12 个 Bot、一个 Chief of Staff、一个共享大脑：私有 GitHub 上的 Obsidian 库 + 周末自查 Routine

场景

多个 Bot 互相不记得对方做过什么，同一个决定要跟不同 Bot 讲好几遍，用量上涨、上下文却在流失

Bots didn't remember what the others had done, so the same decision got explained to several bots, raising usage while context kept slipping

他们怎么做

共享记忆：所有 Bot 和 Cursor 读写同一个放在私有 GitHub 仓库里的 Obsidian 库，每个 Bot 一个文件夹；任务前先读地图和所需笔记，任务后写一条短记录（问了什么、做了什么、我决定了什么）。Chief of Staff：掌握名册（每个 Bot 的唯一职责和禁止事项），收到任务后选负责人、写简报并告诉我谁接了；小事自己做，大事委派；需求凌乱时先复述再开工。周末自查：周五 18:00 读本周 Bot 聊天，标出循环、空跑和没人回答的问题；周六 11:00 按完成度、质量、成本给每个 Bot 打 1–5 分；周日 14:00 修最低分 Bot 的指令或建议退役；无事不发消息。规则：守在本职内；发布、付款、预订、发消息、删除、修改前先问；同一任务失败两次就停下问；能用脚本或 Skill 就不唤醒 Bot。新 Bot 只能经由「设计 Bot 的 Bot」/Chief 创建，否则没有记忆。搭建分两段提示词：先在 Cursor 等能跑 shell 的 agent 里建共享库，再把主 Bot 改成 Chief of Staff

Shared memory: all bots and Cursor read and write one Obsidian vault in a private GitHub repo, one folder per bot; before a task a bot reads the map and only the notes it needs, after it writes a short note on what was asked, done and decided. Chief of Staff: knows the roster (each bot's one job and what it must never do), picks the owner, writes a brief and says who has it; does small tasks itself, delegates big ones, restates messy requests before work starts. Weekend review: Friday 6pm it reads the week's chats for loops, wasted runs and unanswered questions; Saturday 11am each bot gets a 1-5 score on done, quality and cost; Sunday 2pm it fixes the lowest scorer's instructions or suggests retiring it; silent if nothing needs the owner. Rules: stay in lane; ask before posting, paying, booking, messaging, deleting or changing; stop and ask after two failures; prefer a script or skill over waking a bot. New bots are only created through the bot-designer/Chief so they start with memory. Setup is two prompts: build the vault with Cursor or another shell-capable agent, then turn the main bot into the Chief of Staff

结果与限度

「明显减少重复用量」为作者自述，无数字；文中三段提示词以代码块/图片形式嵌入，API 未返回，本条未读到提示词原文。

'Cuts a lot of repeat usage' is self-reported with no numbers; the three prompts are embedded as blocks the API did not return, so the prompt text itself was not read.

同场景的其他案例

• Mac 上的 Claude Code 给每个 Agent 写「合同」，部署成同一台云电脑上的 6 个 Grok BotX, 2026-11-10
• 用 Grok Bot 读 X 数据喂给 Claude Code：运营 273 位创作者的平台与内容引擎X, 2026-11-07
• 只跟一个 Bot 聊：路由到各项目 Bot、委派 Cursor/Codex，散步时打电话推进V2EX, 2026-11-05