学习路径Learning path · 第 3 阶段Stage 3
S3 · 手感调参
S3 · Game feel & tuning
这一步决定什么：从「能跑」到「想玩」。这是全站最关键的一步，也是 AI 最帮不上的一步。
What this stage decides: the move from "it runs" to "I want to play it". This is the most important step in the whole process, and the one AI helps least with.
为什么 AI 在这里失效
不是模型不够强，是这个问题本质上没有可验证的答案：
| 问题 | AI 能不能回答 |
|---|---|
| 「跳跃高度应该是多少」 | ✅ 能给合理区间 |
| 「这个跳跃手感好不好」 | ❌ 不能 —— 没有测试能告诉你 |
| 「为什么我觉得它很飘」 | ⚠ 能列可能原因，但列不出你真正的原因 |
⚠ AI 会给你一个看起来很专业的参数，然后你就照着改了 —— 然后手感还是不对。
因为它优化的是「典型值」，而手感是你的判断。
Why AI fails here
It is not that the model is not strong enough — it is that the question has no verifiable answer:
| Question | Can AI answer it |
|---|---|
| "What should jump height be?" | Yes — it can give a sensible range |
| "Does this jump feel good?" | No — no test can tell it |
| "Why does it feel floaty to me?" | ⚠ It can list possible causes, but not your actual cause |
⚠ AI will hand you a parameter that looks professional. You apply it. The feel is still wrong.
Because it optimises for the typical value, and feel is your judgement.
常见的手感问题与对应参数
⚠ 下面是经验判断，不是实测数据。
| 你感觉 | 往哪个方向调 |
|---|---|
| 飘 | 增大重力；跳跃初速不变；落地时给一点下压 |
| 粘 | 减小重力或跳跃初速；加 coyote time（离开平台后仍可跳的宽限） |
| 滑（停不下来） | 增大摩擦/减速度；注意区分「空中控制」和「地面摩擦」 |
| 不确定（边缘判定模糊） | 改碰撞体大小；玩家角色碰撞盒通常该比视觉小一点 |
| 按键延迟感 | 检查输入缓冲（input buffer）；这不是手感问题，是代码问题 |
Common feel problems and which parameter to move
⚠ What follows is experiential judgement, not measured data.
| What you feel | Which direction |
|---|---|
| Floaty | Raise gravity; keep jump velocity; add a little downward force on landing |
| Sticky | Lower gravity or jump velocity; add coyote time (the grace window after leaving a platform) |
| Slippery (won't stop) | Raise friction/deceleration; note that air control and ground friction are separate parameters |
| Uncertain (fuzzy edge detection) | Change the collider size; a player's hitbox is usually smaller than the sprite |
| Input feels laggy | Check the input buffer — that is a code problem, not a feel problem |
一个反直觉的技巧
⚠ 先改数值，再改代码。
大部分「手感不对」其实是一个数值问题，不是逻辑问题。把代码重写三遍不如把重力从 980 调到 1150 试一次。
One counter-intvious trick
⚠ Change numbers before changing code.
Most "it does not feel right" cases are a numbers problem, not a logic problem. Rewriting the code three times is worse than trying gravity 1150 instead of 980 once.
为什么这条最难教
因为它需要你的直觉，而直觉来自你玩过很多游戏。
AI 帮不了你建立直觉，但 AI 能帮你快速试参数：
- 让它做一个「手感参数调试面板」（滑块 + 实时生效）
- 这样你调参从「改代码 → 重启 → 试」变成「拖滑块 → 立刻试」
⚠ 这是 AI 在这一段唯一真正有用的地方：缩短试错循环，不是替你判断。
Why this stage is hardest to teach
Because it requires your intuition, and intuition comes from having played many games.
AI cannot build that intuition for you, but it can help you try parameters fast:
- Ask it to build a "game feel debug panel" (sliders that apply live)
- Tuning then goes from "edit code → restart → try" to "drag a slider → try immediately"
⚠ This is the only genuinely useful thing AI does in this stage: shortening the trial loop, not making the judgement for you.
常见失败
- 让 AI「优化手感」 —— 它会给你一个参数表，你按着改完仍然不对
- 一次改多个参数 —— 改完不知道是哪个起作用
- 跳过这一步 —— 觉得「能跑就行」，结果玩起来没意思
Common failures
- Asking AI to "optimise the feel" — you get a table of parameters, apply them, and it still feels wrong
- Changing several parameters at once — afterwards you cannot tell which one mattered
- Skipping this stage — deciding "it runs, good enough", and ending up with something nobody wants to play
判断你这一步做完了吗
- 别人玩过，且没有问「这游戏怎么玩」
- 你自己玩 30 分钟仍然不觉得无聊
- 关键参数是滑块可调的（不是硬编码）
Are you done with this stage?
- Someone else has played it and did not ask "how do I play this?"
- You have played it for 30 minutes and still find it interesting
- The key parameters are slider-adjustable rather than hard-coded