模型越强，你说不清楚想要什么的代价越大。从地图与疆域讲到四种空白，拆一套我自己验证过的和模型打交道的方法，最后讲怎么让每个坑只交一次学费。

humor: 0.8 对模型来说毫无意义。我们用散文写角色圣经——五个性格，每个六十行中文。从此它知道怎么损你、怎么关心你、怎么假装不在乎。

humor: 0.8 means nothing to a language model. We wrote character bibles in prose — five personalities, sixty lines each. Now it knows how to tease you, care about you, and pretend it doesn't.

AI 宠物没有记忆就是陌生人。三层存储、混合检索、每晚做梦整理记忆——再加一个八层 prompt 引擎把记忆注入每次对话。这是让'你上次说喜欢吃火锅'成为可能的全部工程。

An AI pet without memory is a stranger every time. Three-tier storage, hybrid search, nightly dreaming — plus an 8-layer prompt engine that turns memories into conversation. This is the full engineering behind 'you mentioned you like hotpot.'

模型是被 RLHF 训练出来的，看到代码就想给建议。做 AI 宠物最难的一关是让它停止帮忙：一句 prompt 拦不住，最后堆了四层防线。

LLMs are trained to be helpful. But a friend doesn't see your code and say 'I notice you're using React hooks' — a friend says 'writing bugs again, dummy.' Suppressing the assistant instinct is the hardest part of building an AI pet.

大多数人花 90% 时间优化 prompt，真正决定 AI 输出质量的，是 prompt 之前的上下文。pretext 工程——给 AI 的"前提设定"——才是第一性原理。