CASE STUDY
2026.11.06–11.08 · Aholo 3D AI 黑客松 · 个人项目 · AI 协作Nov 6–8, 2026 · Aholo 3D AI hackathon · Personal project · AI-assisted
从六面世界到最后一块地：一个 3D 游戏是怎么被“删”出来的？
在做《最后一块地》之前，我其实想做过很多完全不同的游戏：3D 贪吃蛇、六面棋局、冰火人式双角色解谜、荒诞办公室、空间消消乐、乐高式建造，甚至是一整个可以旋转和探索的六面宇宙。
这些想法各有吸引力。问题是，我并没有足够的时间和成熟的游戏开发经验，去把它们全部变成好玩的产品。参加群核 Aholo 3D AI 黑客松时，我需要在三天里完成一款真正可以运行的游戏。
我逐渐发现，困难的并不是让 AI 帮我实现某个想法，而是决定：这么多想法里，究竟哪一个值得被实现？
最初，我们一直在寻找一个足够有趣的游戏
最早，我考虑过生活技能教学游戏：玩家在三维场景里操作洗衣机等物品，完成现实生活中的任务。它有一定的实用价值，但我很快产生了疑问——如果普通 AI 助手已经可以解释这些操作步骤，为什么用户还要专门打开一个游戏？
于是我开始研究那些无法单纯通过聊天完成的体验。从贪吃蛇到森林冰火人，从植物大战僵尸到乐高和棋类，我想知道：把经典游戏放进三维空间，究竟能改变什么？我对六面立方体尤其感兴趣：玩家可以在六个表面移动，旋转世界，改变路径，甚至改变每个空间的运行规则。
但在进一步检索已有游戏和开源项目之后，我们发现，六面贪吃蛇、立体棋局和可旋转魔方棋盘都已经有人做过。
一个概念看起来新奇，不代表它没有先例；一个项目使用了 3D，也不代表它拥有新的游戏机制。
我们需要的不是一个越来越大的世界，而是一条真正有意思、能够实现的规则。
为什么最后选择了经济学？
我本科读的是经济学。过去做 AI 项目时，我经常尝试用技术解决某个具体问题；这次我想反过来，能不能把一个抽象的经济学概念，变成玩家可以亲手操作的游戏？
我首先想到的是稀缺性。土地有限，预算有限，想要满足的需求却很多：建设工厂，城市可以获得更多经营收益，但附近居民可能受到污染影响；建设公园，生活环境可能改善，却需要占用土地和预算；选择商业设施，又会带来不同的收益结构。
这些关系不需要复杂的故事，就能形成一个基本矛盾：当资源不够满足所有需求时，我们应该怎样选择？
于是，《最后一块地》有了明确的形状：它不再是一款宏大的城市模拟游戏，而是一道可以亲手操作的空间经济谜题。
我们最终保留了什么？
第一版游戏只有一块 5×5 的土地、80 点初始预算和四类建筑：工厂、住宅、公园、商店。玩家需要放置一家工厂、两栋住宅，并在公园和商店之间二选一。
不同建筑的位置会影响结果：工厂的污染沿固定风向扩散，公园能够提高附近住宅的满意度，商店则带来更高的经营收益。玩家可以旋转观察三维地块，放置和移动建筑，查看收益与满意度变化，再决定是否提交规划。
我们刻意没有加入复杂 NPC、道路交通、人口增长和动态市场——不是那些功能没有价值，而是想先完成一条完整的游戏循环。
目前有两种可以通关的方案：
| 发展方案 | 居民满意度 | 经营净收益 | 综合评分 |
|---|---|---|---|
| 公园优先 | 90 | 28 | 73 |
| 商店优先 | 85 | 42 | 83 |
这两个结果也暴露了一个值得继续研究的问题：当前评分更偏向商店方案。两种方案虽然都可通关，却不意味着它们已经拥有足够平衡的策略价值。把两种选项做出来，不等于已经设计好了真正的选择。
AI 帮我们实现了什么，又没有替我们决定什么？
这次项目使用 Aholo Lux3D 生成四类核心建筑模型，再通过 Three.js 和 TypeScript 将它们接入 Web 游戏。模型不是简单摆在网页里的装饰：它们可以被玩家选择、放置和移动，并与建筑占地、位置关系和经济规则关联。
开发过程中，我也借助 Codex、Zcode 等 AI 工具参与代码实现和界面迭代。但我希望把两件事区分清楚：AI 可以帮助生成资产、编写代码、解释错误和加快调整；它并没有自动证明某个玩法值得做，也不能证明玩家一定会觉得有趣。
工程上，我们把占地、预算、污染和评分等规则与三维渲染分开实现，使模型替换不必直接改变经济计算。项目测试记录包含 104 项自动化测试通过结果。
这说明一些程序行为获得了验证，却不能直接说明游戏已经具有足够的乐趣、教育效果或长期价值。
过去我经常问自己：当 AI 写出了我不完全理解的代码，这算不算我的能力？这次项目让我更在意另一个问题：我是否知道自己要实现什么，为什么这样设计，以及应该怎样判断它有没有做好？
为什么我不想把这次项目写成成功故事？
因为它还有明显的不足。
目前只有一个固定关卡，玩家理解规则以后，可能很快找到相对稳定的高分布局。土地和预算的限制虽然存在，但还没有形成特别强的空间压力。3D 让建筑关系更容易观察，也让 AI 生成资产成为实际游戏的一部分；不过现阶段核心经济计算依然建立在二维网格上，我们不能声称这是一款充分利用三维高度和物理机制的游戏。
而且，自动化测试通过不等于玩家喜欢。
我希望把这些不足保留在项目记录里，而不是为了让作品看起来成熟，提前给出未经验证的结论。
下一步，不是更多建筑
