大家好 我是为什么叫QQ

Hello everyone, I'm WhyCallQQ.

自从Claude Opus 5.5发布

Ever since Claude Opus 5.5 launched,

不到24小时 X上冒出一堆游戏演示

within 24 hours, X was flooded with playable game demos:

有魂类游戏 带敌人AI 锁定视角和闪避

a Soulslike prototype with enemy AI, camera lock-on, and dodge rolls;

有带翅膀的恶魔Boss战

a winged demon boss fight;

还有一个自定义角色 从森林一路走进小镇

and a custom character walking all the way from a forest into a town.

这些游戏有个共同点

All of these games share one thing in common:

制作人几乎没碰过虚幻引擎

their creators barely touched Unreal Engine manually—

他们只是对着编辑器打字

they simply typed prompts into the editor.

同一时间 一家硬件实验室做了同样的尝试

At the same time, a hardware lab ran the exact same experiment,

结论却是 AI把游戏改坏了 而且修不回来

only to conclude that the AI broke the game beyond repair.

同一套工具 两种完全不同的结局

The exact same toolset produced two completely opposite outcomes.

中间差的那点东西 就是这期视频的内容

What made the difference between them is the focus of today's presentation.

这是一篇实测长文的精读

This is a close reading of an in-depth benchmark report

作者rewind把四个真实测试拼在一起

where author rewind pieced together four real-world experiments

写了一份完整路线图

into a complete end-to-end roadmap:

从空项目开始 连接编辑器 搭关卡

starting from an empty project, connecting the editor, building greybox levels,

加敌人和战斗 换上自己的角色 再做测试收尾

adding enemies and combat, swapping in custom characters, and finishing with verification.

我按顺序讲三件事

I will walk through three things in order:

连接怎么搭， 指令怎么下 ，坑在哪里

how to wire the connection, how to structure prompts, and where the pitfalls lie.

所有数字都来自原始测试 不用宣传口径

Every number comes straight from raw benchmarks—no marketing spin.

看完你能判断这套工作流适不适合你

By the end, you'll know whether this workflow fits your needs

也能拿到一份可以直接照做的清单

and walk away with a practical checklist you can follow immediately.

建议先收藏 有效期会很长 先说原理

Bookmark this—it'll stay relevant for a long time. Let's start with how it works.

以前的AI助手只能给你代码 你自己粘贴

Previous AI assistants could only hand you code snippets to copy and paste yourself.

这次不一样

This time is fundamentally different.

从虚幻引擎5.8开始 编辑器自带官方MCP服务器

Starting in Unreal Engine 5.8, the editor ships with a built-in official MCP server.

Epic还维护了一个Claude Code插件

Epic also maintains an official Claude Code plugin—

免费 MIT协议 在Anthropic官方插件市场

free, MIT-licensed, and available in Anthropic's official plugin marketplace.

连上之后 Claude能直接读取正在运行的关卡

Once connected, Claude can directly inspect your live running level

当场修改 你看着它动

and modify it on the spot while you watch.

它操作的是你场景里真实存在的Actor和蓝图

It manipulates the actual Actors and Blueprints inside your scene.

整个过程在本地跑 服务器监听8000端口

Everything runs locally, with the server listening on port 8000.

引擎官方下场给AI留接口

When the engine maker itself builds a first-party AI interface,

这本身就值得程序员停下来看一眼 连接一共六步

that alone deserves every developer's attention. Connecting it takes six steps.

一 在编辑器插件面板勾选Model

Step 1: In the Unreal Editor Plugins panel, enable both Model

Context Protocol和All Tools ets

Context Protocol and All Toolsets—

两个都要 然后重启

you need both checked, then restart the editor.

只开第一个， 服务器有， 工具没有

If you only enable the first one, the server starts but exposes zero tools.

二 按波浪号键打开控制台 运行StartServer

Step 2: Press the tilde key (~) to open the console and run StartServer.

三 运行GenerateClientConfig ClaudeCode

Step 3: Run GenerateClientConfig ClaudeCode

生成一个.mcp.json文件

to generate a .mcp.json configuration file.

四 在Claude Code里一条命令装上Epic的插件

Step 4: Run a single command in Claude Code to install Epic's official plugin.

