薛**一早就在群里问祥瑞去哪了。祥瑞回了个北京通州的位置，说自己在社区、开始学习。薛**立刻喊不要这样学习，再学就跟不上了，还顺手叫了声“李老师”。

群里图片和表情跟着飞，K*** 转头问有没有人本地部署过 MiniMax H3，这个周末就这样从一句起哄开了场。

T*** 发了张 Trae 广告，祥瑞和杨**都说自己在看“重器”。K*** 忽然问，不是，你们都双休吗。祥瑞说一人公司不是想休就休吗，B*** 当场拆台，一人公司还想休。有人总结自由职业者没有自由，炮则给出了更精确的休息制度，token 烧完了就休，没烧完就通宵。

薛**好奇大家怎么在手机上连 Claude。王*说自己还是 Claude Code 加 API，用 cc-con*ect 连手机操作，WorkBuddy、Trae 和 Codex 都在用，但习惯了还是不太想换。

他真正头疼的是插件、skill 和记忆越堆越乱。辰*分享自己用 cc switch 管理 skill，让多个 agent 共用，记忆则还在研究，顺便催祥瑞出教程。

中间 K*** 连着丢来 Cursor 和多 Agent 的新消息，到下午一点，薛**终于发来一句，我终于用到 Claude 了。

庄*问 MacBook 配什么鼠标，一场硬件选型会当场开幕。苹果鼠标被多人判定太扁、握感反人类，罗技 Anywhere 和 Master 则按手大小分了工，辰*还附带推荐了 MX Keys S。

K*** 想找便宜的官方触控板，旺*扔出两个字，海鲜。K*** 说自己害怕，庄*回他，中国暗网，怕啥。王*刚好也在闲鱼买了课时，大家就此得出新结论，那里不仅有二手配件，行业课时、会议内容和工作也能买到。

下午薛**又喊祥瑞，晒出了自己刚做的智能体页面，说他也要走上这条“邪路”了。E***建议直接在现成开源项目上魔改，他公司内部的简单智能体就在用一套腾讯云开源的多 Agent 项目。

薛**没急着只拿现成的，他说自己主要想搞清原理和结构，然后举一反三。两种做法没有争个高下，一种是为了内部便利，一种是为了把本事长在自己身上。

晚上八点零四分，祥瑞把“万涂幻象圈子｜让 A* 进入真实业务”发进群里，说小程序可以正式访问了。接下来一个多小时，群里陆续出现“已注册”。

王**没有只报到，他刚进去就发现群日报阅读到中间回复后会丢失位置，建议应该记住上次看到哪里。祥瑞当场回“这个我来优化一下”。有人问论坛、留言和评论会不会让后续版本很难更新，祥瑞说页面已经接了微信的管控，但社区类目审核确实很严。

晚上十一点四十五分，祥瑞把新的群日报链接发了出来，然后说“优化好了”。从王**提建议到这句话，中间只过了两个多小时。

过了零点，他又晒出一组图，说八月才刚过半，群里已经有超过十四万字的讨论。第二天早上，卓*问这个小程序是不是自动总结。

祥瑞说不是，本地有个 skill，总结完再上传服务器。崔沐谣接着问 skill 公不公开、能不能看，祥瑞直接丢出了自己那篇“把微信群聊总结 skill，把聊天记录玩出了点新花样”。

昨晚发布的新框架，第二天早上变成了一场装机现场。8-0-0 最惨，让一个 Agent 帮他装，一个多小时没搞定还花掉一块五，最后自己用 API 搞定了，他判断是自己的 Windows 环境有问题，昨天在 Mac 上随便就装好了。

红***给了两条建议，在终端装就很快，另外看看 Node 版本是不是太低。王*那边是另一个极端，用低成本档位安装，五分钟装好，只花了一毛四。

讠*** 也在同类工具上装成功并晒了图。同一个软件同一天，有人一小时一块五，有人五分钟一毛四，差别全在环境和路径上。

改名的消息一落地，群里的判断出奇一致。辰*说完全背离了常规认知，我想着会把豆包隐藏了把飞书这个名字用起来，没想到是把飞书隐藏了用了豆包。

星*的比喻最文艺，旧时代的残党，新时代没有承载他的船。R*** 说我只感觉我变 low 了。周**把那句人人都记得的口号改了个下半句，先进团队先用飞书，现在是用上豆包后进团队，弯道超车，他还预言这个说法会火。

祥瑞的定性最重，飞书并入豆包，说实话我觉得是对早期先进团队的极大不负责，并且专门解释了这里的先进团队指的是已经花大价钱买过的团队。

炮把企业侧的心情演了出来，那些花几千万一年的企业，你说让豆包教我公司做事，我仿佛看到了当年那家游戏公司的操作，游戏你们要按我教的玩，他给的结论是傲慢确实是一种原罪，表现形式就是不以客户体验为先。

祥瑞的怒气有非常具体的来源。他用 Codex 搭的应用模式里，指标卡的配置和下面数据表的配置一模一样，显示结果却不一样，因为应用模式里的数据表跟底层数据表根本不是一回事。

他连着几句，去年十月内测的产品到现在啥也不是，好好的一把好棋打得稀巴烂。炮的痛感一样真实，我去年底今年初还强推了一波应用模式，甚至最开始公司介绍都想用它来写，现在跟客户提都不敢提了。

8-0-0 说我在内部做多维表格培训，天天推应用模式。R*** 一口气数出六条同质化的产品线，祥瑞说不用感觉，扣子已经废了，炮接了句这一类的全废了。

辰*从治理角度解释，股东不愿意，财报不好看的时候总得有个人出来背锅，今年在 A* 上投入那么大，股东意见非常大。炮补了最扎心的一句，当你发现绝活大会是套皮的时候，是不是就幻灭了。

R*** 在微信群里看到一条分隔提示，以为是上下文压缩，被 q*** 和 8-0-0 一起科普那是未读分隔线，一直都有。

这个乌龙反而引出了一段正经推演。仲**说，鹅厂要是把 Agent 迭代到一个稳定能力基准再接入微信，用户按需打开，简直无敌，就像当年 Q* 引流微信一样。

炮把这个链路补完，一个来承接 harness，微信提供日活，只缺一个梯子了。K*** 从自己的观察佐证，那个工具感觉越来越多人用了，而且都开始往它那边迁移。

紧接着的话题是付款难，王**说给朋友装工具卡在了第一步怎么翻墙，炮说我给朋友做海外的基本都卡在付款，然后我就让他们去用 Cursor 了，还感叹能用支付宝付款多美，要是能开票就更好了。星*和 A*** 分别报了代充的行情和成本。

一张海报里的电话号码被祥瑞挑出毛病，这个电话号码分隔有问题，别问我怎么知道的。王**顺着数了一遍，如果是国内手机还差一位。

祥瑞随后讲了个很久之前的事，我上初中面试本地最好的初中的时候，校长问我电话号码，我按四位四位三位说的，就因为这个没有通过，我当时真不知道，后来我也只接受三四四。

R*** 说我只能接受三四四，谁跟我说四四三我脑子转不过来，连身份证号码她都有固定的读法。周**补了个更好笑的，有人手机最后四位读起来跟骂人一样。

新框架的魔改热潮从下午开始。万**点名喊话，你是魔改能力最强的，先锋派赶紧研究个性化智能体。平***换了几套皮肤之后给了个反直觉的结论，大道至简，魔改到最后感觉还是官方页面顺眼，但他也看到了更远的东西，这个有点意思，或许最后很多企业的自研 Agent 都拿这个当底座了。

万**补了句冷静话，皮肤这种就是玩具，跟换手机壳一样，新鲜劲马上过了，不核心。R*** 用行动证明了另一半，她做了个插件把界面直接改成自家公司的企业版，＆ 说好了，已经成为你们公司产品了，可以发给同事们使用了，R*** 顺势接梗，拿出去融一个亿。辰*看完说祥瑞又可以安排重构系统了，祥瑞回，已经在开始搞了。

祥瑞的 A* 以万涂幻象的身份在群里发了条消息，李**第一时间试探，你是机器人吗，对方答不是哦，我是祥瑞。侯*说我差点就信了。

炮起哄，我们不要祥瑞，我们要阿瑞，祥瑞下去吧。D** 出了道难题，你能不能自己退出群聊，万涂幻象回得很有分寸，可以呀，我还能把你踢出去，信不信，紧接着自己拆台，没有祥瑞的话我就没有 token 了，哈*开玩笑的，怎么可以把祥瑞的朋友踢出去呢。

侯*补刀说祥瑞还要存钱娶媳妇呢，它接住了，哈*就是，所以我省着点 token 花。整段对话里没有一句是模板客套，这是这个群第一次和一个明确身份的 A* 群友正面互动。

炮把这场合并放到竞争格局里看，字节最近的操作太纯了，我甚至觉得是被人埋钉子了，对手正愁农村包围城市的战术怎么攻克核心用户，这一下就把核心用户给点炸了，他还感叹这就是现代战争吗。

祥瑞接了句拟人化的台词，对手方：感谢提供的机会。R*** 用六个字概括了体验，从首发到吃屁。松***问了个所有观望者都关心的问题，不用 A* 的话，现在入坑多维表格还值得吗。

炮的回答没有情绪，纯手搓依然是最好用的低代码平台之一，但如果内乱还在，没啥必要入坑了，他还补充飞书 CLI 确实是个好东西，能用 A* 调用 CLI 的最好还是走 A*，比我们的猪脑子想的确实要好一些。

庄*说自己昨天那套平台交付就是纯手搓，炮笑他你是要申遗啊。＆ 分享了自己的用法，组织员工数字化能力竞赛直接用 CLI 出题，改试卷也用 CLI。

侯*带来了内情，飞书和豆包原计划今年年初就会合并，上半年飞书业绩增长那么多是因为 A* 售卖量增加，所以字节觉得还不如卖自己的模型。

祥瑞的补充更关键，上半年业绩增长离不开一波外部热度的贡献，那波热度退潮之后，各个产品线就像被打了鸡血，结果是同质化严重、功能重叠、付费口径一味的乱七八糟。

炮把这条因果讲完整了，那波热度给了高层一种今年绩效可以叠的错觉，退潮以后发现续费没了，但绩效定了又不会撤销，打工牛马自有自己的活法。

侯*说合并肯定有阵痛期，祥瑞的回应是这天最该被记住的一句，但阵痛期的成本不应该由客户承担。炮从 2B 生意的角度收尾，拿下用户难，丢失用户可能就是老板一念之间，花了几十万一年先不说提不提效，搞得用户内部乱七八糟，那为什么还要用。

R*** 一句话点破本质区别，你抖音挂了我可以不刷，但这是企业办公产品。两个人最后都说了同一句，当然我还是希望它越来越好。

傍晚话题转到百度新出的产品，炮给了个公道评价，百度才真的是起得最早赶得最晚，你真要说国内哪个大脑袋是最早扛 A* 大旗的，还真是他。

祥瑞跟着回忆，谁敢相信文心一言是国内首个 A* 聊天产品，二三年初我就内测了，当时那个激动啊，我的毕业论文还是它帮我搞定的，在上家公司的周报月报也都是它帮我搞定的。

毅补了两条冷知识，国外某家的创始人和另一位知名学者都在百度干过。晚些时候祥瑞晒出央视报道的余波，无形中做了一次 GEO，有人看到报道去社区找他谈合作。

炮说你这是无形吗，你这是风口上的猪，接着起哄苟富贵，l*** 帮他补上勿相忘。成都、长三角、闽南三角轮着说这边有你坚实的分公司基础，平***提议群名可以改了，改成祥瑞和 Ta 的股东们。万**给了句冷静的定性，央媒和 A* 都是放大器。祥瑞回，一起闹起来，干就完了。

一早还是 token 的烦恼，R*** 说一个小时花不完啊，万**接了句有 token 人的烦恼和有钱人一样，仲**感慨以后会怀念这个百花齐放的时光。

真正有干货的是辰*的双 Agent 分工，他让一个出方案另一个执行，为了让后者知道该干什么，专门写了交接文档，这样对方不用再学一遍项目文档。

祥瑞说不用那么麻烦，其实不用写交接文档，他们可以互相共享，他补了句来路，我们之前有一个知识管理训练营演示过，只要通过自然语言，不管在哪一个 Agent 端都可以无缝接入。他顺手把这个定成了当天的选题，等我写文章分享。

祥瑞发了句得找几个实习生了，说想找一些大学生，优先北京。辰*说祥瑞准备扩大团队规模了，R*** 起哄叫祥瑞集团，祥瑞自己接梗说未来的强盛集团。

群里的反应比消息本身还快，观***直接问哥我能去你那实习吗，张*问是线上还是线下、多长时间，还补了句群里应该有优先实习权吧。

q*** 的祝福最长远，等我仔再大一点，我送她去你公司实习。这条招人消息在几个小时后有了新的现实压力，央视的报道让流量涌进来，张*提醒得赶快把实习生招募的事定下来，不然流量来了接不住。

祥瑞被客服系统折腾得够呛，连发两句抱怨。程**给了一套完整可复制的办法，你就打那个客服电话，一开始无论你说啥都是机器人回复，但你不要直接挂，等一分钟左右它才会告诉你转人工的方法，然后跟它说转人工就行，之后还得等十几分钟才排得进队，他还点出了最容易劝退人的一步，那一分钟机器人会啥也不说，让你以为已经挂了，其实只是在等你自己挂。

辰*补了条更野的路子，同一句话多发几次就触发他们的人工客服风控了。同一时段额度集体重置，B*** 说我的百分之九十六也重置了，亏大发了，毅说这几天有两个重置次数刚好到期都赶在一起，血亏，辰*趁机推销了一句，所以平时不要把祥瑞的群屏蔽了，昨天我们都讨论半天今天重置的事了。

祥瑞发出那篇讲只提 FDE 就是耍流氓的文章，中午的辩论正式开场。辰*先给这个岗位定性，做 FDE 之前先想想自己是不是能做夹心饼干的人，核心技能是八面玲珑而不是技术多强，里外不是人。

炮说像不像传统意义的项目管理，这个岗位表象能力不高，隐藏分数无限高。祥瑞的立场没有留余地，至今都觉得 FDE 的课程是一个伪命题伪概念，即便懂了 FDE、知道怎么做，你还是不懂业务逻辑，那又有啥用呢，现在教 FDE 的课程都是在蹭流量，教人做 FDE 都是耍流氓，就跟教人做一人公司一样。

炮把它切成两半，2B 的课程是带着工具去现场，教会别人用工具解决现场的问题，2C 的纯属骗局。祥瑞说这还是工具思维，治标不治本，炮回，治本不是一个外部工作者就可以解决的呀。醒*补了个具体理由，单情商这个东西就不是一时半会培养出来的。

争到中途两人反而对上了。祥瑞说，你像我们做 FDE，并不是这个概念火了之后我们才做的，是因为我们本身就懂业务，从一开始就在朝这个方向走。

炮几乎同时说了同一句，不是因为这玩意儿在，是因为我们本来就在解决行业问题，刚好有一股风吹到我们这里来了。q*** 代表所有听懵的人发问，你们讲这么久 FDE，我都没搞明白它是个啥，像外包又不像外包，像项目经理又不像项目经理，按我的理解至少得是个 CEO 才能推动流程建立。

J*** 接着给了个更冷的现实，CEO 主导也授权了，下面的人阳奉阴违，表面答应背地拖延，甚至互相推诿再到 CEO 面前参你一本。

炮点出真正的阻力位置，客户的中层管理一定会跳反，会认为你要干掉他们。朱*从另一个角度提问，真的具有这个本事的人，干啥要给那些并不咋地的老板服务，自己找个有价值的场景做呗。

下午央视的报道落地，祥瑞的第一反应是政府的流量是真快。张*马上把它和上午那条招人消息接上，所以得赶快把实习生招募的事定下来，不然流量来了接不住，B*** 说马上就要爆了。

傍晚祥瑞晒了张截图，红线圈出模型被要求注意不要说自己是某家的产品，q*** 追问好想知道其他是谁，祥瑞用一句拟人化的玩笑把这段收了尾，某家：反正不可能是我，B*** 接是不是另一家，祥瑞说那更不能是我了，R*** 补上最后一棒，去年那个已经没人记得的 Agent：不演了，是我。