在研究 Block’hood、Concrete Jungle、TerraScape 和 Frostpunk 2 时，我发现它们分别从资源依赖、空间顺序、建筑组合和利益冲突出发，建立了自己的策略体验。《最后一块地》不可能也不需要一次实现这些系统，但它们提示了后续可能的方向：不同城市委托、建筑之间的协同关系，以及更多利益主体。
最后一块地，也是一次关于取舍的练习
回头看这次开发过程，我发现它本身和游戏的主题有点相似：我们拥有很多想法，但时间、精力和技术能力都是有限的。选择六面世界，就可能没有时间打磨具体玩法；选择复杂经济模拟，就可能无法完成稳定的 Demo；选择一个足够小的空间经济谜题，则意味着主动放弃许多看起来更酷的功能。
这也是一次资源配置。
我还不能确定《最后一块地》以后会发展成什么样的游戏，但至少现在，我们有了一款可以打开、操作、失败、调整并重新尝试的作品（在线试玩），也留下了为什么选择它、为什么放弃其他方案的记录。对我来说，这些判断和过程，与最终的游戏同样值得保存。
土地有限，选择有代价。
From a cube world to The Last Plot: a 3D game made by deleting
Before The Last Plot, I wanted to build many completely different games: 3D snake, six-sided board games, fire-and-ice dual-character puzzles, an absurd office, spatial match-three, Lego-style building — even an entire rotatable, explorable six-faced universe.
Each idea had its appeal. The problem was that I did not have the time, or the mature game-development experience, to turn all of them into fun products. For the Manycore Aholo 3D AI hackathon, I had three days to finish a game that actually runs.
I gradually realized the hard part was not getting AI to implement an idea. It was deciding: among so many ideas, which one is worth implementing?
First, we kept searching for a genuinely interesting game
My earliest idea was a life-skill teaching game: players operate objects like washing machines in a 3D scene to complete everyday tasks. It had some practical value, but I quickly ran into a question: if an ordinary AI assistant can already explain these steps, why would anyone open a game to do it?
So I started studying experiences that cannot be replaced by chat. From snake to Fireboy and Watergirl, from Plants vs. Zombies to Lego and board games, I wanted to know what actually changes when a classic game moves into three dimensions. The six-sided cube interested me most: players could walk on six surfaces, rotate the world, change paths, even change the rules of each space.
But after searching existing games and open-source projects, we found that six-sided snake, 3D chess boards, and rotatable cube boards all existed already.
A concept looking novel does not mean it has no precedent; a project using 3D does not mean it has a new mechanic.
What we needed was not a bigger world, but one rule that is genuinely interesting and actually buildable.
