in one initial prompt, then corrected it with two short follow-ups.
PLAYABLE GODOT WEB EXPORT
Open the agent-generated Low Poly Racing build in a new tab. The compressed WebAssembly game downloads only after you choose to launch it.
THE PROMPT LOG
The initial request supplied intent and one asset. Each follow-up described only what was visibly wrong.
Initial build
“Use this Unity3D car asset to create a racing game in the Godot game engine. Add a track, player controls, a race HUD, and a reset flow.”
OUTCOME The first playable scene and game structure were created.
Follow-up one
“The car does not have wheels and the player controls do not work. Add the wheels and make the car controllable.”
OUTCOME The visual asset assembly and input path were repaired.
Follow-up two
“Fix the reset button. The directions are reversed: A currently moves right and D moves left. Change the controls so A moves left and D moves right.”
OUTCOME Reset returned to a clean state and the final key mapping matched the player’s expectation.
MODEL ROLES
Worked through DS4 to plan the scene, assemble the project, wire gameplay, and apply follow-up corrections.
Supplied visual inspection so the agent loop could reason about missing wheels, scene state, and the visible result—not only the source files.
GAME LAB / CASE 02 / QWEN38-27B
Qwen3.8 27B built Drift Rush with xhigh reasoning—and thought way too much. The run took more than eight hours to complete the game.
It ran as UD-Q4_K_XL with a 163K context window, mmproj CPU offload, and a homemade harness designed for smooth compaction across the long agent session.