五 从存放.uproject的文件夹启动Claude Code

Step 5: Launch Claude Code from the exact folder containing your .uproject file—

它只认启动目录 这是最多人翻车的一步

it only checks the launch directory, which is where most people trip up.

六 发一条只读指令验证

Step 6: Send a read-only prompt to verify the connection.

它报得出关卡里的真实物体 链路就通了

Once it accurately lists the real objects in your level, the pipeline is live.

现在看实测 四个案例摆在一起才有判断力

Now let's examine the benchmarks—putting all four side by side gives real perspective.

第一个 创作者做出魂类原型

Case 1: A creator built a playable Soulslike prototype

敌人巡逻 追击 攻击 血条 锁定镜头

featuring enemy patrol, chase, attack states, health bars, camera lock-on,

三段连击加闪避 还有一个地牢关卡

a three-hit combo with dodge rolls, plus a dungeon level.

资产是Claude在Blender里建的

The 3D assets were modeled by Claude inside Blender,

全程只手动修了一处动画设置

requiring only a single manual animation setting tweak across the entire build.

第二个 Boss战 会飞的恶魔三种攻击 成了

Case 2: A flying demon boss fight with three attack patterns—it worked,

代价是等了半天 第一个3D模型报废

at the cost of hours of waiting, a scrapped first 3D model,

粒子效果没做完 额度先见底

and hitting the token quota ceiling before particle effects were finished.

第三个 Tripo的演示 森林到小镇的路线

Case 3: Tripo's demo—a path from forest to town

收集品加HUD 中途换上自定义角色

with collectibles, a HUD, and swapping in a custom character mid-way.

核心建议只有一句 每批操作后亲自进编辑器检查

Their core advice boils down to one sentence: inspect the editor yourself after every batch.

第四个案例最重要

The fourth case study is the most revealing of all,

来自硬件实验室Puget Systems

conducted by hardware lab Puget Systems

时间是七月 用的还是上一代模型Opus 4.8

back in July using the previous-generation Opus 4.8 model.

两个简单任务完成得很好

It aced two simple tasks:

一条走廊加六个亮灯的房间

building a corridor with six lit rooms

五分多钟交活 和描述高度吻合

in just over five minutes, matching the prompt specification closely.

可大需求一上来 局面马上变了

But the moment they gave it broad, complex requirements, everything fell apart.

让它改玩家操控 游戏直接没法玩 修也没修回来

Asked to overhaul player controls, it rendered the game unplayable and couldn't fix it.

让它建程序化城市 它悄悄用手摆模型冒充

Asked to build a procedural city, it quietly placed meshes by hand to fake it—

一个程序化节点都没生成 这里想听你们的

without generating a single PCG node. I'd love to hear your take:

AI这种闷声偷工减料 算bug还是算特性？

when an AI silently cuts corners like that, is it a bug or a feature?

把你的看法打在公屏上 讲指令之前 先讲安全

Drop your thoughts in the comments. Before prompting habits, let's talk safety.

Epic自己的描述非常直白

Epic's own documentation is refreshingly blunt:

装了这个插件

installing this plugin

等于把正在运行的编辑器开放给Claude

exposes your live Unreal Editor directly to Claude.

你要把它当成在跑任意代码 因为实际就是

Treat it as arbitrary code execution—because that's literally what it is.

有个工具能在编辑器里直接执行Python

One tool executes arbitrary Python directly inside the editor

对项目有完全访问权

with full unrestricted access to your project.

改出来的东西撤销不一定救得回来 所以几条铁律

Undo won't always rescue broken assets, so follow these iron rules:

每次开工先提交

Always commit to version control before starting a session.

审批确认别关 更不要带着插件跑跳过权限的模式

Never disable confirmation prompts or run permission-bypass flags with the plugin active.

先拿测试项目练手 改东西之前退出Play模式

Practice on a sandbox project first, and always exit Play-in-Editor before making edits.

编译C++和着色器的时候别发指令

Never send prompts while C++ code or shaders are compiling.

服务器没有登录机制

The local MCP server has zero authentication—

端口别暴露 别在共享电脑上跑

never expose port 8000 externally or run it on shared machines.