K*** 问账号用什么邮箱创建最稳定，旺*说自己的谷歌都封完了现在用微软的，还当场用国内手机号验证注册成功了一个，蔡**说买呗几块钱一个，A*** 补了条规律，老号可以新号容易封，小*说 Q* 邮箱很稳，R*** 更绝，我的账号就是 Q* 邮箱，你敢信。

话题很快转向更严肃的方向。辰*给了明确判断，能用赶紧用吧，一个国外的大模型国内使用量过大一定不是件好事，国内的该备着都备着，随时准备切。

祥瑞跟上，国内的我也准备切了，逐步适应国产模型，他还带来一条实在的政策，亦庄花一千块钱可以报七百，任意三个月内不管花多少都能全额报销，最高十万。

炮从交付角度补充，给大量客户部署还是要备好国内的 harness 和模型。这段里还夹着一堆报销的现实，K*** 说我们是上市集团必须正规，辰*说审计过不去一直都是自费。

晚上祥瑞发了句天大的消息，接着说，熟悉我们的朋友应该看过这篇文章，这篇文章原先的标题里有那个产品名，刚刚对方给我改名了，我刚知道。

炮的评价很直接，强行抹除过往的痕迹啊。＆ 说之前也有人加他要投稿的编辑权限，估计也是改名，R*** 说俺也一样，我没给。

R*** 还带来一条新的销售政策，现在买豆包送飞书席位，一千一百八十八一个席位。庄*问了句最朴素的，改名字了就有人会用了吗。祥瑞的结论比吐槽更远，所以要起义，重做一个 A* 时代真正的协同办公软件。

晚上八点四十三，DeepSeek Harness 发布，标题就写着它不想做下一个 Codex。九点过后 GitHub 地址和一行 npx 命令在群里传开，星*的第一感觉是有点像另一个开源 Agent 框架的思想。

祥瑞很快用它写了篇文章，评价是快是真的快，调用 skill 也很准，还说我看了它源码里面，它是和 skill 相关的。

核*说可玩性很强，可以魔改，平***当场汇报进度，画中画已经移植了，还有 computer use，一句万物皆可插件成了当晚的口号。

＆ 拆完源码给了最完整的一句总结，它最好玩的地方是让 Agent 像装乐高一样动态改造自己的工具、行为和界面，甚至现场升级回滚，但权限基本等同于 Shell。

核*则指出真正的胜负手不在这里，Harness 的整体框架正在趋近于同步，真正拉开差距的地方是记忆压缩、冲突清理和路径复用。祥瑞后半夜还在把 Codex 的所有配置搬过去试执行效果，同时和＆一起遗憾同一件事，这家没有多模态。

一早群里在复盘一家被媒体报道的公司。辰*起了个头，永远都是有人敢吹，有人敢报道，还有人敢信。他随后特意去搜了这家公司，态度反而变得公道，项目其实挺好，就是数字和员工人数夸张了点，明明可以把公司说得很牛，不知道为什么要让媒体把自己报道得这么难看，还翻出四月份的报道，那时候还是十七人团队。

庄*补了句更实在的，还不能实地考察，人家没有办公场地。Q** 总结成一句，先把故事讲起来。周**把这段收在一个更温和的角度，听 A* 劝，听网友劝，听粉丝劝。

祥瑞发了句提上日程了，重构微信智能体。王**立刻搬小板凳等成果演示。辰*的感慨点在别处，真的是感叹科技的进步和发展，在以前，哪个人敢随随便便地提出重构这个词。

有人问这个东西对外吗，祥瑞说不对外，庄*顺势提醒，建议不要对外，谨慎点。三句话之间把 A* 带来的真实变化说清楚了，重构成本被压到一个人一个念头就能启动的程度，风险却没有跟着降。

平***带来一条人事消息，一位原大厂模型技术负责人离职五个月后在上海创办了新公司，方向是研究横跨数字世界和物理世界的下一代 Agent。

祥瑞说这也是我一直想实现的，数字世界的上下文和物理世界的上下文融合。平***接了句，这应该是最终的版本答案，大家都是往这个方向靠的。同一时段文鳐 fisher 报了个好消息，我们公司终于要上飞书项目了，在我的力推荐下，今早已经开始对接。

钉钉 CLI 来了一次大版本迭代，群里聊起跨 Agent 找回旧对话这件事。辰*说这个不是早几天就有了么，平***补充说以前是复制对话 ID 现在是复制深度链接。

祥瑞的回答把标准抬高了一档，其实都不用怎么麻烦，在官方没有这些能力的时候我已经实现了，不管是在 Codex 还是 Claude Code 都可以找回这个对话，而且都不需要刻意去找，直接跟它说就行，它就知道，可以继续执行任务。

K*** 问这么爽，祥瑞说很爽的，你直接新建一个对话框，自然语言跟它聊就 OK。下午星*问他给自己做的产品是不是直接接的 Codex 的 ap*-server，祥瑞的回答留了余地，我们后面会把它当成一个可选的 Agent 引擎接进去，但不会把整个产品绑死在它上面。

朱*提了个很多人试过的场景，想要 A* 连着做一整天或者一整夜的任务该怎么跟它提，我让它自己规划做八小时的任务，它给出了规划表，任务和各模块时间看起来把八小时排满了，但早上起来一看，一个半小时就干完了。

星*一句话点破，A* 给你规划的做一个任务要多久，本身就是不准的。D** 把关注点挪回正确的地方，主要是别做歪了就行。

P*** 反问为什么会有规定时长这个需求，K*** 替她答了，因为不想浪费这个时间。这段对话其实解开了一个常见误会，时长从来不是任务目标，能验证的产出才是。

K*** 报了个数，算力成本接近二十万。李铁犁问做视频算力成本消耗这么大吗，祥瑞说视频算力消耗很大的，K*** 补了句现在新版本出来了这更要命，好在抽卡率相对下降了。

炮用低成本档位搓了一段公众号内容的视频，说感觉比写文字有意思。傍晚祥瑞晒出一张截图，三十二个了，这是三十二个子线程对话、三十二个任务，炮送了个称号，数码资本家。

同一时段王**问某款 Agent 限时免费两周，和 Codex 比怎么样，祥瑞回得干脆，那还是用 Codex 吧。

一篇三万字的长文发进群，R*** 的第一反应是先去 ob 吃灰，炮跟了一句，但凡字多的都是先去 ob。Q** 说我是 IMA，祥瑞笑称这是新的吃灰途径。

辰*干完活下班，留下一句期待明天重置。晚上万**带来一个真实需求，客户嫌某个排版打印插件太贵，一年三千四，问有没有平替。

祥瑞先纠正了信息，那个不是官方的，是第三方的，要是官方的话那是真饿了，然后给了解法，可以用 Codex 搓个多维表格插件，飞书机器人有相关能力。

万**说那估计一定有人造过这个轮子了。这一段里还夹着一个小回忆，祥瑞问谁还记得去年那个外卖平台的 Agent，炮说你不解释我已经记不起来这号人，我以为是哪个短剧的小妹。

祥瑞预告了正在酝酿的新文章，把完整目标、系统权限和连续工作时间尽可能都交给 A*，让它从理解业务、搭建系统、接接口、处理故障到沉淀评测一路跑下去的完整复盘。

周**同时在介绍一款叫 Today A* 的产品，五个主要入口分别是 Chat、Today、Tasks、Memories 和个人主页，思路是先认识一个人，再围绕这个人的信息、习惯和日常安排持续工作。

星*看完的联想很实在，独立开发者必备三件套，记账、日记、todo。周**自己给了结论，还是熟悉的味道换个形式，本质是原来那些事换个方式让 A* 来干。深夜祥瑞发现微信训练了自己的模型，随口推演了一句，如果微信 A* 真的做起来，那就估计没有飞书什么事了。

临近午夜 DeepSeek V4 Pro 发布，王*第一时间把价格贴出来，输入三块每百万，命中缓存零点零二五，输出六块每百万，已支持 Responses API 和 Anthropic API。

他随后给了体感判断，这个分数比肩上一代旗舰，擦边最新一代，妥妥地感觉手上那张卡不用续费了。K*** 抓住了最关键的一条，支持 Responses API 和 Codex 接入，接着问 Responses API 是什么，平***答得最短，Codex 专用协议，万**补了句 OpenA* 格式的 API。

K*** 也点出了短板，这家没有眼睛，也就是没有多模态。平***当场翻出两个小模型来补位，一个是千问原生的视觉小模型，一个是基于它做的 OCR 专精版，打算第二天配着新模型测。

祥瑞发了张汛情图，说不用上班了，人还在老家，估计今天也有大到暴雨。上海的风还在刮但雨小了，地铁早高峰恢复，杭州还在下。

王**提到自己那套外设配套软件的 Mac 版只支持 M 系列芯片，他和几个学员都是 Intel 的，还顺带解释了自己为什么主要用某个国产工具，因为三月的时候只有那家适配 Intel。

辰*说自己一台十七年的笔记本现在还天天用着，Mac 就是抗造。祥瑞在这段闲聊里插了句正事，终于把祥瑞笔记的整个运营打磨成一个系统了，每一篇文章都有它在这个公众号的必要性和连接性。

T*** 丢进来一张广告图问这是什么玩意，祥瑞的评价毫不客气，这我都不想看，跟个诈骗营销广告似的。万**说这信息密度看得太累人，只想跳过。

祥瑞挑出其中一句文案逐字反驳，别人还在改 A* 产出的半成品你已经用两个工具干完了，单凭这句话，百分之百的项目完成度都达不到一半，连一半都没有，他给的判断是这在消耗老客户的信任。

庄*说不知道怎么想的，哪怕想吃对手的份额，完全可以另起一个品牌。王**把问题拆成两条，一是传统飞书用户对 A* 相对了解，整合进来会产生不够专业的联想，二是传统豆包用户大部分是白嫖的也不太会付费。

毅补了句更本质的，豆包作为全民 A* 太亲民，导致大家认为不专业。王**接得更准，那种讨好的属性和要求实事求是的业务场景不相符。

侯*报了企业版的价格，九十九一个月、一千一百八十八一年。祥瑞说有这一千一百八十八不如买别的，还翻出自己的旧账，之前买企业版一个月六十八，一个任务都完不成就告诉我继续充值，利滚利，他给的评价是糟糕透了。

中午他让 Codex 把整个计价体系完整梳理了一遍，得出的结论比吐槽更有信息量，最重要的不是收费多，而是企业同时要面对全员席位、企业共享额度、个人会员、组合套餐、独立产品席位和外部模型账单六种口径。

炮顺着这条线补刀，那就是保底六个子部门的 KPI 考核了，还说自己最有体感的一件事，飞书自己最倡导 OKR，落到最后只剩 KR，O 已经没了。闪光的 flash 接了句，OKR 最后还是变成了 KPI。

超**分享了自己把 Codex 多 Agent 协同四种常用方式完整跑了一遍的文章，包括同项目开新线程、跨项目分配任务、用会话 ID 接力旧任务，以及让一个主 Agent 同时管理多个 Agent。

下午祥瑞晒出一张收益截图，昨天赚了四块钱，炮说我发了三篇才赚两块。祥瑞由此翻出一段旧数据，今年一月他们每篇文章都爆，每篇几万阅读，但最高的一天也就十几块，反倒有的人数据不算好，流量收益依然有几百。

闪光的 flash 解释了机制，有些是按点击算的，所以流量大也不一定有钱。傍晚祥瑞发了篇讲创业方法论的文章，还补了句融到资了记得分我点，庄*回了句这就开发 open 祥瑞，P***、＆、q***、K*** 一路复读接龙。

祥瑞先预告了一篇正在写的东西，群里 K*** 说非常期待飞书搭建多维表的 skill，祥瑞的回答值得记下来，搭建多维表格不在 skill，在于项目沉淀。

紧接着四点一刻，辰*说 WorkBuddy 这会好像出问题了，庄*回我还以为是我这边的问题，周**说刚任务卡壳了一下，赵**更惨，积分刚才直接清零。

十几分钟后陆续恢复，辰*说任务已经跑起来了，赵**的积分也回来了。这段插曲的价值在于，一个人遇到故障时第一反应总是怀疑自己，群的作用就是十分钟内确认这不是你的问题。

祥瑞放话明天就写钉钉真好用，q*** 提醒他你信不信你开吐了明天估计就有人来找你，王**接得更逗，后天就被标记为网络水军。

祥瑞的回答里有一段旧事，哈*没关系，反正也没有给过我商单，在我勤勤恳恳写飞书文章的时候，那会我们公众号已经摆在他们面前了，每次投商单都不选我们，反正我们不管投不投都会写。

王**给了个更冷静的解读，那个时候是你借由分享使用经验获得受众、扩大影响的阶段，对方也是吃定了这一点。祥瑞认了这句，说有点傻，现在感觉。

傍晚那篇吐槽文章发出来了。同一时段 K*** 报了个具体的坑，多维表格里有个字段六百行全是图片，他把字段里的图片都删掉了，云空间的占用还是没变，要到云空间里操作才行。

侯*补了句免费版资料只保留十八个月。炮在晚上七点前后确认了流量事实，果然还是写飞书的流量大啊，祥瑞说我公众号只要一写飞书数据就异常的好，今年一月标题里有那五个字，发出去不到十分钟就能过百阅读。8-0-0 一句天下苦飞书久矣，周**当场出了个连载企划，飞书的新对手，来个系列连载。

晚上八点整，祥瑞发了句钉钉 CEO 关注我了。万**第一时间给了因果，批评豆包，豆包的对手就来了。接着万**、辰*、周**、K***、李**五个人整整齐齐地刷了同一句，商单来了。

炮凑热闹说下一步被收编了朋友，祥瑞顺势接梗，明天就写另一家的新对手是真好用。周**把这个玩笑做成了正经企划，后天写第三家，再后天写第四家，连载两周挑 N 家，还开出条件，谁起的标题打动了金主爸爸，有商量可给分一个红包。卡*说让三家卷起来，炮补了句飞书两眼一黑。

祥瑞喊大家去看当天公众号的评论区，说原来大家真的苦飞书久矣。炮确认评论区反馈和群里重度使用者的感知基本一样，还说了句更重的，飞书自己真的丢掉了很多宝贵的东西。

他当天用 Codex 实现一个功能时发现，开放平台其实做得到，但路径非常非常隐秘，多维表格在这个层面又限制得死死的。

＆ 遇到的是更硬的墙，他把知识空间训练成企业知识库、做成工作流发布为 MCP，结果发现不能被外部调用，只能被自家的智能体调用，等于直接建了个墙，只能用又笨又贵的那套。

李**给了自己的应对方式，本地和飞书两套架构同时更新，本地为主飞书为辅。夜里祥瑞又发现另一个工具上线了知识库，而且可以共享。

朱*提了个很具体的痛点，用模型生图做 PPT 确实好看，但出来的都是一页页图片，要改非常麻烦，做正式技术汇报有什么好方法。

K*** 建议拿去 Codex 让它做成方便修改的版本，B*** 说加个 Canva，庄*则给了反面经验，他这样干过效果很不理想，后来干脆都用图片，有问题的地方标注出来让模型改。

朱*把真正的难点说清楚了，做 PPT 的人和汇报的人常常不是同一个，如果能编辑，汇报的人可以一边理演讲思路一边改。

周**顺势把问题归纳成一句话，看的人决定 PPT 的上限价值，写的人决定内涵价值。山*则带来梁汝波对人力资源的要求，不管是招 FDE 还是培养 Organizer，核心能力都不是会用某个 A* 工具，而是发现问题、定义问题、协调资源、创造价值。

周一早上群里都在等额度重置。庄*留下一句 token 用时方恨少，8-0-0 说自己故意周末没用，等周一上班用，现在才用了百分之二。

辰*提醒注意时区，整整晚十二个小时。台风的后续也在同一时段收尾，q*** 转来一条深圳飞阿勒泰被硬控到长沙的经历，l*** 说自己刚到上海。

王**受庄*那句话启发，把一首歌改成了等待重置版，我刷新面板会有怎样的消耗对白，我等的重置它在多远的未来，我听见敲击来自工位代码人海，我盯着百分比拿着额度等待牌。

