Google 开源「大象—金鱼」协作法:先让长会话定设计,再用零记忆会话挑刺
社区介绍的 Elephant-Goldfish Model:长会话「大象」只负责设计文档、禁止写码;零记忆「金鱼」会话只读文档并复述、找边角;复述不过关就改计划。据称 Google 内部广泛使用,Apache 2.0。
社区帖文介绍了一种名为 Elephant-Goldfish Model(大象—金鱼模型)的 AI 协作写法:先用一个长会话「大象」把整份设计装进上下文,明确禁止写代码,只辩论权衡并产出设计文档;再用若干零记忆的「金鱼」会话,只阅读那份文档,尝试复述计划并攻击边角场景。金鱼复述不清楚,就说明计划本身有洞,而不是急着开写。
配图把流程压缩成两栏:大象侧强调长会话、全局设计、争论取舍、零代码;金鱼侧强调全新上下文、只读设计、复述解释、冲击边界。底部金句是:如果金鱼讲不清你的计划,计划就是 bug。
对编程智能体落地团队,这套方法的价值在于把「上下文过载后的自我合理化」拆开:写计划的会话与审计划的会话刻意隔离,降低「自己写自己审」的盲区。帖文称 Google 内部大量使用且 Apache 2.0 开源;本稿未核对内部采用人数与仓库地址,落地前请以官方开源仓库与团队规范为准。适合在架构评审、复杂重构说明、多服务改动的设计文档阶段试用。
Google Elephant-Goldfish Model
Google just open-sourced a way to code with AI that starts with no code at all.
It's called the Elephant-Goldfish Model.
The Elephant: one long AI session that holds the whole design. Writing code is banned.
The Goldfish: fresh AI sessions with zero memory that read the design doc and try to break it.
If the Goldfish can't explain your plan back, the plan is the bug.
Thousands of Googlers use it daily. Apache 2.0.
Honestly, I've been the Goldfish in every code review after a long weekend.
"Who wrote this mess?"
git blame: me. Last Thursday.
Would you let a forgetful AI tear apart your design doc before a single line gets written?