Four AI-Assisted Game Prototypes for Live Streaming: From Guessing Games to 3D Farms

The Huajiao (花椒) team explored AI-assisted game development across four projects tailored for their live-streaming and voice-chat platform. Each attempt combined AI tools — primarily Codex for code generation — with game engines (Cocos, Three.js) and 3D modeling (Blender) to move from idea to playable prototype, then evaluated the path to production.

01 / Everyone Guesses Together: AI in Gameplay and Development

The team built a social deduction game where the system picks an item and players ask questions to narrow it down within 10 turns. The AI answers “yes / no / uncertain” based on the hidden answer. Questions come from a preset pool; the program tracks turns and win/loss.

AI assisted with gameplay design, UI design, code implementation, and model integration. After launch in party rooms and solo-broadcast rooms, the game saw hundreds of rooms opening daily. Real usage surfaced issues: question repetition, difficulty calibration, and answer smoothness — all becoming optimization targets.

02 / Pool: Physics and Match Rules

Pool is a common casual mode in voice rooms. The team focused on ball physics and eight-ball rules. Developers defined scope and technical approach, then used Codex to write and modify code with automated verification. Shot mechanics, collision, pocketing, turn switching, fouls, and win/loss judgment were all implemented. Rendering and input use the Cocos engine; physics and rule code are shared between web and client to keep behavior consistent.

Code runs, but feel requires playtesting. Based on feedback they adjusted table size, aiming during power charge, and foul prompts so players can see clearly and operate more easily. The game is now in the test environment integrated into party rooms, with multi-room support and ongoing verification of full matches and edge cases before a trial launch.

03 / Ghost Defense: From Prototype to Business Launch

This tower-defense style game requires players to accumulate resources, upgrade defenses, and build turrets to withstand waves. Resource allocation decisions directly affect the outcome, making it more complex than the previous two.

Similar to pool, developers used Codex with Cocos to produce a mobile prototype. Playtesters could join rooms, gather resources, tap empty floor tiles to place turrets, and inspect range and stats. Attackers spawn at the door and reduce its durability. The prototype validated core loops; full match flow and numerical balance still need work.

The images and video above show the self-developed exploration prototype; the live version uses a third-party solution.

To meet delivery timelines, the team chose a third-party SUD (Software Unified Delivery) solution for the initial launch, concentrating resources on platform integration and release. The self-developed prototype remains as technical groundwork for future in-house evaluation. The third-party title “Ghost Don’t Catch Me” launched on September 29 in Huajiao chat rooms, with hundreds of rooms opening on day one. The team completed platform integration and business adaptation — gift-triggered skills, game rewards, room UI — and is now in live-experience optimization and operational feedback collection.

04 / 3D Scene: Converting a 2D Farm to 3D

The team was familiar with planting, watering, and harvesting loops from an existing 2D farm. They asked: can we make a walkable, perspective-switchable 3D version while keeping core gameplay?

Developers used Codex to drive Blender for modeling and Three.js for scene and interaction , then refined camera, occlusion, and movement controls. The result is a demo for experiencing AI-assisted modeling and 3D interaction; it has not progressed to production.

Key Takeaways

AI turns ideas into playable versions faster. A working prototype lets the team and business discuss fit, required adjustments, and whether further investment is justified.

Prototype to production still demands full engineering effort. Room integration, networking, feel tuning, and testing must be counted in the schedule — prototype speed alone is misleading.

Self-research and business delivery can run at different cadences. Choosing between in-house development and third-party integration depends on business goals, delivery window, and team capacity; exploration experience feeds future projects.

When AI directly drives gameplay, reliability matters. Model answer quality, exception handling, and invocation cost need continuous observation in real usage.

Continuing Iteration in Production

Going forward, the team will share further experience tuning, technical implementation, and delivery lessons from “Everyone Guesses Together” live feedback and the pool launch process. They will also keep tracking advances in real-time generated games.