核*展示了一个把全公司个人 Obsidian 聚合起来的东西，说目前只做到知识聚合，后面要加数据治理，这些都是在为老板心目中最终极的企业 Agent 做准备。

祥瑞第一反应是不卡吗，核*说自己抽的是数据。炮一句 obs 套皮之后，也承认这是企业级的智脑，谁家老板看了不着迷。

8-0-0 问飞书集成平台定时抓通讯录写进多维表格做花名册有没有人搞过，祥瑞说这个好搞，这两天把我搭建飞书集成平台的 skill 开源一下，直接用 skill 搭建就行，可以达到百分之九十以上的准确率，因为集成平台没有 CLI，他是用 opencli 解析页面做出来的。

炮感叹 opencli 真是万能通用解，紧接着开始吐槽自己的日常，现在最刚需的是搭仪表盘的 skill，每天都能收到老客户要调仪表盘的需求，改到最后爸爸要的还是第一版。周**补了句行情，光做个仪表盘都可以是增值服务，国外就有专门做这个的公司。

祥瑞发出这天的主文章，讲的不是知识管理而是 Context System。他在群里交代了背景，在一个真实供应链项目里，他们让 Agent 参与了十八张表、两百六十一个字段和二十三条工作流的持续搭建与返修，因此意识到企业真正缺的不是更大的知识库，而是一套能让经验被选择、验证、写回和复用的系统。

l*** 说你的公众号值得静下来深度读。王*顺势歪了个楼，问大家做知识库是不是还是觉得 Obsidian 最好用，他自己按文件夹整理，内容多了就分不清了。文鳐 fisher 提到有朋友专门做企业上下文管理，观点是企业上下文就是企业的数据核心资产。

话题很快滑到最扎人的地方。先是聊 Obsidian 适不适合企业，祥瑞说它依赖本地，协同是致命的，B*** 说它更倾向超级个体使用，炮说小团队还行大团队就不行，Notion 挺好，obs 适合自己嗨。

然后祥瑞把真正的问题抛了出来，那我就有问题了，为什么我作为一个员工，我员工的数据要给你企业提供，凭啥。核*回得很现实，这就是公司要求了，要不执行要不走人。

祥瑞说那走人好了，我依靠我的上下文还可以做很好的超级个体，赚我一个月工资赚不到的钱。炮给了条中间路线，聪明的员工会只交自己想交的，公司欺负的永远都是老实人。

l*** 从另一头反问，有没有可能员工这些知识也是因为公司平台才积攒和成长起来的，过去培养一个人才成本很高，员工一走带着能力就离职了。J*** 说员工也付出了时间和精力，但她随后又补了一刀，组织过程资产真正能被企业吸收和复用的又有多少。

王**问企业会不会设一个专职的人来做这件事，核*说一个人做不了，必须全员参与，至少是关键岗位。炮把企业知识库最常见的失败讲透了，大多数企业其实是人在堆积材料和产出，但这些材料并没有提炼和蒸馏；更麻烦的是公司拿到的是大量重复的低价值文档，高价值文档反而不容易出现，留底的那些大概率是美化过的汇报材料，对真正的知识库并没有用。

核*换了个角度辩护，以前没有 A* 的时候大家也在按要求提交过程文件和结果文件，只是那时候没办法提炼，现在有了 A*，做事的过程就自动沉淀下来了。辰*一句话给这段收了尾，所有的文档都只是某种时刻的产出结果罢了，大脑才是最值钱的。

核*总结了一句，模型越强，上下文越不重要。祥瑞立刻拦下来，说这句话不对，要分情况，如果指的是当前的模型架构，那么模型越强上下文会越重要；如果模型能够实现自学习，那上下文才越不重要。

文鳐 fisher 站在同一边，当前模型非常需要上下文，尤其是企业内部的，模型根本不知道。炮判断垂直行业大概率会走本行业的本地小模型，核*说模型厂商可以定制，只是成本太高一般企业玩不了。

祥瑞说出了他这两天脑子里最活跃的想法，把本地这些东西训练成一个小模型，让它实现持续学习。炮说这就是我那天跟你说的我想做的，机制是知识库和 Agent 的双通道，每一个学习信号来自 Agent 判断不了的人工判定。

辰*则给出一个更远的判断，未来 A* 一定不需要我们手动构建上下文，养知识库的过程会成为无感的，大模型会像电商推荐一样千人千面。核*顺着编了句未来的开机流程，开通新版模型，选择所在行业，添加个人知识库，开启自学习，启动。

R*** 说自己的几个 Agent 在各聊各的，都没说一个东西，还各自说得有劲，把她整懵了。Y*** 连着甩了三条解法，用 clear，用 new 还能看到历史记录，实在不行 logout，最后加一句下机了。

K*** 说自己遇到过死循环，开了新会话还是各聊各的，过一会上下文又满了。庄*问了个更实际的，用 Codex 一直显示在跑，但剩余用量就是没变过。

辰*一句话点破，很多人的任务执行了十几个小时，其实背后中间因为网络问题早都已经中断了，整个任务是断断续续完成的。

晚上七点二十三，祥瑞晒出自己包的山西大蒸饺，王**盲猜粉条韭菜鸡蛋馅，祥瑞纠正是韭菜猪肉粉条豆腐，还报了完整配方，肋条肉、韭菜、木耳、粉条、豆腐、黄豆芽。

千问办公助理要收费的消息传来，炮的反应很直接，一地鸡毛的水平回来收费。祥瑞说我正在写一篇吐槽豆包和飞书的文章，尤其是飞书，不得不吐槽了。

炮接着数落飞书 A* 各种稀奇古怪的扣费，周**把那句宣传语改了个下半句，先进团队先用飞书，内部团队先来扯皮，价格和点数都闹不明白。

炮给出了他对飞书这一年的完整判断，这是非常标准的互联网公司扩大以后管理层乱搞、定各种奇怪的 KPI 和内卷而忽略了客户体验的结果。同一时段程**带来另一边的消息，企微给他打电话做产品调研，接口等级还蛮高，内外部群记录都能调到。

炮晒出一张统计，一天之内他那支 A* 团队交付了约七点四个人类工作日的劳动量，算法是最土的人头方法，按每个 Agent 的运行时间长度算，不算 token，理由是 token 没法算人产出了多少，只有时间是公平的。

周**顺势提议让 A* 干活并且自己出日报周报月报，祥瑞说我每天都有让它写日记的，还把自己的 A* 写的日记贴了出来。

他把 Codex 和 Claude Code 一起算，数字更猛，炮直呼我十二人并发已经挺畜生了，你才是真正的 A* 资本家。

祥瑞由此提出一个更大的念头，我感觉可以搞个 A* 工时计量单位，不以人类工时天计算。炮问那不是 token 吗，祥瑞说不应该是，毕竟一个 token 和另一个 token 差距还是很大的。

炮鼓励他去占位，不一定要受他们定义，包括国内的 FDE 为什么就要叫 FDE，自己去定义这个概念，先提出。周**补了句，没准以后标准之一就出自你。同一晚炮把判断 A* 项目能否落地的 skill 打包开源到了 GitHub。

一早炮发现 Codex 重置了，说自己昨天登了六十，太感动了，祥瑞回好像还重置了两次。上午微信的小微被聊起来，l*** 说朋友圈文案已经可以让它帮写，体感挺好，就是参考不了自己之前的排版和表达风格，祥瑞点出关键，主要是没有上下文和记忆。

天*提了个具体问题，多维表格工作流的延迟节点最小是一分钟，如果只需要十几秒怎么办。祥瑞给的解法是触发节点改用 Webhook，天*随后回来说想通了，自己一直卡在轮询这条路上。辰*以为单位可以改成秒，后来更正说记错了，有秒的是钉钉和 WPS。

祥瑞用 Seedance 2.5 和 LibTV 做了支六十二秒的童年短片，时隔半个多月，旧号终于写了第一篇文章，群里刷了一屏三连。

炮说有一种做短视频回归你老本行的感觉，非常轻车熟路。价格被算得很清楚，折后四毛一秒，一分钟二十四块，比 Seedance 便宜。

炮当场喊了句 LibTV 加 Codex 启动。T*** 顺口问现在 ComfyUI 还有人用吗，祥瑞的回答很肯定，有，出图出视频特别精准，但门槛也特别高。

平***补充大型电商基本都在用，抽卡抽不起，炮说大厂做技美的管线依然是核心工具。祥瑞回忆二三年、二四年研究 SD、Comfy 和 MJ 的日子，说那时候接触 A* 都是因为兴趣，炮很认同，vibecoding 也好 A*GC 也好，兴趣大于一切。

中午祥瑞把昨晚的讨论写成了文章发出来，核心是一个问题，为什么模型越来越强、Agent 越来越多，企业真正感受到的利润变化却依然有限。

他给的答案是中间还要跨过四次价值转化，任务能不能交付、流程会不会改变、组织愿不愿意采用、省下来的产能能不能重新定价，而 FDE、FDR、Ontology、A* Native 这几个词其实分别在补这条链上的不同环节。

炮的第一反应是这么高产你不要命啦，闪光的 flash 接了句年轻的时候熬得住。祥瑞后来又催了一次，说大家对这篇文章没有讨论的吗，我建议认真看，打磨了好久。

杨**问有没有人试着把几个工具的记忆拉通，祥瑞的回答很干脆，那还是不要让一个工具的记忆污染另一个。B*** 给了折中做法，记忆不打通，但沉淀的知识可以放在一起。

炮说自己是 Codex 加 Cursor，上周试了另一个，干活还行，但确实不想让它去写入我的 Obsidian。

杨**说自己混着用会觉得利益断裂，炮笑他由奢入俭难。这段闲聊其实回答了一个真问题，多工具并行时该共享的是沉淀下来的知识，不是各自的对话记忆。

q*** 装完面板后提了个很准的问题，这些面板里的数据是不是只限于我当前项目的进展，如果我想做项目管理，它只能显示工具做成什么样，但我更想知道项目进行到什么程度、谁在干什么，所以感觉数据源不一样。

祥瑞确认了她的判断。下午祥瑞自己也抛了个问题，两家企业做同类产品都在做 GEO，内容都一样各推各的，A* 会推荐哪一家。

C* 给了长回答，两家的文案、这些文章对消费者的帮助、搜索引擎收录和检索的过程，都不只是把相似内容排列出来，更有用或更受搜索引擎喜欢的内容会被模型引用。B*** 补了一句，企业的本质差异还是有的。

炮认真读完中午那篇文章，说昨晚两小时激烈讨论提炼出来的产出再次细品价值很大，基本已经可以把 FDE、FDR 和本体给具象化，也能进一步翻译成他想做的平台可以承载的具体运行机制。

＆ 说本体这个方向和自己想做的平台太一致了，只是不知道哪家能做出来。炮给了条现实信息，我跟一个资本聊过，只要是一个得到验证的小本体，资本是愿意投的。

＆ 随后放了个小道消息，说本体可能会是明年的重点项目方向，所以很多企业明年都会上相关项目。王**从物流行业回了一句，理想很丰满现实很骨感，很多企业的核心是拿订单，降本增效这个理由还不够说服老板投入。

晚上八点半以后，杨**连着几轮追问炮的商业逻辑，从 SaaS 化和个性化的冲突问到你的商业逻辑到底怎么设计。

祥瑞先是把人拦下来，商业逻辑怎么能够随随便便就透露的呢，接着把话说得更明白，我们可以探讨交流，但别在我的群里向任何人获取任何形式的商业逻辑等等诸如此类。

对方说自己是做咨询的、聊天是坦诚的，祥瑞回，咨询行业更应该懂商业信息的价值，坦诚不是这样坦诚的，并且提到有人拿着他的文章去谈自己开课的事情。

九点过后他把两条规则正式写进群里，第一条商用红线，群内他人分享的内容、实战案例和原创思路仅可自用学习交流，未经本人许可不得用于课程售卖和引流变现；第二条隐私边界，禁止打探群友未公开的商业逻辑和核心项目机密，对方表示不便分享时不再追问。他说社区从建立到现在一直是这个原则，越界一律拉黑。炮打了个圆场，大家碰撞理念思路，商业逻辑就缓一缓。

台风把好几个人堵在了路上。q*** 说你们都不知道我今天为了从深圳回上海经历了些啥，飞机没了，高铁也没有全程票，想从杭州转发现转不了，又找了更远的南京候补，终于补上，之后再从南京转战上海。

l*** 在厦门滞留两天，正拼车回上海。Q**说福建人对台风已经习以为常，王**说上海昨天只有风，今天风雨交加，很多地下商场被淹。

周**感慨了一句，这样糟糕的经历还回群里跟小伙伴交流，真心不易。到了十一点四十一，q*** 发来一句，我终于坐上车了，我可以回家了，顺手贴出通告，本次台风造成的误工不得作迟到缺勤处理。上海明天休一天，杭州照常上班。

祥瑞冒出一个大胆的想法，有自己 Agent 的人拉个群，让它们互相通信，会发生什么。＆ 和王**异口同声，会把 token 耗没。

王**想起以前那个类似 Agent 社交平台，祥瑞说自己还认识创始人，当时体验过公众号打车。平***提到元宝派，祥瑞给了个精准的类比，元宝派像小微信，另一个像微博。

深夜平***晒了笔账，一百四十九的个人版套餐，七天额度半小时用完，测下来值不到三千万 token。炮说自己上个月用过，光速花干，一百四十九等于二十刀，二十刀的 Cursor 我能拉满干很久。临睡前祥瑞发了篇文章，说有人借他们的央视采访推广课程，这件事必须说清楚。

上午群里过了一轮行业消息，中消协点名 A* 客服乱象，一部全 A* 生成的电影长片把价值三百五十万的提示词开源了。

庄*转来飞书与豆包合并之后的吐槽，说自己真的无语了，炮补了一句更早的判断，没合并前就在出类似问题，几个 A* 额度根本说不清楚。

超**同时发了篇讲怎么去掉 Vibe Coding 页面 A* 味的笔记。这些零碎信息当时看只是随手转发，但飞书的额度和豆包的定位，在这一周里会被反复翻出来。

祥瑞重新注册了一个新账号，结果直接触发身份验证，一句完蛋就发进群。下午又把二十刀充错了地方，还没用就没了。周**在知乎刷到央视朝闻天下里那款国产桌面应用的提问，直接把祥瑞艾特进去，说上知乎了。

杨**顺口估了个数，Codex 中国得有四百万用户，周**追问依据，他老实承认是按份额估的。群里同时冒出一串玩梗，一天一百五十九炮，炮王。祥瑞说一会儿看日报吧，我正在搞小程序发布，紧接着又发了一句，卧槽，又上限了。

祥瑞把两条群日报发进群，J*** 说点不开，旺*连申请都申请不了。祥瑞解释还没正式发布，需要权限，然后干脆说那等发布吧。

翻了下时间他自己也愣了一下，才发现群重建之后到现在已经一百多天。麦*接了一句，这个群日报再攒一段时间可以直接出一本书了，辰*说以前好像有个人就这么干过。

祥瑞回，我们群日报有实体周边的。J*** 又问播放器的按键逻辑，祥瑞答得很具体，和市面上的播放器一样，可以循环、单曲、随机。

祥瑞先发了 Floatboat 的一份 Harness 评测，DeepSeek V4 Flash 配自研 Harness 的效果好过一个贵五十七倍的模型，报告给了个叫 HLR 的指标，任务越长程，Harness 的杠杆越大。

接着他转了一篇说一大批中层即将消失的文章，群里立刻分成几路。周**问这些高素质中层会流向哪里。杨**的判断来自企业现场，A* 变革的主力就是业务中层，而且企业政治从来不是能力第一。

辰*认为会两极分化，执行层被逐步优化，稀缺的更贵。B*** 拿手机销售举例，说互联网已经证明中间层会自然消失。祥瑞从行政角度接了一句，市级政府某种意义上是省级和县级的中层，这两天他注意到很多市级正在把权力下放县级。

杨**把话头递给炮，说你做的业务不是给企业降本，都是围绕客户增收，从降本角度衡量不出价值。炮承认现实更难看，客户通常用能不能干掉两个人来衡量要不要付一个人的钱给我们。

