AlterU AlterU Part 6 · 跟 Agent · 先看再做 Part 6 · Agents · Explore First
大纲Outline
Agent · 先看再做

让 agent 先看
再让它动手。

Agent 最常翻车的不是写错代码，是没看现状直接动手—— 脑补一份风格、脑补一个 API、脑补一个 pattern。

  • ✅ "先 explore 这个游戏的 X，再决定怎么改" — 让它 Read 现有 asset / 现有代码 / 现有 skill
  • ✅ "看一下 [[score-beat-run-end-vs-sensory-toy]] 这个 pattern 怎么写的" — 用 skill 名当 anchor，比口头描述准
  • ✅ "参考 Kiss Wall 的 chemistry-bath develop 过程" — 给已上线游戏的 URL，比 PRD 强
  • ❌ "做一个像 X 的功能" — 漏掉了"先看 X 怎么做的"

真实事故：Kiss Wall v1 prompt 脑补不看库 → ship 立刻被骂；The Bidding Paula 不指定 head/body ratio → 11 岁角色被画成婴儿。 都是"动手前没看"的代价。

A designer talking to an AI agent across a desk
Agents · explore before you build

Let your agent look
before you let it act.

The #1 thing an agent gets wrong is not bad code — it's acting without reading the current state: imagining a style, an API, a pattern.

  • ✅ "First explore X in this game, then decide what to change" — make it Read the assets / code / skill before deciding
  • ✅ "Check how [[score-beat-run-end-vs-sensory-toy]] is implemented" — skill names are sharper anchors than verbal descriptions
  • ✅ "Reference Kiss Wall's chemistry-bath develop process" — pointing at a shipped game beats writing a PRD
  • ❌ "Build a feature like X" — missed step: first see how X is done

Real incidents: Kiss Wall v1 prompts written from imagination → roasted on ship; The Bidding's Paula without head/body ratio → 11-year-old rendered as a baby. Both = "act before look."

A designer talking to an AI agent across a desk

备忘 · 按 S 关闭Notes · S to close

这条是整个"跟 agent 沟通"章的总规则。其他规则都是它的展开。

讲法：让同学回想自己最近一次让 ChatGPT / Claude 做事翻车的场景——大多数时候是"我没让它先看，它脑补了"。 agent 写代码时也一样——你不让它 Read 库里现有 asset，它脑补一个看似合理的风格 → ship → 被打脸。

实际用法是给 agent 明确的"先 explore + 再 decide + 再 implement"三阶段。 explore 用 read-only agent (Explore subagent)，decide 用 Plan agent，implement 才 Edit/Write。

This is the master rule for the whole "talking to your agent" chapter. The rest are corollaries.

Have the room recall the last time ChatGPT / Claude got something wrong — most cases come down to "I didn't make it look, it imagined." Agents writing code repeat the pattern: if you don't make them Read the existing library, they invent a plausible-looking style → ship → roast.

The practical move is to give the agent an explicit three-phase script: explore (read-only agent) → decide (planning agent) → implement (edit agent).

35 / 41