AI 做游戏Vibe Gaming

跟 AI 聊天把游戏做出来Make a game by talking to AI

这个站只回答一件事：AI 已经改变了做游戏的哪部分，还没改变哪部分。This site answers one question: which parts of making a game has AI actually changed, and which have it not.

核心判断The core claim

AI 让做游戏的成本下降是不均匀的：写代码降了 70~80%，做美术 / 配乐 / 关卡几乎没降。The cost of making a game with AI falls unevenly: code drops 70–80%, while art, music and level design barely move.

于是「一个人做游戏」的瓶颈从「写不出来」变成了「凑不齐」。So the bottleneck for a solo developer shifts from "cannot write it" to "cannot assemble it".

本站怎么组织How this site is organised

按做游戏的实际顺序分五阶段，每阶段固定四块：做什么 · 怎么做 · 常见失败 · 我们的口径。Five stages in the order you actually build a game. Each stage has four fixed parts: what, how, common failures, and our reading.

选引擎Pick an engine哪个引擎对 AI 最友好Which engine AI handles best
核心循环The core loop怎么把玩法变成代码Turning mechanics into code
手感调参Game feel为什么 AI 给的参数不对Why AI’s parameters feel wrong
资产合规Asset licensing美术 / 音频从哪来，哪些能商用Where assets come from and what is usable
导出发布Ship it怎么让人玩到Getting it in front of players

→ 进入学习路径（含完整教学与常见失败）→ Enter the learning path (full walkthrough and common failures)

表 1 · AI 改变了做游戏的哪部分Table 1 · What AI changed about making games

⚠ 「成本变化」一列是**我们的判断**，不是实测数据 —— 这个领域没有权威统计。判断依据写在第三列。⚠ The "cost movement" column is **our judgement, not measured data** — no authoritative statistics exist for this. The reasoning is in the third column.

环节Stage	成本变化Cost	成熟度Maturity	判断依据Reasoning	性质Nature
引擎搭建与架构Engine setup & architecture	↓ 大幅down a lot	成熟mature	场景图、状态机、ECS 都有官方文档与大量可参考实现；AI 生成的是结构清晰的样板代码，不依赖引擎私有约定。Scene trees, state machines and ECS all have official docs and many reference implementations; what AI writes is boilerplate with clear structure, not code depending on engine-private conventions.	可回溯官方源traceable to official sources
玩法原型（成熟品类）Gameplay prototype (established genres)	↓ 大幅down a lot	成熟mature	平台跳跃、卡牌、塔防、贪吃蛇这类玩法的核心循环是确定性代码，AI 能可靠产出。The core loop of platformers, card games, tower defense and the like is deterministic code, which AI produces reliably.	部分可核验partly verifiable
手感与调参Game feel & tuning	↓ 部分down partly	半成熟half-mature	AI 能给参数（重力、摩擦、跳跃高度），但「好不好玩」无法自动判定 —— 没有测试能告诉你手感对不对。AI can supply parameters (gravity, friction, jump height), but whether it feels good cannot be judged automatically — no test tells you that.	我们的判断our judgement
2D 美术资产2D art assets	↓ 少down little	半成熟half-mature	单张/单组素材可用；成套资产的风格一致性仍是生成模型的缺陷。多轮对话不会让它更一致。Individual assets or small sets are usable; a coherent art style across a full set is still a weakness of image models. More dialogue does not make it more consistent.	我们的判断our judgement
3D 美术资产3D art assets	↓ 很少down barely	不成熟immature	角色/场景的一致性、拓扑可用性、贴图一致性都难以保证；且商用授权存在法律灰区。Character and scene consistency, topology usability and texture coherence all remain hard; commercial licensing is also a legal grey area.	我们的判断our judgement
音频（BGM / 音效）Audio (BGM / SFX)	↓ 部分down partly	半成熟half-mature	单点音效与短BGM 可用；长段配乐的风格稳定性不足。Single sound effects and short BGM work; style stability across longer pieces does not hold up.	我们的判断our judgement
关卡与数值设计Level & numeric design	↓ 少down little	半成熟half-mature	AI 能提出数值方案，但平衡需反复试玩迭代，而试玩成本无法压缩。AI can propose numbers, but balance needs repeated playtesting — and playtesting time cannot be compressed.	我们的判断our judgement
剧情与对白Story & dialogue	↓ 大幅down a lot	可用（需校对）usable with editing	这是 LLM 的本职。⚠ 但仍需人工校对（事实错误、语气一致性、分支收敛）。This is what LLMs are for. Still needs human editing: factual errors, tone consistency, branch convergence.	我们的判断our judgement
导出与发布Export & release	↓ 少down little	成熟mature	引擎自带导出流水线，AI 帮助有限 —— 这块是确定性工程，不是生成任务。Engines ship an export pipeline; AI helps little here — this is deterministic engineering, not a generation task.	可回溯官方源traceable to official sources

表 2 · vibe coding 的哪些前提在游戏里不成立Table 2 · Which premises of vibe coding do not hold for games

vibe coding 的方法论不是万能的。这张表列出它依赖的四个前提，以及游戏开发中哪些不成立。vibe coding's method is not universal. This table lists the premises it relies on, and which of them fail for game development.