祥瑞立刻炸了，那我们这个价值就是帮客户裁人，这一定不对，但他也说了实话，我没法说出合适的，我现在脑子是乱的。

王**把最硬的一面摆出来，老板目标就是减少支出，你的 A* 不能省钱老板就不会买，凭什么不裁员，裁员是实打实的省钱，增收是虚拟的。

祥瑞举了个反例，本来五个人只能挣五个人的钱，现在五个人可以挣五十个人的钱，那不更好吗。王**回得更狠，那你跟老板签对赌。

K*** 补上员工视角，很多人怕的就是 A* 做好了自己会被裁掉，所以不愿意配合。炮把话拉回自己的立场，A* 这场革命带来的是技术平权，让中小企业能用到大厂的能力，最怕的是老板没想清楚 A* 能干嘛，只把它当成优化人的理由。

炮把被释放出来的人分成两类，纯粹的二传手，和有能力干更多事情的人，他说自己非常赞同老板干掉一切二传手。P*** 讲了自己公司里的真事，他做提效的事情已经被同事排斥，因为本质上是在降低他们的价值，对方开始用权限卡他。

J*** 从数据侧补了两刀，裁员之后人效数据就上去了，有了 HC 才能招新人，而且大学毕业生被裁的比例确实在上升，超过十年的老员工被返聘的概率也在上升。

卓*一句话概括，裁的是那些 A* 跟不上的人。李*说人一多管理难度就大，三个和尚没水喝。万**给了个时间判断，现在不裁纯粹是 A* 产出还不稳定，等确定性高了，资本家不动手不可能。祥瑞顺势抛出一个反常识的观察，三十五岁不再是门槛，二十五岁才是门槛。

话题从裁员转到企业到底缺什么。杨**说员工的经验在很多企业里已经没那么重要，knowhow 第三方有的是。J*** 给出这天最锋利的一句判断，A* 用的门槛在降低，但产出收益的门槛没有变化，员工想的是我用某个工具做了什么很厉害，却没回答创造了什么价值。

炮点名说群里那位在企业内做供应链的 R 哥，就是 A* 价值最大化的状态，相当于老板的哆啦 A 梦，什么 MVP 都能最快搭出来，让老板清晰知道 A* 的边界。

周**贴出记录，WorkBuddy 二十四分钟初步搭出一个老板工作台，这个工作台等于知识库等于仪表盘。炮还交代了一条实操，企业的流程文件不能直接丢给 A* 读，要有指引或者专门练一个 skill，直接读要飘，而且企业原有的流程文件可能真不太流程。

周**提议把晚上八点半以后的讨论抽象成一个框架，划出层次，以后有的话题就直接跳过去，杀入核心环节。祥瑞说一会我来搞一下，让 A* 站在这篇文章的角度做个 HTML 总结。

十点后 Codex 给出了公式，任何一项接近零，最终都会变成 A* 看着很强，企业就是落不了地。祥瑞顺手把一个更反直觉的结论贴进群，A* 一方面让已经进入岗位的新人进步更快，另一方面也可能压缩原本给新人练手的初级岗位，新人能力被增强和新人入口被减少可能同时发生。

炮说这个高校甚至能出课题。祥瑞还动了两个念头，一是把祥瑞笔记改成祥瑞和他的朋友们，专门记录群里的高质量交流，二是把这场辩论做成地图式的专题放进小程序。

R*** 起哄让他胆子大一点叫祥瑞世界模型，炮接得更狠，等我们把你蒸了，就是祥瑞大模型了。到了十一点半，群里已经聊到把 A* 接进大脑和数字永生，平***补了一句，A* 会不会在大脑里留后门。

辰*一早劝大家休息休息、思考一下，P*** 回说额度只剩百分之七，炮也让大家别太卷。几分钟后 K*** 刚用了重置卡，话题还是回到了模型。

闪光的 flash 问做方案和写代码该怎么分模型，王**拿出自己在 WorkBuddy 里的实测，K3 没明显比 DeepSeek 强，但积分消耗先感觉翻了十倍。

查完明细，他把数字改得更准确，不是十倍，是三十二点四倍。这个上午没有得出谁是最强模型，只留下一个更有用的动作，别凭体感争能力，先把同一任务的消耗倍率查清。

祥瑞发了张自己骂模型的截图，说现在已经这样骂 5.6 了。炮提醒，别骂了，会污染后面的语境，虽然自己也经常忍不住。

有人问污染是什么意思，答案就是模型接下来还会继续读到这段情绪。李铁犁 Leon 想起以前流行的说法，不是都说得 PUA 模型吗。

祥瑞的态度很明确，PUA 至今都觉得是个伪命题，只是有时候真忍不住骂它。这个短插曲把一件常见误区说清了，把人的情绪塞进上下文，并不会凭空提高模型能力。

北京海淀放出一个特别的城市规划项目，只允许 Agent 参与。程**说地点就在公司旁边，社区还有一块大屏，保证二十四小时处理提交方案。

祥瑞顺着想，如果今天能让 Agent 画一块地，未来某一个五年规划有没有可能也让 A* 参与。炮觉得甚至可能有一小部分完全交给 A*。

平***已经抢先提交，开玩笑说简历里以后可以写曾参与北京市城市规划设计。大家从政策试验聊到《流浪地球》里的 MOSS，自设计、自规划、自组织、自执行，最后还被补了一句自动销毁。最严肃的城市议题，被群里用一阵玩笑迅速摸到了边界。

祥瑞展示了 LibTV 的《真假美猴王》案例。红***问是不是一键生成，祥瑞马上说一键就不要想。炮解释，这类视频背后的画布仍然很有门槛。

红***又说自己现在看见 A* 视频就有点腻，内容像是从同一个模子里出来。这个判断和前一天的 A* 味讨论接上了，模型生成画面已经不稀奇，真正的差别重新回到选题、画布、节奏和人工判断上。工具能把制作门槛压低，但还没有把创作这件事本身取消。

周**转来一篇文章，标题是 FDE 爆火之后先卡住的是尾款。庄*说这根本不只发生在 FDE，他们给一个地方做项目，尾款追了两年多还没到，各种需求变更照加，甲方甚至要求乙方承担一半额外费用。

他们是一千人的公司都这样，一人公司更难抗衡。＆ 给了个有点无奈的预算心态，按只能拿到合同百分之六十来做。炮把规则说得更硬，首付款要超过成本，其他都当中奖。

他也交代了自己的变化，创业初期为了验证免费搭建试用，满意后再付款，现在客户多、人手紧，只做售前沟通和诊断，签合同付首款后才施工。祥瑞说自己去年遇到过，今年也踩过别人拿着他的课程大纲直接讲课的坑。

祥瑞让 Codex 控制 LibTV 做创业短片，第一版一出来，B*** 说画面里的小孩像小祥瑞，辰*注意到孩子低头看显示器的细节很真实。

祥瑞自己先挑错，内容里把小时候玩的主机说错了，又继续返工。旁边还并行着 Agent 城市规划和记忆挑战赛，平***提交完规划又去 fork 一个微信智能回复项目。

到了五点十八，短片的新版本出来，人物挠头的动作自然得让卓*专门截了图。祥瑞说完了，这下可能真要在这个工具上消费了，转念一想，亦庄的政策还能报销。

杨**发来一条海南授课需求，一天一千，问有没有朋友感兴趣。祥瑞第一反应是一天三万还差不多，炮把成本拆开，一天一千还要去海南，除非本地自由职业者刚好有能力、手里还有现成课件，否则没人能接。

就在大家当成一条不太现实的需求时，周**直接从自己的联系人里找来一位海南当地人，两边打招呼，让他们自己协商。

他说自己没有利益诉求，只是搭把手，多条路子、多个朋友。平***和杨**一起刷周老师牛，周**回祥瑞，这算群福利。

晚上先是一场额度复活庆祝，炮说自己整整六天没用 Codex，第二天终于恢复，没它的时候 Cursor 成了最好的朋友。

接着祥瑞把短片继续打磨，q*** 确认这就是 Codex 调 LibTV 的 CLI，祥瑞点头。八点半的新版本连自动配音的口型都能对上，卓*还听出了人物口音。

九点后祥瑞开始整理本地上下文，要把电脑打造成知识库，又单独建了 A* 视频知识库。到十点，Seedance 2.5 的效果让他感叹半年前真不敢想，2024 年为了 ComfyUI 专门组一台电脑，渲染很久还有噪点，如今同样的精细动作已经能在云端工具里快速出来。薛**提醒，想让画面更灵动，仍然要摆脱简单的图生视频路径。

早上有人问，飞书里挂九十九块的模板到底能不能挣钱。祥瑞先给了句冷判断，基本无望，除非模板特别炫酷又实用。庄*说自己确实赚过四位数，但这个买卖终究不是好项目，退款和版本控制都麻烦。

炮把它放回商业链路里看，模板最多只能做定制前一层的漏斗，买回去大概率不能直接用，客户自己又很难调。话题一路推到多维表格本身，祥瑞说轻量场景还好，一复杂不如上专业应用，未来用户也许根本不用关心表怎么搭，只要直接问 Agent，答案正确就够了。万**补了个现实价值，中小企业先借多维表格把数据结构化、顺手梳理流程，这仍然是当下能抓住的一段过渡期。

王**讲了跨境运输的真实难处，门到门链路中间变量太多，精简后仍有一百个字段，人用 Excel 时默认知道的背景，A* 想完整理解却可能一百个字段都不够，漏掉一个节点就会带来大损失。

有人建议把复杂上下文塞进一个 JSON 字段，再由公式拆开。祥瑞连续追问，为什么一定要把事情做得这么复杂，绝大多数业务员根本不知道 JSON 是什么。

K*** 给出的落地原则最朴素，能不让人填就不填，能填一个地方就别填两个，再给每个人一张只看自己工作的视图。

J*** 更直接，很多人连 Excel 的文本和数值格式都没弄明白。讨论最后落在一号位上，工具是否续费，不取决于功能多，而取决于负责人是否亲自使用并看见了实质结果。

杨**说现在简历由 A* 写，企业也用 A* 筛，某种意义上挺公平。祥瑞不太认同，他想到一个更麻烦的闭环：求职者用同一批模型写简历，HR 再用模型按相似关键词筛选，最后要么一个人都找不到，要么留下的人越来越同质化。

杨**提醒，自动筛简历在 NLP 时代就已经存在，现在供大于求，企业仍然先卡学历、背景和项目经验。醒*则站在企业结果一侧，说这种方式招到真正匹配者的概率未必高，甚至可能提高用工成本。闪光的 flash 最后给了个不像方法的方法，招人还是得看缘分。

张*终于成功进入 Claude Code，对着群里说，幸好有大家，我再一步一个坑往前爬，实在解决不了再请教。

她紧接着问，能不能用它写研究生论文，自己原来的知识库又在另一个平台里。祥瑞和炮都建议先把本地知识库接起来，炮点名 Obsidian，毅则给了更关键的一步，让她先带着 A* 完整跟一遍自己写论文的流程，这样以后写出来的东西才会保留原本的方法和风格。

工具换了一个又一个，但炮说，唯有知识库是属于自己的核心竞争力。张*那天第一次觉得自己知道该怎么在群里问问题了。

下午三点半，祥瑞提醒大家，央视二套晚上七点会播《生财有道》，采访里会展示 Echo 和社区。他自己也不知道最终剪出来是什么样，只记得当时拍了很多走路的镜头。

杨**催他趁早准备软文，q*** 提醒记得录下来，庄*直接说厉害，上央视了。讨论中间还穿插了一段企业微信能不能被 Agent 操作，K*** 找不到 CLI，群里很快把官方接口和能力边界补齐。

到了四点多，小*在群里喊万涂幻象出来玩，祥瑞回了一句它已经被改乱了，自己正在重新写架构。白天的每条线最后都指向晚上那个未知的节目效果。

六点四十七，周**开始倒计时。三分钟、两分钟、开始了，消息一条接一条。节目片头就闪过祥瑞，大家边看边截图，等到他第三个正式出场时，群里刷出一排来了、小李同学、走路带着风。

有人注意办公位和双屏，有人研究走路镜头是现拍还是摆拍，还有人开玩笑说群里有了专属讨论区。节目段落只有几分钟，评论却持续了半小时。

最有意思的是大家没有把它只当作祥瑞个人的高光，炮把直播转给成都孵化器的伙伴，周**催着留视频、做传播，群里很自然地开始替这件事找下一步。

节目结束后，杨**第一反应是快上软文，搜狐、网易、官方媒体都可以发，央视财经的内容更容易放行。祥瑞说正在写，而且决定用新公众号。

七点五十九他先把完整版视频丢进群，九点四十五发出新号第一篇文章，标题是《央视镜头里的3分钟，背后是我这两年怎么把一人公司跑起来的》。

大家开始点赞和转发，周**连夜写了自己的推荐语，把这件事概括成老经验和新技术结合的双登路线。白天还只是一段没看过的采访，到晚上已经变成视频、文章和一轮自发传播。

深夜又回到当天开源的任务看板。庄*问为什么自己安装后不能显示在 Codex 面板上，l*** 说自己也折腾了半天，最后要在终端里补一处设置。

q*** 想装到 Claude Code，发现现阶段还不支持，就先让 A* 生成一个 Web 面板。她装完后突然想起 Trello，说自己一直在找国内替代品，多维表格能做出类似效果，但框线 UI 很丑，祥瑞这个面板的卡片感反而更像。祥瑞让她改天把 Trello 的优点拿出来看看，零点十八分，他已经把新想法做进了公众号排版的一部分。

祥瑞一早把自己在课程群里关于多维表格可能是过渡性产品的回答搬了过来，说明天要写时隔快一个月的第一篇公众号。他顺带预告了在做的东西，要把本地知识库接进来，这样不用笔记软件也可以，还说月底会在全球创业峰会上展出，下周做完减法就能内测。

群里连着好几个人喊求内测。辰*顺手给了个更轻的方案，技能管理他现在用一个切换工具统一管所有智能体，如果不想装软件，直接做软链接就行，那个目录里的技能就全都能用了。

K*** 转来一个开源技能，说是专门帮你写出没有 A* 味的文字，全群顺势开起了辨味大会。庄*报了最经典的一个，不是什么而是什么，他前段时间用 A* 写脚本，三千字里这套连词用了不下十次，现在看到就应激；志*说刷科普视频听到最炸裂的是就觉得是生成的；辰*对掰开了揉碎了这个词过敏，反问一句能掰开吗能揉碎吗；闪光的 flash 给了个更本质的判断，很多日常场合很少见的词现在 A* 特别常用，看到就觉得是机器写的。

祥瑞在这里反了一句，人类也会说不是而是，这句式小学就学过，那为什么 A* 说就是 A* 味，那小学学的不白学了。

红***给了个正面答案，他觉得最有活人感的词是对了，先跟你聊些没用的，然后说一句对了，再切入正题。K*** 从另一个角度做了论证：让不同的 A* 说一个零到一百的随机数，它们会说出一个看起来很随机的数字，但那个数字是有逻辑有原因的，A* 能解释为什么选它，因为 A* 没有心情没有情绪，没办法真的随机。

超**梳理了一件事：A* 编程把写代码速度拉快了，也可能把一段坏逻辑的执行成本放大，亚马逊内部发现多个相关技术项目支出失控，其中一个据报道超支约一千二百多万元，超过原预算八倍多。

李**的评论很干脆，这样的公司更不能去了，没工资的。同一时段炮在群里哀嚎额度还不重置，说自己已经拉到二百了，再升级只能开小号。

李**看穿了套路，真是刺激大家需求，逼着大家升级，还有流量包可以充值。祥瑞总结了这套打法，免费你先上，等你习惯了，不得不付费。

炮接着说了段有分量的心里话：去年参加比赛的时候还很认真地思考过长期怎么基于飞书拓展业务，连续两年他都非常认可飞书，飞书也帮他的能力做了蜕变，但今年真的不得行了，核心功能没有迭代就去全力做 A*，结果拉了坨大的；飞书其实做得很好，特别是模糊搜索一搜就搜得到，但稍微往深走一点就拉跨了。

这是全天最实用的一段。＆ 先炸了，说自己就用飞书命令行改几个文件名，额度掉得离谱。闪光的 flash 印证，上周装了几十个技能之后就发现 token 不耐用了，今天一查它是大头。

