AlterU AlterU Part 3 · LLM Part 3 · LLM
大纲Outline
能力 · game-chat (LLM)

让游戏说话。

  • OpenAI 兼容格式 · system prompt + user → assistant text
  • 多轮历史可维护（Pulp Hour 短篇对话）
  • 关键：system prompt 决定 AI 嗓音 · 永远写"角色化、有性格"
  • 反例：默认 ChatGPT-friendly 平淡语气 = AI slop 直接的 UI

真实嗓音案例：

  • Trash or Treasure：温柔毒舌 · "KEEP it. You'll regret tossing this."
  • Hour Capsule：博物学家 · "Manufactured 14:32, Tuesday. Limited run of one."
  • Pulp Hour：黑色一刻 · 6 段视觉 hit 短篇
Pulp Hour
Capability · game-chat (LLM)

Let the game talk.

  • OpenAI-compatible · system + user → assistant text
  • Multi-turn history if you need it (Pulp Hour short pieces)
  • Key rule: system prompts must be in-character. Default ChatGPT-friendly = soul-dead UI.
  • Real voices: Trash or Treasure (snarky-but-tender) · Hour Capsule (curator) · Pulp Hour (noir) · Daily Arcana (tarot reader) · The Locksmith
Pulp Hour

备忘 · 按 S 关闭Notes · S to close

这一张关键是让设计师意识到 LLM 不是"返回正确答案的盒子"，是"返回角色台词的盒子"。

system prompt 写好整个游戏会有灵魂。写糟糕就是 ChatGPT 套皮。 Trash or Treasure 的"温柔毒舌"语气定下来后，整个游戏的气场就有了。

实际能多深？建议同学课后挑一个游戏写一段 system prompt，看 AI 回的是不是"那个游戏角色应该说的话"。

Land this: the LLM is not "a box that returns correct answers." It's "a box that returns lines a specific character would say."

A good system prompt gives the whole game a soul. A bad one is ChatGPT in a costume. Have students write a system prompt for one of our games after class and check the output for in-character voice.

16 / 41