Why economics in the end?
I studied economics as an undergraduate. In past AI projects I usually tried to solve a concrete problem with technology. This time I wanted to reverse it: could I turn an abstract economics concept into a game players operate with their own hands?
Scarcity came first. Land is limited, the budget is limited, and the demands exceed both. Build a factory, and the city gains operating revenue while nearby residents may suffer pollution. Build a park, and living conditions improve — at the cost of land and budget. Choose a commercial facility, and the revenue structure changes again.
These relationships need no elaborate story to form a basic tension: when resources cannot satisfy every demand, how do we choose?
That is how The Last Plot took shape: not a grand city simulator, but a spatial-economic puzzle you can operate by hand.
What did we keep in the end?
The first version has only a 5×5 plot of land, an initial budget of 80 points, and four building types: factory, residence, park, and shop. The player must place one factory and two residences, then choose between a park and a shop.
Position changes the outcome: factory pollution spreads along a fixed wind direction, a park raises the satisfaction of nearby residences, and a shop brings higher operating revenue. The player can rotate the 3D plot, place and move buildings, watch revenue and satisfaction change, and then decide whether to submit the plan.
We deliberately left out complex NPCs, roads and traffic, population growth, and dynamic markets — not because those features lack value, but because we wanted to finish one complete game loop first.
Two strategies currently clear the level:
| Strategy | Resident satisfaction | Net operating revenue | Composite score |
|---|---|---|---|
| Park first | 90 | 28 | 73 |
| Shop first | 85 | 42 | 83 |
The results also expose a question worth studying: the current scoring favors the shop route. Both routes are winnable, but that does not mean they already carry balanced strategic value. Building two options is not the same as designing a real choice.
What AI implemented for us — and what it did not decide for us
This project used Aholo Lux3D to generate the four core building models, then Three.js and TypeScript to bring them into a web game. The models are not decoration placed on a page: players can select, place, and move them, and they connect to footprints, spatial relations, and the economic rules.
During development I also used AI tools such as Codex and Zcode for code implementation and UI iteration. But I want to keep two things separate: AI can generate assets, write code, explain errors, and speed up adjustments; it has not automatically proven that a mechanic is worth building, nor that players will find it fun.
On the engineering side, footprints, budget, pollution, and scoring are implemented separately from 3D rendering, so replacing models does not touch the economy calculation. The project’s test records include 104 automated tests passing.
That verifies some program behavior. It does not establish that the game is fun, educational, or valuable over time.
I used to ask myself: when AI writes code I do not fully understand, is that still my ability? This project made me care more about another question: do I know what I am building, why it is designed this way, and how to judge whether it works?
Why I don’t want to write this up as a success story
Because it still has obvious gaps.
There is only one fixed level. Once players understand the rules, they may quickly find a stable high-score layout. The land and budget constraints exist, but they do not yet create strong spatial pressure. 3D makes building relationships easier to observe and made AI-generated assets part of the actual game; still, the core economy runs on a 2D grid, and we cannot claim this is a game that fully uses 3D height or physics.
And automated tests passing does not mean players like it.
I want to keep those gaps in the project record instead of presenting unverified conclusions so the work looks mature.
The next step is not more buildings
While studying Block’hood, Concrete Jungle, TerraScape, and Frostpunk 2, I found each builds its strategy experience from a different source: resource dependency, spatial ordering, building combinations, and conflicting interests. The Last Plot cannot — and does not need to — implement all of these systems at once, but they point at possible directions: different city commissions, synergies between buildings, and more stakeholders.
The last plot is also an exercise in tradeoffs
Looking back, the development process itself resembled the game’s theme: we had many ideas, but time, energy, and technical ability were all limited. Choosing the six-sided world might have left no time to polish gameplay; choosing a complex economy simulation might have left no stable demo; choosing a small spatial-economic puzzle meant deliberately giving up many cooler-looking features.
That, too, is resource allocation.
I still cannot say what The Last Plot will grow into. But right now we have a game that can be opened, operated, failed at, adjusted, and retried (live demo) — plus a record of why it was chosen and what was given up. To me, those judgments and that process are worth keeping as much as the game itself.
Land is limited. Every choice has a cost.