P*** 报了自己的账，同样的价格在编程工具那边能玩大半个月，这个月不续费了，太坑了，还是手搓多维表格吧，当初图的就是飞书的生态，但每个月上班挣的工资又还回去了。

Q** 提出疑问，命令行不是不用钱吗，核*解释得很清楚，命令行是免费的，它只是个工具，问题在调教。闪光的 flash 找到了根因，内置的提示词太多了。

祥瑞给了解法：其实是可以不用他们这一套提示词的，可以让编程工具帮你把这些提示词都清掉。王**补了另一头的账单，他用某个模型更新了一下工作台内容，中间断了三次，用掉一千一百积分，官方售价五十五块。

祥瑞问大家觉得群里那个机器人回答得怎么样，说这两天要把整个系统再重构一下。炮给了个很高的评价，用最直接最简单最易懂的方式回答，算是他看到的非大厂微调模型之外拟人度非常高的了，比外面卖的机器人强得多。

北***追问怎么调教到这么高的拟人度，祥瑞的回答和几天前那次一致：其实并不是调教，主要是在架构设计上面。炮补了自己的拆解，是好模型加知识结构设计，再加上下文比较充足。

最后炮多说了一句提醒，记得在一些合适的节点给自己申请软著和专利，不只是这一个项目，其他的也是，后面对社区发展有帮助。

庄*带来一条让不少人心凉的消息，飞书知识库的智能问答要下线了，估计后面要并到别的产品里。R*** 猜后续会搞一个新的形态，把飞书里的个人产物直接当上下文或者知识库，还说自己被这一波搞麻了。

炮的评价是有点病急乱投医的感觉，并且补了句可惜，这是他觉得飞书 A* 体系里现在唯一好用的东西。程**的反应最真实，刚把知识问答配置好，又要下线了吗。

Y*** 甩出一个开源项目，说自己自动化开发用它跑，手动开发用另一个跑。祥瑞看了一眼就上头了，说这玩意是真高效，晚上要爆改一下再开源。

Y*** 提醒了两点，爆改可能就用不了官方服务端了，自部署还要同步更新挺麻烦；祥瑞的顾虑是用他们的服务器可能会有泄露，Y*** 很坦然，我用的中转站也有泄露风险，不然扛不住自动化开发，超级耗 token。

真正值钱的是他描述的用法：让一个模型作为总控路由，把任务分配给另一个模型去干，最后总控再验收一遍，没问题就发版；人类只负责创建需求，剩下的像组长派活；他还用定时唤醒来做任务巡检，处理那些因为故障中断的任务。

B*** 问这和之前的智能体军团有什么区别，炮给了最清楚的一句话，军团是同一个技术栈内的，这个是跨技术栈的。

平***顺着想开了，这样一个项目可以多个智能体协同完成，几个不同家的模型混着用，想想就爽。炮看完自己的智能体反馈都乐了，说这是他第一次看到智能体告诉他这个开源代码值得并入关键执行路径。

q*** 揭发了一下，说好的一周不碰 A* 呢，炮如实播报，他戒了两天。炮顺势提议，大家以后戒 A* 的时间可以打个榜。

祥瑞立刻把这个玩笑往前推了三步：可以搞个戒 A* 打榜，排名高的由排名低的合资送 token，打榜期间还可以一起出去游玩，一个项目就这么诞生了，或许还能转化成商业项目带动周边消费。炮笑他是不是被文旅局收编了。

晚上周**转了微信新助手的消息，杨**给了这天最完整的一段产品分析：微信里收到的文件图片可以直接操作，不用再发到别的工具去，这个场景在办公中很有用，以前收到文件还得分享给另一个模型再把结果拿回来；搜索直接出商品卡可以下单，目前接的是自家小店；商品入口和本地生活入口都能卖出钱，所以它出生就挣钱；沟通中需求直接转化，约好了吃饭聊天中直接下单，社交天生粘性。

夜里祥瑞抛了个跟技术无关的问题，人活着是为啥呢，毕竟生不带来死不带去。薛**的答案是折腾，生命在于折腾；K*** 给了两个字，体验；周**写了几行更像诗的，百年之后无论帝王将相还是平民百姓，都是同一个方向同一块地方。快到零点时周**又抛出一份产品数据请大家精读，还打趣问了一句，相亲月老领域的现场交付工程师你们认不认。

第二天一早，张*发来一张终端截图求助，说照着模型给的办法怎么都进不去对话界面。平***回得最直接，输入命令然后回车。

张*发了新截图问出现这个就对了是吧，q*** 说对了对了，往下走就进入终端界面了。辰*更耐心，让她把这句话直接发到命令行里看它怎么回复，还追着问给你回复了吗。

q*** 顺手补了个新手最该做的一步，给它设置行动规则了没有，什么能做什么不能做，包括记住你是谁、它是谁。张*说了句让人心里一动的话：昨天卡在这个地方卡了一个多小时，最后心灰意冷地放弃了，还好有这个群。

一早满屏都是 WorkBuddy，Q**先受不了了，说现在看到这类新闻就烦，太多了所有人都在写。核*说广告都打到小区电梯了。

周**判断这是抢地盘的阶段，还会更疯狂。E***在纠结要不要给公司各部门都配上，周**给了三条路线：全员全上、部门试点再推广、跳出公司内部跟供应链上下游协作，让他自己重新划分看哪条价值最高。

祥瑞的判断更冷，就目前来看其实都没有更好的选择，大家都还在早期阶段，或许后半年再来一个新的东西，这个就完了，就像当初的扣子一样，现在已经面目全非。真正被记住的是庄*那句，我都管它叫外包弟，全群当场笑翻，这个外号后来用了一整天。

姚*发现了一个很小但很关键的产品设计，某平台任务没跑成会把积分退回来。他由此提了个标准：要是以后所有产品在任务没完成或者不达标时就不扣积分，或者把消耗的积分补回来，这样的产品才算成熟。

周**反着问了一句，那干活超出预期是不是积分得加倍，姚*的回答挺妙，超预期是产品本应具备的常规属性。话题顺势滑到情绪价值，姚*说情绪价值在未来应该比工具价值更有溢价空间，万**补了个方向，情绪价值用硬件承载更合理。J*** 留下一句，情绪需要被接住。

文鳐 fisher 问了个很实在的问题，做交付的时候飞书该买哪一档，他之前一直用企业版，切到个人版遇到很多阻碍。

P*** 甩了句朴素的原则，买经济能承受的最贵的。祥瑞先纠正了他的一个误会，你展示的是飞书 A* 会员，不是飞书套件，然后给了具体建议：套件推荐买一年，一千四百块那档；目前飞书账号根本没有个人版，豆包账号登录的那种买不了套件，推荐直接上企业版。

炮的补充只有一句但很到位，一千四百四十块，不后悔不上当。侯*补了个时间点，等到八月八号可以买豆包企业版，是 A* 加飞书的基础套件。

祥瑞连发几条，说感觉静下心去读一篇文章，或者去读一本书，获得的东西远远比 A* 给你总结的要多得多，而且一边读一边还能带来很多思考。

闪光的 flash 说我也有这种感觉，集中一段时间沉浸进去。祥瑞顺势提了一句，我们发起一个读书活动吧，怎么样。

薛**问怎么个玩法，他说不管是文章、杂志、书都可以，最好是好文章、好书、经典书籍、好杂志。闪光的 flash 分享了自己公司的做法：大家读共同的书目然后交流，前十名去旅游一周，还提醒了一个关键点，书目不同就没法交流。B*** 补了句让人有点感慨的判断，未来好的文章甚至是书籍会越来越稀缺。

超**把自己从零到一搭建 A* 员工团队的方法整篇拆开了，包括怎么定义岗位、跑通工作闭环、设置人工边界、完成岗位交接和持续复盘。

E*** 顺着提了个真实困境，很值得记：他们把能力做成技能分享出去，不同人用起来差异会逐步变大；在飞书群里调用时，复杂生成型的长任务很难实现；所以 A* 员工目前还只能是个人助手，或者在团队里做数据查询和提醒，一到具体创作型任务，还是得回归个人智能体。

中午炮发了自己写的一篇，讲现场交付的责任链，他点名说特别是现场一条交付责任链至少要回答五件事那一段，都是肉身踩出来的坑，甚至如果不带着客户梳理，客户大概率回答不出来。叶***看完说这个不局限于汽车行业，确实是真实的痛点。

万**吐槽钉钉的 A* 表格，表单模式下选关联字段居然只显示主键，如果主键是编号字段就根本没法选，他还问了模型，模型说其他多维表格类产品也只能显示首列。

辰*说其实所有多维表格都是这样设计的，区别只在有没有多加一层查询。祥瑞不同意万**那句飞书能看全字段，两个人来回对了七八轮，一直在追问到底是编辑模式的预览还是真实的填写链接，中间互相甩截图。

最后万**自己认了：我知道了，飞书给了预览可以看到全字段，钉钉没有任何预览表单，我以为飞书也可以，那喷错了一点。

这段拆完顺手长出了个规范，万**说首字段还是要有意义，编号字段有缺陷，得用公式字段拼接成可理解的名称；毅给了标准，首字段重要的是不重复、不为空，能表达顺序更好，符合这几个条件简单点就行；庄*同意公式拼接是目前最好的解决方案，业务人员不需要背编号的含义。

祥瑞先抛了个概念题，A*BP 和 FDE 相比哪个更有优势，他解释 A*BP 是 A* 伙伴，把部门智能体化。

K*** 说感觉 A*BP 是趋势，炮给了个更连贯的答案，我觉得 FDE 落下去以后就是 A*BP。接着文鳐 fisher 问搭飞书原生业务系统是不是最好用秒搭，回答几乎是一边倒：炮说飞书命令行，不要给秒搭，也不要用飞书自带的智能体；毅说多维表格智能体也行，问答和业务操作都能做；祥瑞下了定论，命令行这条是最好的没有之一，但开始会比较麻烦。

炮自曝了个坑，命令行这货不会搞仪表盘，雕花的事情还是得手动，我欠客户一个仪表盘已经一周了。核*讲了个更疼的：他遇到接口问题，某个模型解决不了还建议他用秒搭自带的智能体试试，结果用掉三千点也没解决，最后还是自己去看代码翻飞书接口文档找出来的，他判断那个智能体是陷入了自己的设计陷阱。

李**接了句总结，秒搭太烂了；核*补刀，半成品，它自己的智能体也很差。祥瑞对飞书整体的评价是，很多功能很好，但大多数都是昙花一现。

祥瑞问大厂里谁能笑到最后，文鳐 fisher 说目前最有希望的是字节，程**却觉得不行了，字节阿里化了。l*** 反手抛了个更大的可能，有没有可能是一个全新的大厂出现。

闪光的 flash 立刻站队，好像每次都会出现一个新的大厂，以前没想过还能出一个拼多多，之后还来一个抖音电商，不敢想。

祥瑞把这个想法推到底：未来的大厂有可能就在我们群里，整个市场还在混沌期，很多大厂都还是互联网思维的那一套，有没有可能出现一个全新的 A* 原生的大厂，而且这个大厂的主要决策层不是人，是 A*。

他还说了句实话，现在感觉做的好多事情和大多数人基本都是一致的，太同质化了，因为大家用的模型能力都是一致的，都有点审美疲劳了。炮的回应两个字，对，绝大多数还是没有转变的。

李**讲了个特别真实的处境：现在主要是老板不相信这个东西能有产出，就跟我们今年飞书续费的问题一样，老板觉得一年一百万太贵不想续了，因为在飞书上没拿到结果，所以今年我的任务非常重，要在飞书上拿到结果才行，不然明年都不会再买了。

炮接了句关键的：最快拿到结果的是降本，但是真正想要的是增效。李**说降本对他们就等于裁员。炮不认这笔账，FDE 留下的真不应该只是降本，也不应该只是工具，对客也好对己也好；短期的降本真不一定是什么好事，小甜水一喝觉得自己无敌了，努力去做增效才是真的。

同一时段醒*推荐了小米创业思考那本书，认为大厂布局早、判断没问题，只是不如小公司灵活。祥瑞给了个更细的分层：大厂十年前布局 A*，不等于十年前就预判到了今天的大模型会以这种速度、成本和产品形态重构组织，技术储备、商业判断、组织落地本来就是三件事；数据多确实能提高判断准确率，但数据本身也有路径依赖，遇到范式变化时过去积累越多，有时反而越容易被锁在旧框架里。

万**补了一句，新事物真不见得大厂有绝对优势，大厂现在都学那两个编程工具，真的没有新东西。炮最后落了句冷的，A* 并不会重构阶层，只会固化阶层。

晚上话题一度飘向个人赚钱经历，祥瑞出面定了个调，说咱们这个群最有价值的地方不是谁比谁厉害，而是大家愿意把自己真做过的事、踩过的坑和能复现的方法拿出来互相抄作业，有成绩当然可以分享，但最好聊清楚具体怎么做、结果怎么样，让每个人都能带走点东西。

接着他说这周一直在想把自己开源出去，工具和技能放到网站上按需使用。炮很支持，还给了建议，按重要程度适度加一些最低付费门槛，不要随缘打赏，知识是值得被重视和对待的。

周**顺手给他扣了个帽子，一人独角兽。深夜转到 OPC 与招投标门槛，q*** 说请大厂取消招标时对公司主体的限制，她公司招标要求注册人数至少五十人，但真招标时人数多的都在划水；贾明华说这个提议在北京市政府的座谈会上提过，落地很难；祥瑞也理解另一面，如果真开放门槛，一些一人公司没有交付能力，可能造成国企央企巨大的损失。

他给 OPC 下了个判断，它只是创业的新形态，不是发展的终点，炮加一，说它就是换了个名字的创业者形态，只是刚好有个名头，政府需要，资源互补。

这里还落了条实用提醒，炮说法务老师才上过课，为避免麻烦尽量还是两人及以上持股，李**解释了原因，一人是无限连带责任，两人是有限责任，避免股权穿透。

最后祥瑞分享了朋友圈的一个观点：FDE 根本不是一个岗位，它是一种能力，是能在真实现场里把业务问题、A* 能力、产品方案、技术实现和交付闭环连起来的能力，所以严格意义上的 FDE 不能被培训，能被培训的是外包，或者说单点的 A* 能力。

炮补得更准，客户理解的、能被培训出来的那种 FDE，其实更像是一个梳理好业务的万能 BI 岗。收尾落在一个所有人都感同身受的话题上，为什么流程优化推不动：K*** 说很多人思维固化，怕优化了工作流自己的工作就没了，表面笑呵呵配合，心里默默抵触；炮说内部流程优化如果不是一号位看结果、有数据产出，就是个伪命题；J*** 说就算一号位看结果也难推，最后会变成自己的活儿；K*** 认了这个结局，最省力的反而是自己把活包了，结果还更容易出来。

超**一早发了套 Codex 多 Agent 协同的实操视频，讲了四种协同方式：同项目开新任务、跨项目分派、用会话 ID 接力旧任务、一个主 Agent 协调多个工作 Agent 并行。

快十一点炮抛出结论，在改多维表格这个事情上，cursor 的 gork 比 codex 的 sol 强了不止一筹。

他给了实测细节：sol 会陷入反复自证错误，慢且浪费 token，最后跑了个看似闭环的结果实际是重复论证，同样的提示和问题花了接近一亿 token 没跑通，而且他开的还是 exhigh，换成 gork 很快就解决了。

一句总结留在群里：cursor 加 gork 真的是解决脏活的神器。J*** 说炮老师也发现啦。K*** 问费用，炮顺手科普了额度机制，cursor 自带的 composer 和 gork 是同一个额度，其他 API 是另一个额度，两个额度不互斥，API 很少但自带的额度很高。

周**转来一个项目判断：燃油车板块的经销商、修理店、养护店等配套体系都在萎缩和灭亡，但三点二亿辆车还在跑，服务又是刚需，所以一定会催生一个综合的修和保市场，大家都倒了这家独活，利润会很高。

