学习路径Learning path · 关键词Keyword
上下文窗口与 AI 写游戏
Context window and AI-written games
一句话：游戏代码的形态（长期演化、多文件耦合）与 AI 的上下文限制正好相冲。
In one line: the shape of game code (long-lived, multi-file, coupled) runs directly into the limits of AI context.
⚠ 为什么这是个真问题
| 游戏代码的特点 | 对上下文窗口的影响 |
|---|---|
| 单个文件很大 | 玩家控制器、敌人 AI、关卡逻辑都几百行起 |
| 跨文件耦合 | 改一个参数要看三个文件才知道影响 |
| ⚠ 长程一致性 | ⚠ 游戏逻辑要前后一致（这个变量在第 3 章出现过） |
⚠ 模型一次只能看到一部分代码 —— 所以它会：
- 重复定义已经存在的函数
- 用不同的命名做同一件事
- ⚠ 忘记自己在第 200 行时约定过什么
⚠ Why this is a real problem
| Property of game code | Effect on the context window |
|---|---|
| Large single files | Player controllers, enemy AI and level logic run to hundreds of lines |
| Cross-file coupling | Changing one parameter means reading three files to see the effect |
| ⚠ Long-range consistency | ⚠ Game logic must stay consistent (this variable appeared in chapter 3) |
A model sees only part of the code at a time — so it will:
- redefine functions that already exist
- use different names for the same thing
- ⚠ forget what it agreed at line 200
实操建议
1. 单文件控制在几百行内
⚠ 超了就拆。⚠ 这不只对 AI 有好处 —— 人类也受益。
2. 每次只让 AI 处理一个模块
✅ 好：「把敌人 AI 从 80 行扩到 200 行，保持现有的状态机结构」
❌ 差：「加上敌人 AI、关卡、UI」
3. ⚠ 关键约定写进注释或文档
AI 看不到「上次的约定」，但它能读注释：
```python
约定：所有伤害计算走 damage() 函数，不要直接改 hp
⚠ 这个约定是项目级的，不要在单个函数里临时改
```
4. ⚠ 定期让 AI 总结当前状态
⚠ 让它写一份「这个项目现在有什么」的文档，
然后每次新会话先让它读这份文档。这比硬塞代码进上下文有效得多。
Practical advice
1. Keep single files to a few hundred lines
⚠ Split them when they grow past that. ⚠ This helps humans too, not only AI.
2. Let AI work on one module at a time
✅ Good: "expand the enemy AI from 80 to 200 lines, keeping the existing state-machine structure"
❌ Bad: "add enemy AI, levels and UI"
3. ⚠ Put key conventions in comments or a document
AI cannot see "what we agreed last time", but it can read comments:
```python
Convention: all damage goes through damage(); do not modify hp directly
⚠ This convention is project-level — do not break it inside a single function
```
4. ⚠ Have AI summarise the project state regularly
⚠ Ask it to write a document listing "what this project currently has", then have it read that document at the start of every new session. That works far better than stuffing code into the context.
⚠ 一个诚实的说明
本站没有实测不同模型在不同上下文长度下的表现。
以上是基于「上下文窗口是硬约束」这个事实的推断。
⚠ 具体某个模型会怎么退化，请自己测 —— 这恰好也符合
本站的精神：说「未知」比猜好。
⚠ An honest caveat
This site has not measured how different models degrade at different context lengths.
What follows is an inference from the one hard fact: the context window is bounded.
⚠ Test how your model of choice actually degrades — which is itself in the spirit of this site: saying "unknown" beats guessing.