怎么定一个两周能做完的范围How to scope a game down to two weeks

这篇解决什么：首版做太大的问题。AI 降低了写代码和出素材的成本，但没有降低「决定做什么」的难度 —— 范围仍然要人裁。

⚠ 先说清口径：「两周」是我们建议的练习周期（方便设检查点），不是实测的完成时间，也不是行业标准；下文所有工时数字都是虚构示例。

What this guide solves: first versions that are too big. AI lowered the cost of writing code and making assets; it did not lower the difficulty of deciding what to build. The scope still gets cut by a human.

⚠ Framing first: "two weeks" is a practice cycle we suggest (handy for checkpoints) — not a measured completion time, not an industry standard. Every effort figure below is a fictitious example.

适用条件

• ✅ 个人或两三个人的验证型项目：目标是「证明这个玩法有没有戏」
• ✅ 你已有一个能跑的核心循环原型，或准备在周期第 1 周得到它
• ⚠ 不适用：签了交付合同的商业项目 —— 合同范围不由这篇决定
• ⚠ 不适用：还没定引擎工具 —— 先看 S1 · 选引擎

When this applies

• ✅ A solo or two-three-person validation project: the goal is "does this mechanic have a future"
• ✅ You already have a running core-loop prototype, or you will have one in week one
• ⚠ Not for contracted commercial work — the contract decides that scope
• ⚠ Not before the tool is picked — start at S1 · Pick an engine

步骤

1. 写验证问题，不写愿望。 一句话形如「到第 10 天（虚构），陌生人玩 3 分钟后能说出这游戏的目标，并且想再来一局」。判断标准：这句话可证伪。

2. 把所有想要的东西列全。 玩法、界面、存档、教学关、成就、多平台 —— 先不裁，列到不想加为止。

3. 三档分类。 必须（验证问题缺了它就不成立）/ 可有（有它体验更好，不影响验证）/ 本期砍掉。检验「必须」的唯一问法：没有它，第 1 步那句话还怎么做？

4. 给每个「必须」写完成条件。 形如「拖拽放置方块，吸附到网格，放置失败有反馈」—— 写不出完成条件的项，多半是「可有」伪装成「必须」。

5. 估一遍并砍到留余量。 按天估每个「必须」（估法是练习，数字虚构）；全部加起来明显超过周期六成时（虚构阈值，我们的口径是留缓冲），继续从「必须」里砍，或把大项拆小。

6. 第 6 天中检（虚构日）。 有任何一项「必须」还没动过：第一动作是删「可有」，不是加班。

7. 周期末导出一次。 哪怕粗糙，必须走通导出 —— 见 怎么做一次发布前检查；发出去收集反馈走 先发极简版。

The steps

1. Write a validation question, not a wish. One sentence: "by day 10 (fictitious), a stranger plays for 3 minutes, can state the goal, and wants another round." Test: it must be falsifiable.

2. List everything you want. Mechanics, menus, saves, tutorial levels, achievements, multi-platform — no cutting yet.

3. Sort into three piles. Must-have (without it the validation question collapses) / nice-to-have / cut-this-cycle. The only test for must-have: without it, how does the sentence from step 1 get answered?

4. Give each must-have a done-condition. Like "drag a block, it snaps to the grid, a failed placement gives feedback" — items you cannot phrase this way are usually nice-to-haves in disguise.

5. Estimate, then cut to leave slack. Day-size estimates per must-have (the exercise is invented, the numbers are fictitious). When the total clearly exceeds about 60% of the cycle (invented threshold — our reading is: keep a buffer), cut deeper or split big items.

6. Day-6 checkpoint (fictitious day). Any untouched must-have: the first move is deleting nice-to-haves, not overtime.

7. Export once at the end of the cycle. Rough is fine; exporting must actually happen — see How to run a pre-release check; collecting feedback afterwards is Ship the smallest version first.

一个可走查的例子

⚠ 功能清单、天数、通过线全部为虚构示例。

• 输入：一款网格解谜游戏。验证问题：「第 10 天，陌生人 3 分钟内过关 1 关并说出规则」。候选功能五项：核心推拉机制、关卡选择界面、云存档、音效、新手教学关。
• 处理：分类 → 必须 = 核心机制 + 1 关内容（关卡数据组织见 关卡数据的组织方式）；可有 = 音效、教学关；砍 = 云存档、关卡选择。第 6 天（虚构）检查：核心机制已能玩，第 1 关拼图中——音效一行没动，属「可有」，不慌。
• 结果判断：按验证问题逐项问「缺了还能验证吗」—— 五项里只有两项答「不能」，范围成立。
• 失败后下一步：假如第 8 天（虚构）核心机制仍卡住 → 不是砍功能（已经没可砍），而是回查改动质量：按 怎么验收 AI 生成的游戏代码 走一遍，通常问题在那一层。

A walk-through example

⚠ The feature list, the days and the pass line are all fictitious examples.

• Input: a grid puzzle game. Validation question: "by day 10, a stranger clears level 1 in 3 minutes and states the rule." Five candidate features: core push-pull mechanic, level-select menu, cloud saves, sound, tutorial level.
• Processing: sorting → must-haves = the mechanic + one playable level (level data: How to organise level data); nice-to-haves = sound, tutorial; cut = cloud saves, level select. Day-6 checkpoint (invented): the mechanic plays, level 1 is being assembled, zero audio touched — that is the nice-to-have pile, no panic.
• Result: one question per item — "can the validation still happen without it?" Only two of five answer "no", so the scope holds.
• After the failure: if by day 8 (invented) the core mechanic is still broken → there is nothing left to cut; the problem is change quality — run How to accept AI-generated game code instead; that is usually the layer at fault.

失败分支

Failure branches

下一步

• 定完范围进入完整循环：S2 · 核心循环
• 周期末的发法：先发极简版

Next steps

• Scope set, enter the full loop: S2 · The core loop
• How to release at the end of the cycle: Ship the smallest version first