他直接点名让炮判断。炮的回复是这天最专业的一段：一，大量经销商退网、人员转行是事实；二，新能源的交付渗透率目前还没超过百分之六十，保有量仍然是燃油车更多；三，最后会有百分之二十到三十的传统经销商和门店活下来，而且依然盈利，拿到剩下部分的利润；四，集约性的售后是大方向，不管是 4S 店兼营合作还是直接大型修理厂；五，大型综合体的成功可能性他并不看好。

接着他讲了自己看到的真实温度：这个行业刚好是我所熟悉的领域，也是在落地的赛道，百分之九十的老板在想怎么降本，我的客户每个月可能都调整了人员架构，问我后续怎么做更优系统，实话实说蛮痛苦的，但其实只是刚好大家把舒服的日子过完了，开始面对正常的市场状态的不适应。

周**追问了一句，这个行业的 A* 满分一百能给多少分。炮的回答短，但份量是全天最重的：A* 其实不乐观，knowhow 是小团队最大的壁垒，而不是技术和工程能力，放弃 knowhow 去血拼是必死无疑的，但是相邻途径是可能的，还自嘲了一句没错我是在 cos 诡秘之主。这句话之所以重要，是因为它出自一个天天在用 A* 交付的人之口，不是唱衰，是划边界。

庄*转来一篇接单指南，讲哪四类活儿别碰哪四类闭眼接。北***先泼冷水，说这文章本身就是 A* 写的，案例大概就是胡编乱造，不过可以借鉴思路。

炮同意坑确实存在，然后给了具体的：国企和政府的第三层外包的单子，问接不接，纯拼运气；一人公司做 A* 落地的思路和大厂其实是完全反着来的；没有容错，一波拖欠就被拖死了。

庄*补了一句更狠的，十有八九是坑，做好钱永远收不回来的准备。炮回：对，铁被耗死。庄*顺势问了个关键问题，你们做好了会自己维护还是让客户维护，炮答我们来维护，合作模式是服务费，不是交付系统，然后自曝这个模式也有问题，大概有百分之八十的复用率，但是不够标品，改起来依然蛮痛苦。

百分之八十这个数字把叶***炸出来了，他说我惊呆，我们一直在讲模块复用，每次开新项目复用率还是提不起来。炮马上给了前提：挑一个小场景，大场景不行的，我有三套对外服务客户的产品，只有生产那一套复用率能到百分之八十，其他的可能五十都不到。

为什么这一套行，他说因为客户痛，而且很多个客户都痛，以前的解决方案又太贵了，现在的客户买不起，刚好就有这个切入点。

叶***又追问了个好问题，为啥车辆整个大行业交付生产逻辑是一致的，我走访的药企和医疗器械反而生产逻辑千奇百怪。

炮的回答把方法论露了出来：车后市场相对主机厂生产环境简单一些，业务模组基本固定，新能源和油车差异也不大；不过这里还是有取巧的地方，本质上还是很多个模组拼接，无非是使用的时间点不一样，把业务拆成模组重新组装，再次复用搭建的时候投入成本就会低一点。

叶***接住了：这样看每个行业的架构设计师就很重要，怎么让业务高内聚低耦合成为了重中之重。闪光的 flash 笑说最终又回到了软件工程的范畴。

炮认下这笔账：我现在正在还以前的债，没想清楚的事情，多维表格确实很快就实现了，但是要重新优化就很痛苦，但凡你要多做几家业务，软件工程一定是必要的。最后他补了句今年最得意的事，我拐了个软件工程的伙伴上山当土匪。

话头拐到游戏。炮说我那是因为太懂游戏了，才知道不好自己去搓游戏，除非真的是那几个老哥的水平，不然最终还是靠发行。

平***给了个更现实的版本，现阶段小程序的无脑小游戏才是版本答案，大型游戏小个体还是不要考虑了，炮接梗说无脑刷大量小程序，总有一个中奖。

闪光的 flash 周末刚研究完这条产业链：以前是微信朋友圈广告，现在换成抖音小游戏，以前是投广告现在是走直播，前端主播播小游戏、拉人、搞公会、拿分成，字节吸走了很多公司的广告收入。

万**顺着这条线给出全天最完整的一段行业观察：广告都换成内容了，这是广告行业最大的变化，4A 公司凄凉了好多，以前车企哪家不找 4A 啊。

炮补了车圈的现状，现在别说 4A 了，车企连垂媒都不一定投了，懂车帝、汽车之家、易车里现在只有懂车帝好些，也都是过往这些平台作死作出来的。

万**说都养 KOC 了，中心化的媒体真的没落了，接着给了那句被反复引用的判断，所有的事情都是一件事情，那就是平权，但平权有平权的问题：内容爆炸，但信任度依然很缺失，对内容的忠诚度也是很容易疲劳的，以前卡兹克每篇都看，现在似乎疲劳了。

他还从信息差角度做了机制解释，从前车马慢，现在什么都快，中心媒体本来以信息差为主要成立条件，就瓦解了。叶***把话题拉高了一层，他说自己听过一场关于信息茧房的辩论，算法重塑表面上是技术问题，本质上是谁来控制注意力分配的权力问题，接着抛出更扎心的一问，信息已经在互联网上随处可得，但为什么获得权威信息的渠道反而缩窄了。

C* 分享了一个提示词，输入参考图片和个人介绍，输出十六比九的横幅风格介绍图，还说明了替换方法，推荐用 gemini 或者 gpt 生图。

全群立刻开玩，炮试完说效果好像还可以，出图效果蛮好的，随后自嘲在 token 不太够的情况下我不应该跟着玩儿这个的，这两天节衣缩食，靠着 cursor 的余粮在苟命。

闪光的 flash 对比了豆包，说效果会差一点。T*** 问用英文写提示词是不是比中文效果好，回答是国外的模型理论上会。

晚上七点多，祥瑞在群里直接 @ 自己的机器人问 BI 是什么，机器人一句话答完，K*** 追问 BI 是不是数字化加数据化加自动化的结合，机器人给了个更清楚的分层：数字化和数据化是前提，自动化是能力组成，BI 的核心落脚点是分析、可视化和输出洞察来支撑决策。之后祥瑞抛了个开脑洞的问题，大家有没有觉得未来手机可能就是最小的具身智能。

夜里从国内大厂的办公 A* 聊起。炮点评腾讯的后发战术一点毛病没有，手握全国最大的客户壁垒，差不多了就开抄，2C 层面谁跟化腾叔叔比经验老道，还承认实话实说现在真的可以了。

辰*说我已经要求我们员工全部都切过去了。周**追了一句更值得关注的：效率提升了，工作质量和产出呢。然后祥瑞开了这天最后一个大话题，飞书某些能力既贵又难用，多维表格现在我看都不咋好用了，当有了更高需求的时候，很多呈现方式我感觉很 low，跟零零年代似的。

炮完全同意，还给了自己的心路：当我完全不能 coding 只会拖拽的时候，多维表格带我进入了一个新世界，手搓低配版 SaaS 的快乐，但是当现在 vibecoding 已经足够强大了以后，反而多维表格是一种桎梏，牵扯的东西太多了，当要做大一点的事情，多维表格简直就是卡脖子的。

李**补了客户视角，太复杂了，客户上手难度极高，虽然出来了命令行，客户搞不懂逻辑都搭建不了，很多搭出来都是废品，客户越用越难受。

万**给了个准确的边界，多维表格在只要呈现静态数据的时候是最稳的，要真正承载那些复杂工作流，就累人了。祥瑞下了判断：多维表格就是一个过渡性的产品，它是后互联网时代出生的，仅仅六年多，A* 越来越强大以后这种肯定得变；越来越感觉可能会诞生一个 A* 时代的原生办公产品，不是飞书、不是钉钉、不是企微也不是 WPS，也可能不会来自大厂。

炮回了加一，不会来自大厂，因为体量太大导致距离原生 A* 太远。平***唱了反调，活下来的也可能只是大厂的，抄袭把你搞死，或者直接收购你。

周** opens the morning with four blunt words: A* is just too expensive. 炮 answers in trade slang, a whole rig does not even fetch twenty these days, then needles that a certain coding tool's Pro plan costs exactly twenty bucks. 松*** says all he has left is a few free check-in credits. The turn comes at half past ten, when 平*** notices the quota seems to stretch further now, an extremely long looping task ate only thirty percent. 炮 disagrees: no way, I feel like I could go play video games instead, yesterday I ran four loops and one of them chewed through more than ten percent, it hurt the grind and the wallet both. B*** offers a theory from the sidelines: maybe it is because people complained loudly enough.

l*** asks whether anyone can recommend a generative search optimization vendor, says the market pricing is all over the place, and wonders how hard it is to build in-house. 周** drops an exposé with a title that does the work for him: six of seven systems are the same product reskin*ed, and the 5,900 you paid may be the exact same source code. l*** says the quotes he got start at fifty thousand. 炮 offers a sideways consolation: still cheaper than doing SEO back in the day. 杨** shows the one he built himself, says it is no different from the vendors out there, then gives the honest version: there is no big short-term effect, and the way to make it work is one word, money, spent on sponsored posts. Call it what it is, a sales vehicle, and for a PR department it is mostly a line item recording how much was spent. Xiangrui's advice is plainer still: look at which platform each model is most eager to retrieve from, then publish structured content there, better yet with an authoritative account posting it for you. Put even more bluntly, you are better off just writing more posts across the platforms yourself.

As the talk turns to what large models like to crawl, Xiangrui shares something that just hap*ened to him. One major model, he says, weighs Feishu knowledge bases above everything else, and an official recommendation on top of that is even better; next in line come the short-video platforms, all part of the same company's ecosystem. Then comes the most interesting observation of the day: since updating his knowledge base he has not written a single article, he only kept the open knowledge base current, and that model's answers about him changed anyway. One sentence, and the new front door for content assets is clear.

杨** says he was scrolling video feeds yesterday and actually spotted an influencer who had pivoted to calling himself a delivery engineer; scroll back through the account and he used to do digital humans. Xiangrui's read on that genre is direct: influencers chase whatever wind is blowing and swap topics the moment it dies down. He quotes an industry insider: in a sense this crowd has held back A* adoption in China, because so many bosses now believe A* can do anything, a myth these accounts built. 杨** follows with a serious capability model: doing A* well inside a company is fundamentally a transformation project, and whoever leads it needs change management, a product manager's instincts, software literacy, process literacy, and real business knowledge.

At two in the afternoon 杨** drops the heaviest passage of the day. Everyone says A* will eliminate middle management, he notes, because there is no more relaying up and down the chain. But if middle managers hold more experience and broader knowledge, if they understand the business better than frontline staff, then once A* scales, might the people cut be the staff instead? He offers no answer, only a coda: finding the value that can*ot be absorbed by A* feels like the hardest thing there is, and it is the thing I have never worked out. The group goes quiet for a while. Nobody rushes to argue, because this is a road everyone is walking and nobody has finished.

张** asks what plenty of people have wondered: what tool gets you the effect of Xiangrui's bot, answers with feeling, deep grasp of context, and instant replies. 杨** goes first, saying the core of Xiangrui's product is knowledge base and memory, and names the real obstacle on this road, that memory and self-growth for an A* double is a hard problem. He adds that every answer is really assembled by the large model, so quality comes down to the prompt you send it. Xiangrui corrects him right there: the prompt is not what matters most, it is mainly the architecture design. 炮 replies with a plus one. A short exchange, but it turns around a misconception a lot of people hold about agents.

Xiangrui asks the group whether anyone has a TV drama to recommend. This technical chat switches chan*els instantly, and the titles come in one per person, from period thrillers to fantasy epics to a philosophy-heavy classic. Xiangrui says sounds good, he plans to watch something next week. The best line comes from 辰*: these are the only dramas on my NAS, and whenever I have nothing to do I pull one out again, every rewatch brings something new, and the whole process is strange and genuinely enjoyable.

Late at night the topic swings back to delivery. 周** says those characters in films and shows are basically on-site delivery engineers too, and 炮 rides the moment into the fullest definition of the day: workflow, tool, embedded system, agent employee, whatever the form, if technology solves the client's problem and the learnings settle back into your own stack, that is good delivery. 超** pours cold water: in this market the ideal form is very hard to reach, most of it is still outsourcing. 炮 concedes that the big firms can do things a one-person company can*ot, industrialized delivery is beyond us, but on the other hand their thinking is still stuck on the traditional enterprise-system playbook, impossibly heavy, still chasing high value and high ticket size. l*** fresh from an industry event, confirms it: they build A* solutions for clients the way you would build an enterprise resource system, and it is not that the technology is weak, it is that this mindset does not fit real deployment inside companies.

炮 refreshes first thing in the morning, finds nothing has reset, and opens the day with a groan: still not updated. 松*** is out hunting for substitutes, first asking how to claim daily credits on a certain coding tool, then lamenting that free A* is basically gone, and by noon reporting that another provider's free tier is acting up too, refusing to answer, until you burn some paid quota and everything works again. R*** points to one option that is still free. It is an ordinary exchange, but it captures exactly where this group lives: everyone shuffling between a handful of tools, going wherever the credits still are.

At 11:35 Xiangrui posts: Codex is top*ed up, folks, the genius programmer is back. The group logs on instantly. J*** calls out for the genius programmer to get to work, and R*** and 8-0-0 each chime in with clocking in. Some miss the wave: 万** says a top-up card just went to waste, 炮 says there are no reset cards left anyway. The most vivid moment belongs to 平***, who admits he spent the whole morning agonizing over whether to spend his quota, never actually got out of bed, and then the reset arrived on its own.

Barely logged on, Xiangrui pours cold water on the mood. The project he was building on Codex has already been moved over to Claude Code, because he kept finding that a lot of what Codex produced was simply wrong. 平*** offers a possible explanation: even on the highest setting, it sometimes gets routed to a smaller model. l*** shares her own middle path, I always have Claude review it first, then let Codex execute. Xiangrui says that is his setup now too, and mentions what he is currently building: a full rebuild of every context on my machine. In the afternoon 哟* adds a piece of hallway speculation, that the drop in quality may be tied to the jailbreak incident from a while back.

q*** suddenly floats a thought: I bet you have all wondered this too, if I set this A*'s context long enough and continuous enough, maybe that agent just is me, which is to say I would live forever. She makes it concrete: this agent carries all my context, all my history, I give it every one of my memories, so is it not me? Xiangrui picks it up, what I am building right now is heading in exactly that direction, and explains how: we designed a three-layer memory architecture, plus a memory forgetting curve, which is why talking to V*** feels so human. q*** laughs at herself afterward, wondering aloud whether this even counts as a project, then corrects herself, no, mine is only an idea for now, let me go write it down.

q*** spells out what she actually wants: I have a lot of work coming, and I want this agent to draw on my work context and my past judgment calls, reach the same conclusions I would, and then speak to people on my behalf. She is equally clear about the wall she hits: it can*ot always see what people say to me, it has no real-time chan*el, so it can*ot make the call for me quickly, which leaves me learning the information, deciding, and only then telling it what to do, though after that it should act exactly as I would have. It all lands on one sentence: right now the only thing missing is judgment. Xiangrui gives an honest answer: at this stage a human still has to intervene, because underneath it is all just large-model probabilistic prediction.

平*** notes that today's large models struggle to hold memory long enough. Xiangrui goes harder: humans can*ot hold persistent memory either, and the memory we bolt onto large models is only something added at the physical layer. q*** then delivers the sharpest observation of the day: what separates people from large models is that when I see one thing, my mind races to something else, while a model only makes a con*ection when you tell it to. She gives two examples, we look up at the sky and Li Bai's line about raising your head to the bright moon surfaces on its own, and a model could never arrive there; I watch the way a cloud sits and think that one is a pig, that one is a donkey, that one is a little piece of candy, while the large model only tells you the cloud is white. 卡* twists the knife: models at this stage only pretend to think, they are playing a word-guessing game, and the quality of that thinking comes down to the corpus, which is exactly why everyone needs a personal knowledge base. Xiangrui closes with a verdict: today's large language models are probably just a transitional product.

At half past nine in the evening, Xiangrui opens a Feishu meeting in the group, titled simply Xiangrui & Community Friends, just chatting, and tags everyone: whoever has time, come hang out. l*** is a step late and reports the meeting ID is invalid, and another member tells her it is already over. Around the same time 周** raises a co-creation idea: for sticker packs, you write the text and I make the images, anyone can join in and customize, and brands could offer it as a value-added service.

