36/41 · Part 6 · 跟 Agent · prompt 套路

Agent · 4 条 prompt 套路

让 agent
说话像那个游戏。

1 嗓音

system prompt 角色化

不要让 LLM 默认 ChatGPT-friendly 平淡嗓音。

✅ Trash or Treasure 温柔毒舌 / Hour Capsule 博物学家 / Pulp Hour 黑色一刻 / Daily Arcana 塔罗师 / The Locksmith 锁匠

2 锚点

4 层 elevator + skill 名当 anchor

把 idea 说成"身份钩 X + 范式 Y + skill [[Z]] pattern"。

例："做一个 sensory-toy + 身份钩 ④（网络）+ [[score-beat-run-end-vs-sensory-toy]]" 比"做一个泡泡纸玩具，可以看到好友戳的次数"准 10 倍。

3 参考

给已上线游戏 URL + reference video

"像 Crossy Road 但角色是 X" / "像 Suika 但合成 X" / "像 Murmur 但触发是 Y"。

Build-a-Boyfriend = Suika fork · Block Hop = Crossy Road fork · Hour Capsule = Seal Press fork。让 agent 站在巨人肩膀上，不是从 0 写。

4 反馈

反馈语言模板化

同样的反馈短语对应不同 pattern：

"太复杂" → 砍规则 / 砍 setup
"social 太弱" → 加 verb + notify
"看不下去" → 砍卡通脸 sticker
"没耐心" → 砍引导 / 用 1 个手指 icon
"太平淡" → 加 visceral hit

Agents · 4 prompt patterns

Make the agent
sound like that game.

1 voice

In-character system prompt

Never let the LLM default to ChatGPT-friendly mush.

✅ Trash or Treasure snarky-but-tender / Hour Capsule curator / Pulp Hour noir / Daily Arcana tarot reader / The Locksmith

2 anchor

4-layer elevator + skill names

Restate the idea as "identity hook X + paradigm Y + skill [[Z]] pattern."

e.g. "sensory-toy + identity hook ④ (network) + [[score-beat-run-end-vs-sensory-toy]]" is 10x more precise than "a bubble-wrap toy that shows your friends' tap counts."

3 reference

Shipped URL + reference video

"Like Crossy Road but characters are X" / "Like Suika but the merge is X" / "Like Murmur but the trigger is Y."

Build-a-Boyfriend = Suika fork. Block Hop = Crossy Road fork. Hour Capsule = Seal Press fork. Stand on shoulders instead of writing from zero.

4 feedback

Template your feedback language

The same player short-phrases map to specific patterns:

"too complex" → cut rules / cut setup
"social is weak" → add a verb + notify
"can't watch this" → cut cartoon sticker face
"no patience" → cut tutorial / 1 finger icon
"too flat" → add visceral hits

备忘 · 按 S 关闭Notes · S to close

讲法：每条都给一个真实例子。第 4 条最有用——这些短语本身已经成为我们沟通的 ontology，agent 听到立刻知道往哪改。

建议同学把这 4 条 prompt 套路存进自己的 system prompt 模板，下次开 idea 时直接套。

Walk each one with a real example. Pattern 4 is the most actionable — these phrases have become our shared ontology; the agent hears them and immediately knows the direction.

Suggest students save these four into their personal system-prompt template, ready for the next idea.

让 agent说话像那个游戏。