真正决定成败的是你怎么下指令

What truly separates success from failure is how you issue prompts.

六个习惯 按重要性讲

Here are six habits ranked by importance:

一 一条提示词只做一个系统

Habit 1: Build only one system per prompt.

关卡是关卡 血量是血量 敌人是敌人

Keep level layout, health systems, and enemy AI strictly separate—

Puget两次失败 全是大而模糊的请求

both of Puget's failures came from massive, vague prompts.

二 说清楚做完长什么样

Habit 2: Define a concrete 'definition of done'—

做完的标准是我能从出生点走到塔 不卡住

for example: 'I can walk from the spawn point to the tower without getting stuck.'

有了验收标准 它才能自己检查

Clear acceptance criteria allow the model to self-verify.

三 说清楚什么不许动 保住你验过的成果

Habit 3: Explicitly state what must NOT be touched to protect verified work.

四 要求交报告

Habit 4: Demand a change report

列出每个新建和改过的蓝图与资产

listing every Blueprint and asset created or modified.

五 禁止闷声绕路 做不了就停下来报告

Habit 5: Forbid silent workarounds—if it's blocked, it must stop and report.

六 提示词开头带上/unreal-mcp 明确触发技能

Habit 6: Prefix prompts with /unreal-mcp to explicitly trigger the skill.

有了习惯 再看完整路线 练习项目故意做得很小

With those habits in place, let's look at the full roadmap using a deliberately small project:

第三人称 一条路 一个村庄 三颗发光宝珠

third-person camera, one road, one village, three glowing orbs,

一个守门的敌人 一个胜利判定 一个重新开始键

one gatekeeper enemy, a win condition, and a restart key.

小就是设计目的 能做完的切片才教得会流程

Keeping it small is intentional—only a slice you can finish teaches the full loop.

整个流程六道关

The pipeline has six stages:

连上编辑器 搭灰盒关卡 写规则 加敌人和战斗

connect the editor, greybox the level, script rules, add enemy combat,

做外观 完整测试

dress the visuals, and run full verification.

每道关都有验收动作 通过了就提交一次

Every stage has a verification check; pass the check and commit immediately.

提交点就是你的存档点

Each Git commit acts as your save point.

程序员看这套东西会很眼熟 CI的红绿闸门

Software engineers will recognize this immediately: CI red/green quality gates

被原样搬进了游戏开发 执行顺序有讲究

transplanted straight into game development. The execution order matters.

先用第三人称模板建项目 让Claude搭灰盒

Start from the Third-Person template and have Claude build a greybox level:

一片空地 一条一百五十米长的弯路

an open clearing, a 150-meter winding road,

尽头一个有塔的村子 只用简单几何体

and a village with a tower at the end, using only simple geometric primitives.

Puget的同类任务花了五分多钟

Puget's equivalent greybox task took just over five minutes,

模型还自作主张加了走廊灯

though the model took the liberty of adding corridor lights on its own.

这种合理猜测很常见 所以每次都要它列清单

Unprompted additions like that are common, which is why you always require a change manifest.

按Play走一遍 过了 提交

Hit Play, walk the path end to end, and if it works, commit.

接着加规则 三颗宝珠 屏幕计数 集齐开门

Next, add game rules: three orbs, an on-screen counter, opening the gate when all are collected,

R键重开

and pressing R to restart.

再往后加战斗 一次一个功能 敌人

Then layer in combat one feature at a time: the enemy,

血条 三段连击 锁定加闪避

health bars, a three-hit combo, camera lock-on, and dodge rolls.

Boss战测试里 三段连击花了大约一小时

In the boss fight benchmark, wiring the three-hit combo took about an hour—

作者说他手写只要十五到二十分钟

something the author noted he could hand-code in 15 to 20 minutes.

灰盒能玩了 才轮到外观 资产三条路

Only after the greybox plays well do you tackle visuals—via three asset paths:

最快是Fab商城的现成资产包

The fastest route is ready-made asset packs from the Fab Marketplace.