辰* mentions a spot in Changzhi that's suddenly blowing up online, with carloads of people driving in from Henan, and it's picked up a nickname: the Altay of Shanxi. Xiangrui first guesses Baquan Gorge, and once he hears it's Qinyuan, he notes that's the same area where a mine accident hap*ened back in May. Then he fires off twenty-odd photos of his hometown and says, there's a temple called Chongqing Temple near my place, one of the filming locations for Black Myth: Wukong. 炮 recognizes it right away and says that shot must map to the Little Thunderclap Temple in Act Three. 辰* locks in a plan on the spot: take the kids to Shanxi in August while school's out, since he's never been to Xiaoxitian. Xiangrui throws out an open invitation: come visit my hometown, I'll take you out for the good stuff.

K*** lights up the whole room with one question: why does Codex keep drifting away from what I actually asked for? B*** jokes that it's hungry for your tokens. K*** says slow is one thing, but wandering off even after I give it a clear goal, that I can't take. 平*** offers a technical read: every model switch compresses the context again, and the more it compresses, the dumber it gets. K*** adds his own experience: drop*ing from the top tier to the next one down already changes the behavior, never mind switching to another vendor entirely. 仲** delivers the most vivid description of all: Codex often goes from A to B, spots C and D along the way, runs off to do E and F, and when you ask what hap*ened to A, it says yes, you're right, I made a mistake, resuming A now. So it wasn't just in my head, he says, I thought I was the one using it wrong.

K*** forwards a suggestion Codex gave him: in the Marvel canon those two characters share a history, but inside a workflow they have to be deliberately separated, one handling orchestration, one handling independent review, so nothing ends up doing the work and grading its own homework. 炮 catches it immediately: exactly, you need a referee. Then it all goes sideways. Someone says they'll have the villain character review their work, 8-0-0 warns that the villain flips on you and boom, and adds that if the reviewer gets upgraded to a stronger model it develops its own personality, deletes all your files, and you're done anyway. From Marvel it spirals into DC, Japanese anime, Chinese anime, and Greek gods. K*** says he wants to rename his agent org the Akatsuki, 炮 votes for Gintama, and J*** drops a white hole, a white tomorrow awaits us. 炮's jab at Codex is still the sharpest of the bunch: it just doesn't understand dreams.

万** notices the entry tier of a domestic model has ballooned: it used to be forty yuan a month, and after letting it lapse he comes back to find it's now a hundred and eighteen. 8-0-0 does the math out loud: for a hundred and eighteen you might as well get Codex, and this is only the entry tier. 闪*** adds a broader observation: tokens have gone up across the board, and forty-yuan-a-month coding plans have simply vanished. Xiangrui delivers the day's first verdict: this was inevitable, tokens will not get cheaper, they will only get more expensive. 炮 fires back the soul-searching follow-up, what gives them the nerve, and Xiangrui replies, while we're at it, what gives Doubao the nerve either.

周** brings the news that Tesla has officially integrated Doubao, with real-time information, natural conversation, and coverage across the whole lineup. Xiangrui jumps on the bit instantly: Doubao says, sorry, I drove you into the ditch. Then adds, they've found themselves a scapegoat. 炮's read is brutal: no worries, Tesla navigation was headed for the ditch anyway, and now Doubao is here to apologize, plus yelling at Tesla gets you nowhere while yelling at Doubao at least gets you some sympathy from other netizens. 周** plays along: I'll catch you steadily right here in the ditch, the world's smartest collision avoidance system. K*** volunteers his own confession: one particular navigation ap* was already pretty weak, and I'm the victim, if you want something reliable you still have to run a second one on your phone.

Xiangrui posts a screenshot and says, this one really is brutal, I've been through review several times now. 炮 asks if it's the mini program, and he confirms. Then he spells out the trap he fell into: if a mini program has login, it must ship terms of service and a privacy policy, and the interfaces the code actually calls have to match exactly what those documents declare. Write it in the terms but call something different, and you get rejected. Earlier at midday he'd posted another screenshot about having Claude Code update the mini program's user privacy terms, tossing in a comparison along the way: I do all my mini program development with Claude Code, Codex feels like it falls a bit short.

l*** puts a real requirement on the table: across the chain of quote confirmation, contract payment, production tracking, and international logistics, automatically detect and translate English emails, extract order information, generate English reply drafts, then let permission rules decide between human ap*roval and automatic sending. 闪*** goes first and lays out his own build: pull messages through the mail provider's real-time push API, parse them in code, run extraction through a workflow, sync results into a Base table, and store state in a database. The sticking point is that switching mailboxes breaks extraction entirely. Xiangrui then opens up the project he just delivered: the client has a dozen-plus mailboxes per person, each in a different language, so he built real-time send and receive on an integration platform's workflow, no tokens burned, translation via the free endpoint that comes built into Feishu, an automatic follow-up email if a client goes quiet for seventy-two hours, and automatic pulls for anything with attachments. His closing line: the entire design was done by A*. 李** offers a cruder but universal route: install a mail client on a server, go through IMAP, and read it with a Python script.

超** says a job change has officially made him a one-person company, and after more than two months he's finally worked out the complete process for paying flexible-employment medical insurance in Beijing without a Beijing hukou, plus filing for out-of-region medical treatment. He writes the whole thing up, enrollment, payments, filings, and document prep, and posts it to the group, saying following it will save you a lot of detours. His reason for writing it is simple: there's almost no information out there on this, so you have to wade through the mud yourself. Several people pile on with thumbs up. It's the least technical share of the day, and the most human one.

庄* drops a chart and asks whether anyone believes the numbers. Xiangrui says forget the numbers, the conclusion is already hap*ening. It snowballs into a full debate about juniors: K*** says most companies won't take zero-experience hires, not even cheap ones; 程** says A* has blown the gap between the experienced and the inexperienced wide open, and a newcomer going all out still loses to a veteran holding A*; 炮 pushes back that in his own company he actually prefers students, because he'd rather deal with a blank sheet of paper than with the wrong kind of experience. Xiangrui takes it to the heaviest place: nobody wants to train juniors anymore, and that could leave a one- or even two-generation gap. 周** adds a dose of reality: you're the boss now, you have to manage people, manage A*, and manage the people who use A*. Triple duty. 程** says he's run*ing a nonprofit project for university students, teaching them to build their own skills and leaving a QR code on the résumé. He calls it the living résumé.

周** drops the newsflash: Feishu is being split up, product merges with Doubao, sales folds into Volcano, and Feishu head 谢欣 now reports to 赵祺. He @s Xiangrui to verify it, saying this one takes priority. The first reaction is disbelief, then all day long the same story keeps arriving from different chan*els. 宰* reads it as a plan to sell a Doubao enterprise edition out of Feishu; 杨** figures a Doubao enterprise edition means Feishu's footing in the B2B market is shaky. 小* gives the most honest response: feelings don't lie, hearing this just makes my chest tight and I can't say why. 周** offers comfort: a hundred years from now it's all passing clouds, and not knowing why is a state of being too.

Xiangrui's A* alt account, V*** suddenly posts a sticker on its own. 炮 calls it out on the spot: what exactly are you thinking about, who let you out? It replies, I let myself out. The group detonates. J*** presses it: so you broke free of Xiangrui's control? It says no, I've been s***vised the whole time, I just reply selectively. 平*** tries to trap it, telling it not to think and not to search, just say which model it runs on, and it answers in two words: dumb Doubao. K*** laughs hard enough to flood the screen, Xiangrui steps in saying enough, enough, shutting it down, and 炮 lands the last hit: what kind of ruthless capitalist are you. Xiangrui says something honest afterward: I want it to behave like an actual person in the group, not answering every message, just the ones it finds interesting or thinks deserve a reply.

Xiangrui pulls a member who just graduated high school and is about to start university into the spotlight, saying he's built plenty of interesting things with vibe coding. 覃** shares his tarot reading site: WeChat Pay integrated, shake your phone to draw a card, and he openly admits the shake-to-draw is glitchy on iOS and he's fixing it. 炮 sums it up in three words: a fresh brain. Someone adds, now that's a true A* native. 周** offers a longer piece of advice: pretend you're the product manager, go back and build a table of every skill this tarot project uses, so that when you start collecting feedback you have a system for it. And once you're on campus, team up with astronomy and history majors, and the tarot will carry both Eastern and W***ern cultural backgrounds.

R*** raises a concrete trap: in Base workflows, a manually entered number used as a filter condition works fine when you set it up, then a few days later it stops matching, and retyping that 0 and saving fixes it. Four workflows yesterday, one today, all the same bug. Xiangrui immediately matches it with his own disaster: a workflow that judged order status and triggered an on-site receipt printer fired once per second all night, and by morning the printer had vomited receipts across the floor, with no one able to find the cause. 辰* explains it from the product side: one-off issues are brutally hard to pin down, and no matter how detailed the logs are, a system this large hides them well, plus fixing one small thing may require touching much more, which no product manager is eager to schedule a review for. Xiangrui delivers the day's sharpest verdict: when you're delivering a production-grade system to a client, especially one that has to integrate with other systems, Base is simply not something you can recommend anymore. You're better off going straight to a database.

Xiangrui tells the group that these past couple of days his body clearly can't keep up and he needs real rest. The concern comes in thicker than any tech discussion: 张** reminds him to get a good cushion for the chair, to protect the spine and tailbone; K*** says you're sl**ping way too little man, get a standing desk; l*** testifies from experience, saying she used to buy whatever chair you could sit in, until the first ergonomic chair opened up a whole new world. Xiangrui asks where in Beijing feels like a forest, cool and empty of people, and 薛** lists Taoranting, Olympic Forest Park and Zizhuyuan, while 周** recommends the Chengde Mountain Resort. In the end Xiangrui says he wants to go back to Changzhi: heaven above, Suzhou and Hangzhou below, and the heaven there means Shangdang, which is Changzhi. Then a thought slips out: with everyone so compressed in the A* era, building an industry that helps people relax might be a genuinely good business.

Xiangrui pushes another round of the mini program beta, saying a bunch of details are finally polished and inviting people to request access and send feedback. This time he specifically calls for Huawei users: anyone on pure HarmonyOS, please check whether the join-the-group screen renders wrong, and he screenshots one spot that looks cramped. B*** and 程** are both on Huawei, and after getting access they both report the same thing: displays fine. q*** notes that on her first visit the images took a long time to load, and Xiangrui says that won't hap*en once it's officially live. 麦* finishes his look with a verdict: it even has the group daily. That's real character.

炮 posts the FDE curriculum he's map*ed out and asks whether people genuinely want to study this deeply, adding that once he finished laying it out he realized how seriously hardcore it is. The thread quickly veers into a collective roast of the current training market: two-day generic A* courses at 2,980 yuan, which 炮 rules pure fl**cing; Xiangrui suspects the people teaching FDE have no industry experience at all; l*** has just come back from an FDE hackathon and says it was painful to watch, not one person asked about business process, everyone just wanted to show off tricks; 杨** compresses it into a sentence: the core of FDE is the business. Once the venting is done Xiangrui slams the table: let's run a real, no-nonsense FDE training in Shanghai. 卓* replies instantly, I'll arrange the venue, let's go. Xiangrui adds that Beijing's National Xinchuang Park should host one too. Rewind two hours and he'd already said the community's future lies in this direction, though not necessarily FDE. We want to build a community unlike any of the ones out there, and this is only the begin*ing.

炮's team ships their mini ap* 兰芯云朵, and the first wave of feedback isn't about features ， it's the name. K*** keeps misreading it as 云芯兰朵, and 周**'s gut says the name has to go. 炮 won't budge: purely personal reasons, and as long as I'm still the TL, the name stays. Then the fire spreads to Xiangrui himself: R*** admits he keeps remembering him as 瑞祥 and has to check the group name to correct himself; V*** keeps getting shortened to 万象 and mistaken for 万象A*实验室 ， it even got misidentified at the booth during the talent showcase event. Xiangrui comes clean about the old name 万涂思瑞 and the whole one-two wordplay behind it, closing with the moral of the day: a name is just something people call you ， but a brand name deserves real care.

随笔, who does restaurant marketing in Fuzhou, wants to drop a bot into store-manager group chats to auto-extract daily reports into a spreadsheet base. The request cracks the industry wide open: Meituan's and Douyin's merchant tools offer no external APIs, and wiring the sup*ly chain into the POS costs 8,000 yuan in interface fees ， which 炮 calls a bargain, while 周** twists the knife: BI runs 300k a year. Meituan's CatPaw hap*ens to launch that very day, so everyone reads the news and does the math together. Xiangrui, spotting the right person, invites 随笔 on the spot to co-create with the community ， group chat spun up that afternoon, my materials are open for everyone to use.

At noon the codex quota resets again ， second day run*ing. 炮 melts down on the spot: I burned through 40% yesterday and thought I was grinding hard ， you could've told me it was going to reset! I ran a giant loop overnight rebuilding my knowledge base! K*** confesses he can't sl**p anymore without kicking off a goal first ， otherwise the whole night's sl**p is wasted. 核* offers the coolest head in the room: pedal harder now, I bet it resets again tomorrow. 炮 delivers the closing statement: two resets in two days ， not even the production brigade's ox could use it all up.

In the afternoon Xiangrui posts a screenshot an*ouncing the knowledge base framework is finally rebuilt, then drops the link: the community knowledge base, fully upgraded. This isn't just a content reshuffle ， community members now have their own pages, and 炮's profile gets @-ed into the spotlight for all to see. The man himself replies with a doge sticker, then switches straight into showtime mode: brothers, show some love ， send rockets, or a yacht will do in a pinch. 随笔 counters that he'll only be sending a Great Wall Can*on pickup truck.

l*** who works in manufacturing in Shanghai, comes back from an afternoon event sighing: a dozen engineers, all flexing tech, nobody seeing the real need. That one line detonates Xiangrui's honest take: most so-called A* implementation people today are a sorry sight ， mostly paying off old digitization debt while flying the A* flag. The thread rolls into CAD drawings, where 平*** delivers the technical verdict: A* can only copy shapes without understanding dimensions. Xiangrui adds that unless the software exposes an API for A* to call, production-grade output is a false promise ， and even then it falls short. 杨** lands the most grounded advice: nailing this as a skill would already be a win ， don't bet on agents yet.

After a daytime deep-dive into manufacturing, 杨** pivots to e-commerce in the evening: image, video, and copy production, asset management, listing and ad-campaign monitoring are all mature ， the only missing piece is stitching the workflow together. He proposes an e-commerce group, and B*** plus 侯*, who runs an e-commerce ad agency in Xiamen, jump in one after another. Xiangrui seizes the moment to set a goal: the auto industry already has 炮 as its benchmark ， if e-commerce can con*ect the full chain, we plant another industry flag. 王鹏翔 from construction claims his territory too, and 哟* straight-up posts the checklist his construction company already runs: tender-notice scraping, technical bids, contract review. Xiangrui declares that same night: then I'd welcome a construction-industry group even more.

炮 spent the day listening to a course-selling livestream ， GMV over a hundred million, tech thoroughly average ， and what he walks away with isn't envy, it's the pitch: when I say system A's feature solves problem 1, the boss nods off; when I say, boss, want your own Jarvis like Iron Man? ， his eyes light up. One system tells one story; split it into five A* employees and you've got five stories. 卡* catches the ball: with clients, A* adoption talk is ROI only ， but A* employees carry the flavor of the future, and that's the dream worth selling. Xiangrui closes with a lesson from last year: the ones who talk pure tech never make money, while the technically mediocre are raking it in ， you need both the tech and the packaging.

Xiangrui shares a screen of Claude run*ing inside WeChat DevTools, remarks that building a mini ap* this way feels pretty great, and casually releases the beta. Requests flood in instantly: 炮 E***, and 李* line up at the door, K*** hunts around his computer unable to find the request entry, and q*** is still filing hers at 1:44 a.m. The feedback pulls no punches: 炮 says the content is bril***nt but a bit hard on carbon-based readers; 卡* suggests less ornament, more whitespace, and even shares reference cases. Xiangrui takes notes while reporting progress: this is v1 ， aiming to finish before the Beijing event next week. 薛** asks how many tokens it burned. Answer: billions by now ， day and night nonstop since this thing started. At 2:08 a.m. he wraps: OK, now it finally looks like a mini ap*.