vibe coding 依赖vibe coding assumes	在游戏开发中In game development	后果Consequence
反馈周期以秒计Feedback in seconds	以分钟到小时计 —— 你得玩二十分钟才知道这个跳跃手感对不对Counts in minutes to hours — you have to play twenty minutes to know whether the jump feels right	迭代节奏慢一个数量级，「快速试错」的优势被削弱Iteration is an order of magnitude slower, which eats most of the fast-try-error advantage
结果可自动验证（类型检查、测试、编译）Results are automatically verifiable (type checks, tests, compile)	「好不好玩」无法自动判定 —— 没有一个测试能告诉你这个游戏有趣Whether it is fun cannot be judged automatically — no test can tell you this game is enjoyable	⚠ 最关键的一条：AI 能写出通过所有测试但不好玩的代码⚠ The key one: AI can produce code that passes every test and is not fun
失败会报错Failures raise errors	失败常常是沉默的 —— 游戏能跑、能通关、只是不好玩Failure is often silent — the game runs, it is completable, it is just not fun	你可能以为做完了，实际上只是没发现问题You may think it is done when you have only failed to notice a problem
代码是唯一的产物Code is the only output	代码只占产物的一部分 —— 美术、音频、关卡、平衡都是产物Code is only part of the output — art, audio, levels and balance are outputs too	⚠ 瓶颈从「写不出来」变成「凑不齐」⚠ The bottleneck shifts from "cannot write it" to "cannot assemble it"

表 3 · 什么适合、什么不适合Table 3 · Where it fits and where it does not

⚠ 这张表是**本站的判断**，不是数据。它的用途是给读者一个起点，不是标准答案。⚠ This table is **the site's judgement**, not data. It is meant as a starting point, not an answer key.

适合Fits

Game Jam（48 小时等短周期）Game Jams and other short time-boxes时间盒强制你放弃资产 perfectionism，AI 的代码提速正好补在刀刃上A time box forces you to drop asset perfectionism, and AI's speedup lands exactly where it helps
成熟品类的原型验证Prototyping an established genre核心循环是确定性代码，AI 可靠The core loop is deterministic code, which AI handles reliably
玩法驱动的独立游戏Gameplay-driven indie games资产需求低，玩法是卖点Low asset demands — the mechanics are the product
教学、可视化、内部工具Teaching, visualisation, internal tools不需要美术品味，产出是功能不是体验No artistic taste required; the output is a function, not an experience

不适合Does not fit

3A 级项目AAA-scale projects资产量与打磨需求远超 AI 能覆盖的范围Asset volume and polish demand far exceed what AI can cover
依赖独特美术风格的作品Work that depends on a distinctive art style风格一致性正是生成模型的弱项Style consistency is precisely where image models are weakest
需要大量手工调优手感的产品Products needing heavy manual game-feel tuning「好玩」无法自动验证，AI 帮不上判断这一环"Fun" cannot be verified automatically, so AI cannot help with that judgement
强竞技的多人在线游戏Competitive multiplayer online games网络同步与反作弊是确定性工程 + 安全问题，不是生成任务Netcode and anti-cheat are deterministic engineering and security problems, not generation tasks

附·引擎的可核验事实Appendix · Verifiable engine facts

→ 按「对 AI 的友好度」看这五个引擎的判断与依据→ See the five engines by how AI-friendly they are, with the reasoning

下面每一条都取自 GitHub 官方 API，快照时间 2026-10-05。上面三张表是判断，这张表是事实 —— 两者不要混读。Every row below comes from the official GitHub API, snapshotted 2026-10-05. The three tables above are judgements; this one is fact — do not read them as the same kind of claim.

仓库Repository	类型Tier	脚本语言Scripting	★	许可Licence	最后提交Last push
`godotengine/godot`	通用 2D/3Dgeneral 2D/3D	`GDScript / C#`	118,127	`MIT`	2026-10-04
`bevyengine/bevy`	原生 3Dnative 3D	`Rust`	48,611	`Apache-2.0`	2026-10-04
`libgdx/libgdx`	Java 跨平台Java cross-platform	`Java / Kotlin`	25,422	`Apache-2.0`	2026-09-24
`pixijs/pixijs`	Web 2Dweb 2D	`TypeScript`	48,284	`MIT`	2026-10-03
`mrdoob/three.js`	Web 3Dweb 3D	`JavaScript`	116,227	`MIT`	2026-10-04
`cocos/cocos-engine`	2D/3D 跨平台2D/3D cross-platform	`TypeScript + C++`	9,845	`NOASSERTIONNOASSERTION`	2026-09-21
`love2d/love`	2D 轻量lightweight 2D	`Lua / C++ / Fennel`	8,796	`NOASSERTIONNOASSERTION`	2026-09-20
`raysan5/raylib`	2D/3D 轻量lightweight 2D/3D	`C`	34,971	`Zlib`	2026-10-04

⚠ cocos/cocos-engine 与 love2d/love 的许可在 GitHub 上是 NOASSERTION —— 机器无法判定，须读官方 LICENSE 与官网条款。本站不替它们判断。⚠ The GitHub licence field for cocos/cocos-engine and love2d/love reads NOASSERTION — machine-undeterminable. Read the official LICENSE and the site's terms. This site does not judge for them.

我们不说的What we do not claim

本站不跑benchmark，不给质量与速度结论We do not run benchmarks and draw no quality or speed conclusions
不做引擎排名，不做作品排名No engine rankings, no work rankings
成本变化幅度是我们的判断，不是数据Cost movement figures are our judgement, not measured data