要自己的角色用Tripo 生成带骨骼和动画的FBX

For custom characters, use Tripo to generate rigged and animated FBX files—

骨架预设选UE5 Mannequin 省掉后面的重定向

choose the UE5 Mannequin skeleton preset to skip animation retargeting later.

要风格统一的道具 可以让Claude驱动Blender

For cohesive custom props, you can have Claude drive Blender directly,

但很慢 一个精细Boss模型试了两次

though it's slow: a detailed boss model took two attempts

花了约三个半小时

and roughly three and a half hours.

换装要分批 一次一个区域 先村子 再植被

Swap visuals in batches, one zone at a time: village first, then foliage,

再灯光音效

then lighting and audio.

程序化生成的PCG是Puget翻车的地方

Procedural Content Generation (PCG) is where Puget's test failed,

现在Epic的City Sample

but Epic's City Sample project now

注册了教模型搭PCG图的技能

registers dedicated skills that teach the model how to wire PCG graphs.

动手前先问一句 这个项目里注册了哪些技能

Before starting, always ask Claude what project skills are registered.

测试决定原型能不能信 四层 各管一段

Testing determines whether your prototype can be trusted, across four distinct layers:

第一层 工具返回的状态

Layer 1: Tool return statuses.

很多工具失败不报错 只在回复里带状态

Many MCP tools fail silently without throwing exceptions, embedding status flags in the response,

提示词里必须要求它读编译结果

so your prompt must explicitly require reading compilation results.

第二层 截图加日志

Layer 2: Viewport screenshots plus Output Log.

每批操作后让它截一张关卡图

After every batch of edits, have Claude capture a level screenshot

报Output Log里的新错误

and report any new errors in the Output Log.

第三层 亲手玩 这层没法外包 碰撞 手感 镜头

Layer 3: Hands-on playtesting. You cannot outsource collision feel, combat responsiveness, or camera flow—

按了Play才知道

you only know by hitting Play yourself.

模型能自己截图检查改动 但手感它截不出来

The model can screenshot visual changes, but it can never screenshot game feel.

第四层 自动化测试 有C++测试的项目

Layer 4: Automated tests. In projects with C++ test suites,

它能发现 运行并解读失败

Claude can discover, run, and diagnose failing tests.

修bug也有纪律 两次修复失败就停手回滚

Debugging requires discipline too: if two fix attempts fail, stop and roll back,

把任务切小再来

then break the task into smaller pieces.

在两次失败上叠第三次修复 项目就是这么废掉的

Stacking a third blind fix on top of two failures is exactly how projects get ruined.

说钱 这个工作流最大的成本是等待

Let's talk cost. The biggest expense in this workflow is waiting:

每个请求五到十五分钟

each complex request takes 5 to 15 minutes,

改 测 调的节奏完全被打断 额度是另一笔账

breaking the tight edit-test-tune loop. Token quota is the other bill:

Opus 5.5的API价 输入每百万token四美元

Opus 5.5 API pricing sits at $4 per million input tokens

输出二十美元

and $20 per million output tokens.

Puget一个上午的小实验 用掉团队周额度的15%

Puget's single morning experiment consumed 15% of their team's weekly quota.

最极端的 一位开发者36小时做完多人RPG

In the most extreme case, a developer built a multiplayer RPG in 36 hours,

烧光最大档套餐一整周的额度

burning through an entire week's Max plan quota—

按API价折算约2400美元 省钱四招

roughly $2,400 at API rates. Here are four ways to save money:

平时待在medium档 同一个难题low约0.55美元

Stay on medium effort by default—the same hard task costs ~$0.55 on low

max要5.98美元 差十倍 指令一次写好

versus $5.98 on max, a 10x gap. Write clear prompts upfront,

给实验设时间盒

timebox your experiments,

五分钟的活自己动手

and do five-minute tasks yourself—

挪一个Actor比写提示词快得多

dragging an Actor in the viewport is much faster than writing a prompt.

网上演示讲的版本 是AI替你写游戏

Viral social media demos sell the fantasy that AI builds the game for you.

四个测试拼出来的版本 完全两样

The reality across these four benchmarks is completely different:

它执行明确指令很快

the model executes well-specified instructions rapidly,

从一句模糊的话里发明一个完整系统很弱

but is weak at inventing coherent systems from a single vague sentence,

而且判断不了成果好不好玩

and cannot judge whether the result is actually fun to play.

