学习路径Learning path · 贯穿流程Across the stages
实战流程
The build order
五个阶段不是五篇文章，是一条线。 每个阶段页各自讲清了「做什么」，
但阶段之间的交接没人讲 —— 而返工几乎都发生在交接处。
⚠ 本页是本站的口径（方法论），不是实测数据。
The five stages are not five articles — they are one line. Each stage page explains what to do,
but the handoff between stages is what nobody covers — and that is where rework happens.
⚠ This page is the site's own method, not measured data.
流程总览
| # | 步骤 | 产出物 | 交接检查（过了才准往下走） | 最常见的崩法 |
|---|---|---|---|---|
| 1 | 选引擎 | 一个选型决定 | 你能说出「为什么不是另一个」 | 先定美术风格再挑引擎 —— 顺序倒置 |
| 2 | 核心循环 | 一个能跑的原型 | 「能操作 / 有目标 / 有反馈」三条都是「是」 | 跳过原型直接堆内容 |
| 3 | 手感调参 | 可玩的手感 | 一次只改一个参数，且改动有记录 | 把 AI 第一次给的数值当最终值 |
| 4 | 资产合规 | 一份资产台账 | 每条资产的来源与授权都能当场指出 | 先用起来，发布前才查授权 |
| 5 | 导出发布 | 一个可分享的包 | 在别人的机器上能打开 | 最后一刻才发现目标平台导出受限 |
The line at a glance
| # | Step | Output | Handoff check (pass it before moving on) | The usual way it breaks |
|---|---|---|---|---|
| 1 | Pick an engine | One selection decision | You can say why it is not the other one | Settling the art style first, then looking for an engine — the order is inverted |
| 2 | The core loop | A prototype that runs | All three of "operable / goal-directed / responsive" are yes | Skipping the prototype and piling on content |
| 3 | Game feel | Feel that is playable | One parameter changed at a time, and the change is written down | Treating the AI's first numbers as final |
| 4 | Asset licensing | An asset ledger | Every asset's origin and licence can be stated on the spot | Using assets first, checking licences before release |
| 5 | Export and ship | A package you can share | It opens on someone else's machine | Discovering at the last minute that the target platform's export is limited |
每一步的失败点
⚠ 下面这些不是单个阶段的失败，而是交接处的失败 —— 所以它们在阶段页里看不到。
Where each step breaks
⚠ These are not failures inside a single stage — they are failures at the handoff, which is why the stage pages cannot show them.
1 · 选引擎
- ⚠ 顺序倒置：先定了美术风格或玩法体量，再回头找引擎 —— 于是引擎的导出与性能限制变成硬伤。
→ 先按 选引擎 的判据定工具，再谈风格。
- ⚠ 把「AI 友好」当成「引擎好不好」：本站的友好度只回答「AI 写错代码的概率」，不回答引擎强弱。
- ⚠ 没看导出：网页与移动端导出各有硬限制，见 Web 导出可行性。
1 · Picking an engine
- ⚠ Inverted order: settling the art style or the scope first, then hunting for an engine — and its export and performance limits become hard blockers.
→ Pick the tool by the criteria in Picking an engine first, then talk about style.
- ⚠ Reading "AI-friendly" as "a better engine": our friendliness rating only answers "how often AI gets the code wrong", not how strong the engine is.
- ⚠ Ignoring export: web and mobile exports each have hard limits — see Web export feasibility.
2 · 核心循环
- ⚠ 没有目标感：角色能动，但玩家不知道该干嘛 —— 这不算原型。
- ⚠ 一次让 AI 写太多：提示词里塞进敌人、UI、存档，结果拿到一堆半成品。
→ 见 提示词的结构。
- ⚠ 把长会话当版本控制：关键决定散在对话里，改到后面自己也不知道哪一版是对的。
2 · The core loop
- ⚠ No sense of goal: the character moves, but the player does not know what to do — that is not a prototype.
- ⚠ Asking AI for too much at once: stuffing enemies, UI and saves into one prompt returns a pile of half-finished features.
→ See Structuring the prompt.
- ⚠ Using a long chat as version control: key decisions scatter across the conversation, and later you cannot tell which version was right.
3 · 手感调参
- ⚠ 一次改三个参数：改完不知道是哪个起的作用，下次遇到同类问题还是不会调。
- ⚠ 把 AI 给的数值当结论：AI 能给一个起点，但「好不好玩」没有测试能自动判定。
- ⚠ 在错误的地方调：先确认是手感问题还是代码问题，见 手感问题清单。
3 · Game feel
- ⚠ Changing three parameters at once: afterwards you cannot tell which one mattered, and the next similar problem is just as hard.
- ⚠ Treating the AI's numbers as a conclusion: AI can give you a starting point, but no test can decide whether it feels good.
- ⚠ Tuning in the wrong place: establish whether it is a feel problem or a code problem first — see The game-feel checklist.
4 · 资产合规
- ⚠ 先用后查：到发布前才发现字体或音乐不能商用 —— 这是返工成本最高的一种。
- ⚠ 没有台账：素材来源散在下载目录里，谁也说不清哪一条能用。
4 · Asset licensing
- ⚠ Use first, check later: discovering before release that a font or a track cannot be used commercially — the most expensive kind of rework.
- ⚠ No ledger: sources are scattered across download folders, and nobody can say which asset is usable.
→ See The AI asset ledger and The font licensing trap.
5 · 导出发布
- ⚠ 在开发机上能跑就以为能发布：目标平台可能不支持你用到的特性。
- ⚠ 首版做太大：见 先发极简版。
5 · Export and ship
- ⚠ "It runs on my machine" taken as "it can ship": the target platform may not support what you used.
- ⚠ A first release that is too big: see Ship the smallest version first.
三条贯穿全流程的原则
| 原则 | 为什么 |
|---|---|
| 一次只改一个变量 | 同时改多个就无法归因 —— 调参、修 bug、换引擎都适用 |
| 先能跑，再好看 | 粗糙原型逼你先验证玩法，而不是被「看起来还行」误导 |
| 台账从第一天开始记 | 合规成本随资产数量增长，事后补台账等于把资产重查一遍 |
Three rules that run through everything
| Rule | Why |
|---|---|
| Change one variable at a time | Change several and you cannot attribute the result — true for tuning, bug fixing and swapping engines alike |
| Runnable first, pretty later | Crude prototypes force you to validate the gameplay instead of being seduced by "it already looks decent" |
| Keep the ledger from day one | Licensing cost grows with the number of assets; rebuilding the ledger later means re-checking every asset |
想直接抄代码
五个引擎的最小骨架（能操作 / 有目标 / 有反馈）见 最小可玩原型代码清单。
⚠ 那些骨架未经本站实测，是起点模板，每个都附了官方入门出处。
If you just want code to copy
Minimal skeletons for the five engines (operable / goal-directed / responsive) are in The minimal prototype code list.
⚠ Those skeletons are not tested by this site — they are starting templates, each with a link to the official tutorial.