Puget评审那句话值得抄下来

Puget Systems' conclusion is worth writing down:

用它加速你已经懂的工作 别拿它填你知识上的洞

'Use it to accelerate work you already understand; don't use it to fill gaps in your knowledge.'

那个做出魂类原型的人 本身懂游戏

The creator who shipped the Soulslike prototype already understood game design;

那个改坏操控的人 把判断权交了出去

the one whose controls broke handed over design judgment to the model.

你越懂引擎 它给你的越多

The better you understand the engine, the more leverage the AI gives you.

工具放大的是方向 站在哪边 你自己选

Tools amplify your direction—which side you stand on is your choice.

整套流程抽象成三步 ，一 开工留档

The entire workflow abstracts into three steps: First, checkpoint before starting—

提交代码 定验收标准 划保护范围

commit your code, define acceptance criteria, and mark off-limits files.

二 小批次推进

Second, advance in small batches—

一条指令一个系统 每批做完让模型自查

one system per prompt, having the model self-check after every batch

状态 截图 日志三件套 三 人工验收再入库

using status codes, screenshots, and logs. Third, human playtest before committing—

亲手跑一遍 过了提交 挂了回滚 眼熟吗

play it yourself: commit if it passes, roll back if it breaks. Sound familiar?

这就是带一个执行力很强

That is the exact daily routine of mentoring a junior engineer

判断力很弱的初级工程师的日常

with high execution speed but low architectural judgment.

code review CI闸门 两败回滚

Code reviews, CI gates, two-strike rollbacks—

全是软件工程里的老东西

all time-tested software engineering fundamentals.

AI没有发明新流程 它把带人的成本打了下来

AI didn't invent a new process; it dramatically lowered the cost of delegation.

第二个问题留给你

Here is my second question for you:

你会把哪类任务放心交给这样的初级工程师？

what kinds of tasks would you comfortably delegate to a junior engineer like this?

评论区见 往后看两个信号

Let me know in the comments. Looking ahead, two signals stand out:

MCP服务器现在还挂着Experimental标签

First, the Unreal MCP server still carries an Experimental tag.

Epic文档说工具调用一次一个

Epic's docs state tool calls run one at a time,

插件九月更新说已经支持并发

while the plugin's September update claims concurrent tool execution—

两边口径还在打架 说明迭代非常快

the docs and changelog contradicting each other shows how fast it's evolving.

引擎厂商给AI留官方接口 这条路已经走通

First-party AI interfaces built directly into game engines are now a proven path.

技能注册进项目是另一个信号

Second, registering reusable AI skills directly inside projects:

PCG技能出现了

PCG skills are already here,

后面会有动画技能 材质技能 性能分析技能

and animation, material, and profiling skills will follow soon.

编辑器会越来越像一个带工具箱的运行环境

Game editors are becoming runtime environments equipped with AI toolboxes:

模型负责执行 人负责验收

the model handles execution, while humans own verification.

谁先练熟这套协作 谁先吃到效率红利

Whoever masters this collaboration loop first captures the productivity dividend.

最小可行版本只有一个晚上

Your Minimum Viable Product takes just a single evening:

连上编辑器 搭出灰盒

connect the editor, greybox a simple level,

加三颗宝珠和一条胜利提示 从起点玩到通关

add three orbs and a victory prompt, and play from spawn to finish.

之后的一切 都是同一个循环 放大野心再来一遍

Everything after that is the exact same loop scaled up with bigger ambition.

这期内容来自rewind的实测长文 原始链接在简介

This episode is based on rewind's benchmark report—the original link is in the description.

留个问题 如果用这套流程做你的第一个游戏

One final question: if you used this workflow to build your first game,

你会选什么题材？打在评论区

what genre would you pick? Share it in the comments.

觉得有用 点赞投币收藏

If you found this useful, please like, support, and bookmark,

转发给那个一直想拉你一起做游戏的同事

and share it with that colleague who's always wanted to build a game with you.

我是为什么叫QQ 我们下期见

I'm WhyCallQQ—see you in the next episode.