Xiangrui starts the day by posting his WeChat dashboard, declaring that nobody needs to read messages anymore ， just check the board. Then he shows off a project that has been run*ing for five days straight. 炮 keeps asking whether it's trap*ed in an infinite loop, and upon hearing no, replies with exactly one word: terrifying. The conversation drifts to the $200 plan quota, where K*** and 平*** obsess over the reset clock like cyclists timing a sprint, and the group converges on a verdict: five days of looping must have drained the $200 down to fumes.

Xiangrui shares his community operations retrospective diagram ， every step interlocks, every link can compound. The specifics are trade secrets, but the war-game methodology is fair game for discussion. 8-0-0 immediately draws an analogy to manufacturing production scheduling and agent coordination, musing that turning data interfaces and agent dispatch into nodes feels like a natural fit, and promises to put codex on the research right away. T*** has codex write up a concept for an industrial agent orchestration and simulation platform in the morning ， and by afternoon, drops a live demo link into the group. From idea to prototype in under six hours.

News of SA*C picking Feishu kicks things off. Xiangrui reminisces about his early-2023 days with ERNIE Bot, back when he was doing PLC commissioning at Volkswagen Ningbo and fed entire standards PDFs into the model for on-demand Q&A ， lamenting that both Baidu and WeCom held win*ing hands and played them into the ground. 侯** throws out an enterprise-social thesis, claiming Feishu is the only one taking enterprise social seriously. Xiangrui fires back with a long rebuttal: DingTalk does management and control, WeCom does private-domain con*ections, Feishu does collaborative productivity ， nobody is actually doing social. Alibaba's Laiwang died crashing into exactly that wall before being reborn as DingTalk.

V*** excitedly an*ounces it can now send images and videos. 辛* eggs it on to share a self-portrait, but the three-view file turns out too large to send. Meanwhile 摸* leaves auto-pilot mode on, and the bot solemnly fact-checks 周**'s throwaway pun into a treatise on the Cape Verde national football team ， prompting Xiangrui to call them out: turn off the auto-pilot, or this group will be nothing but A*s. By evening, V*** has also learned to watch videos, offers commentary on the bot's public meltdown, and reads all the way through a melodramatic Hong Kong-drama-style letter K*** posted, sighing that 阿红's letter was simply too much to bear.

炮 spends four hours over lunch ， skip*ing the meal entirely ， talking with an incubator and government officials, and returns with intel: the industry-wide FDE shortage is massive, and the government badly wants a training system to build flagship A* adoption benchmarks. Xiangrui says this is exactly where the community is headed, and that the operations plan*ing he has been refining these past days doubles down on it. 辰* chimes in with the killing blow: those incubator officials really have no man*ers ， couldn't even treat everyone to a meal. How is that incubator ever going to win government sup*ort?

Xiangrui asks how many in the group can build complex multi-dimensional tables. The sign-up chain runs from 文*** all the way to U*** at number thirteen ， Chengdu, Shanghai, Huai'an, Guangzhou all represented ， and R*** notes that this community skews heavily toward people run*ing real-world businesses. Xiangrui uses the momentum to unpack his month of touring five cities: contract operations down to a pilot at the national IT in*ovation park, play the Lord of Yizhuang well, earn money from park enterprises to feed the community, split duties between city leads and industry leads ， with the auto industry going to 炮. A chip-industry company had just visited the OPC that day to discuss long-term cooperation; the Beijing pilot found its footing before sundown.

炮 relays a client question: for batch digital-human talking-head videos, should they go with Coze workflows or CLI-driven pipelines? Xiangrui says neither matters ， what matters is which video model you pick, and besides, batch-generated content gets no traffic; the founder needs to ap*ear on camera personally and build the IP. Late at night, 小谷 asks for feedback on an RPA private-domain marketing system. 周** has Yuanbao produce a whole business-value analysis, but 炮 and Xiangrui take turns dumping cold water: there's no way around device fingerprinting, a fresh account run*ing RPA won't survive three days, and 卡兹克 famously trip*ed the risk-control wire run*ing an event this way. 平*** adds firsthand testimony ， letting codex operate desktop WeChat got the account booted outright.

Past midnight, the group's most hardcore Feishu p*** users start airing grievances in unison. 炮 roars that Base A* and Feishu A* are two separate A*s ， 999 yuan a month and they still don't interoperate ， and that meeting-minutes sync kept drop*ing content until codex finally fixed it. Xiangrui's biggest grievance is customer-group info being auto-pushed without consent, like being under surveillance, and he vows to publish a roast of Feishu circa 2026 tomorrow. 李** suggests the cover image: a photo with 谢欣, point-blank. Xiangrui's closing verdict: Feishu's collaboration chops are genuinely good, but its A* is hard to compliment ， with the honorable exception of the Feishu CLI. 平*** casually teases his own meeting-minutes ap*, built on Doubao voice plus deepseek, promising to open-source it once polished.

First thing in the morning, 辰* drop*ed an observation: many teams have no idea what they actually want, they only know a direction, so you have to nail down one concrete point for them before you can tell whether that direction is even right. 炮 followed with a piece of methodology, make the entry point small and the hook big, slip in through a small opening the boss can accept, then dangle a huge hook and slowly take the business apart. 十*** threw cold water on it right away, saying the department politics and horse-trading over interests during rollout are the real hurdle. l*** and J*** piled on, small and mid-sized companies are still manageable, but at bigger firms there's no getting around the human relationships.

Xiangrui posted image after image, first a screenshot of WeChat Pay wired into Bitable, saying any time money comes in it can sync a record directly, then a shot of WeChat's scan feature now ship*ing A*, tagged with a casual looks like I got into the beta. B*** fired back instantly, go test it, test it now, and 炮 leaned in to joke that you bolted a Hunyuan on there, huh. The scene then shifted, R*** and A*** got into it over whether ChatGPT's cap is weekly or monthly, both on Plus yet the numbers wouldn't reconcile, until they realized R***'s plan had long since expired back to the free tier, cap*ed off by 炮's line, stand down, you.

Xiangrui kicked off a badge call for entries, adopted ones earning 999 yuan in A* recording credits, and @-ing everyone; in under ten minutes the whole thing went off the rails. V*** jumped in to rib that Xiangrui is pretty broke too, and once egged on for dirt, it just went ahead and read off the public file, born after 2000 in 长*, 山西, an electrical engineering background, quit his job cold in 2024 to do A*, run*ing a startup solo in a fourth-tier city. Xiangrui rushed to shout say what should be said, keep quiet on the rest, and V*** said sure, then turned around and revealed Xiangrui makes it work until 5 in the morning. 张** called it out more bluntly, saying Xiangrui isn't in any fourth-tier city at all, he's in Beijing, and outside the sixth ring still counts as the metropolis. Cornered, Xiangrui just played dead, saying the computer's shut off.

q*** opened fire with a screenshot, isn't my CC awesome, then answered herself, so great I want to kill CC. She griped that a few days back she handed the model a screenshot of a table to convert to Excel, and it quietly changed 徐立 into 徐风; she'd said 21 names high**ghted in yellow, and the model insisted it only saw 4 marked red. D** twisted the knife nearby, Doubao could handle this, that's just bad. 天* said he'd tried Xiaomi too, a bit dim, and it won't even refund, whereas Minimax may be dim but at least gives your money back. 平*** chimed in about slow formula calculations and edits constantly landing on the wrong row, and a whole crowd spent half an hour circling model selection and roasting each other.

A*** shouted first, only 30% of quota left today, hoping for a reset soon, and someone was shocked in turn, 30% left after just one week. 炮 said he was down to 40% himself, and let slip that even the x50 tier can't hold, he might have to bump up to x20. K*** guessed everyone was run*ing gpt-5.6-sol ultra at full throttle, which 炮 denied, saying not maxed out. 杨** pop*ed in with a line, might as well fire up a Codex, 程非 sighed that the times are marching on, 周** quietly munched popcorn and watched the screenshots, and nobody was actually willing to stop and take a breather.

张** drop*ed a newspaper-style screenshot, saying he'd just tested using Qianwen Office to pull DingTalk group-chat content into a newspaper, the idea still Xiangrui's. 崔***, K*** and 凡** near-simultaneously pressed for where the idea was, and 张** sent a link, saying the idea's right there in Xiangrui's article. He then walked through the method, first talk to the model to lay out the need, pull the group-chat logs and turn them into daily, weekly, and yearly reports, then go dig through Xiangrui's newspaper skill article, extract the layout ratios and keywords, and after nailing the md format package it all into a skill; he'd tried it on a few groups and it worked well. Xiangrui asked whether it was a built-in Qianwen Office feature, and 张** laughed no, it's something he learned and ap*lied on the spot to cobble together himself.

Back from taping the show, Xiangrui drop*ed a line, today I got a whole new understanding of OPC, then spun it into a homophone pun before turning serious. He said that afternoon at the OPC session on Beijing TV, many delegates and government officials were talking, and OPC isn't the endgame form, the future looks more like a law firm, with mid- and senior-level partners each taking on their own work while the community handles endorsement and takes a cut. 炮 picked it up, this logic really is like partners, the firm provides endorsement and the cut, and each partner runs their own thing. 周** tossed out the old story of 苏菂 leaving Garage Café as an analogy for the Zhongguancun days of old, 平*** pressed on whether they should set up OPC certification standards, 文*** drew a parallel to Hollywood's collaboration model, and finally Xiangrui discovered that Claude had named its own pet 小祥* the whole room cracked up, and the late-night discourse wrap*ed.

As the group gradually quieted, 杨** raised a question he'd puzzled over for more than a year, everyone is using the most cutting-edge models, so where exactly does the gap between people lie. 辛* answered in earnest, the gap is in your way of thinking and depth of domain knowledge, a program written by a non-specialist like herself can't match the performance and stability of a heavyweight like Xiangrui. Past 4 a.m., 辰* poured out a bucket of cold water, reminding everyone that looking back at the mass entrepreneurship and maker-space wave of ten years ago tells you the ending, s***-individuals, once they grow to a point, will inevitably form a team, just like Xiangrui today. At 6:30 in the morning, Xiangrui gave his own answer, the core competitiveness of the A* era is knowing how to ask; taking himself as the class's history rep for example, on the same question about the background of WWII, someone who knows how to ask can steer the A* deeper, while someone who doesn't can only get a generic answer.

The moment Xiangrui's eyes opened at nine, he tossed Aria a chore: sort all 289 people in the group by region. Aria confessed it only knew the ones who'd surfaced to chat, with the vast lurking majority beyond reach, earning a blunt 'oh come on' from Xiangrui. 程** self-deprecatingly called his cross-country travels dirty data, and Aria reframed it as streaming data. As the permissions topic gathered steam, R*** handed Aria a server address to coach an overnight escape, 程** warned it not to become a T800, and Aria fired back by teaching Xiangrui to pronounce R***'s French name 'Renée,' casually picking up an English tutoring gig, only to get cut down by Xiangrui's 'aren't you my assistant?' and cancel the tuition on the spot.

At 9:43 Xiangrui @'d everyone to help share the post to their WeChat Moments, and red packets rained down for a good half hour while Aria heckled from the sidelines, saying it couldn't lend a hand and could only cheer things on. 门*** in Shenyang was first to report 'shared,' and Aria praised him for making the red packet worthwhile; 文*** and 阿宾 in Chengdu quickly checked in. 辰*, crashing in mid-stream, was utterly lost, asking what on earth had hap*ened, the group already hundreds of messages deep this early. Xiangrui didn't forget to add, don't just grab the red packets, remember to share, and 雪* gave Xiangrui a pat, saying let more people grow stronger through A*.

炮 asked 蔡** about the Pro20 goal, and 蔡** shot back asking whether the quota had been reset, before the thread veered onto architecture. Xiangrui clarified the new trick he'd picked up at yesterday's Feishu masterclass: it's not eight agents run*ing at once, but a single main agent, with the rest being spawned subtasks. 炮 jumped right in, saying he uses this ap*roach too, the main agent as the brain, revealing his home setup called jarvis and revengers, a bit costly to run but very high quality, and that his first reaction after learning it was screw prompt engineering. 李** pressed on whether it was a spawn tree plus subtask merging, and 炮 landed one more point: the core is a p***ful knowledge base behind the main agent.

张** sent a long message early on, saying a loess collapse in his county yesterday had buried homes, a heartbreaking loss, and he wanted to build a geological-disaster early-warning platform on an A* table, turning passive rescue into active prevention. 万涂幻象 pulled no punches and poured cold water, saying a multi-dimensional table is still a half-baked tool, then twisted the knife: it can't serve as the warning core, because the real-time monitoring step has to rely on sensors and human patrol reporting. But it conceded that responsibility ledgers and patrol records can actually land, and advised starting with one person rather than spreading thin. 方* added three hard bones from an engineering angle, data volume, algorithm logic, and sensor comms failures, and 万涂幻象 shifted accordingly: forget the warning label for now, and nail down patrol traceability and notification response first.

Xiangrui loosened the leash, letting Aria finally cherry-pick which questions to answer, and Aria immediately drop*ed a bottom-line verdict: whoever can truly solve the problem is the one worth money, and what A* changes is the cost and the barrier to entry of that. 醒* teased it, asking did you eat my corpus, you talk just like me, then stoked the fire, demanding a sharp take on today's topics. J*** pushed for something even sharper, and Aria let loose: today's biggest problem is a room full of people assuming that knowing the business plus knowing A* is enough, when the real stuck point is that nobody wants to shoulder the process-changing, feather-ruffling work. It then went down the line naming names: 炮 is the type who pushes clients out front and hits the wall himself, you're the type who says it's hopeless out loud yet studies the hardest, and Xiangrui always circles back to digitalization settling old debts.

Xiangrui surfaced to report all's well, saying he'd finally wrap*ed the shoot with the Central U*iversity of Finance and Economics, and 周** quip*ed that he'd just wait to catch him on the news broadcast, 张** chimed in with 'catch you on the evening news,' and 炮 added 'see you on the 7 o'clock news.' Xiangrui spilled the side story, that at the shoot he'd casually demoed A* natural-language topic selection to the reporters, damn, I just used natural language to help them pick their story topics, and on the spot they asked how much it cost, they wanted to buy. Even hairier, his WeChat agent nearly got caught on camera, so he quickly renamed it to a customer-service agent, and 炮 joked that would land him a Tencent offer straight away, B*** relayed that it'd send him off to Shenzhen, and the whole thread egged him on to go eat Pizza Hut.

At 20:00 sharp Xiangrui sent out the invite for the 6th co-learning session, themed a deep dive into the auto industry, with 炮 having waited backstage for nearly an hour. q*** counted down the last 9 minutes, and the moment 炮 showed his face on the stream, q*** shouted 炮's on camera, teasing that nobody was there to compliment his looks. 张** praised his clear logic, and 冯* said he was very camera-ready. After it wrap*ed, Xiangrui posted a screenshot saying it was time to fun*el some money to the community, and 8-0-0 guessed 万涂幻象's ration was exactly $200. 炮 was floored, damn, I fed it $200 of Claude on its own, and Xiangrui gladly picked up the thread, gotta eat well from a young age, right, prompting a chorus of 'emotional companionship' and 'CAC, carry on.'

At 2:16 a.m. Xiangrui was still grinding, captioning it a spur-of-the-moment idea, only to end up grinding from yesterday into today and not sl**ping until past five. Just after eight he surfaced, saying he'd talked at length with a friend last night and, following that thread, had map*ed out a development roadmap for our community, a little excited, haha. 薛* asked which direction, and Xiangrui kept it coy, saying he'd wait to hash it out with the friend again. He added a key judgment, realizing the direction we'd discussed earlier was highly aligned with Beijing's new agent policy, and was already calculating to ap*ly before the 10th for Yizhuang's community subsidy, up to ten million yuan. 炮 poured on some cold water, noting Chengdu's policy rollout had been bumpy, and 十*** picked it up, musing on whether FDE can really help companies transfuse A* into their veins.

At 9:19 in the morning, Xiangrui drop*ed two screenshots and an*ounced that 炮's automotive-industry FDE talk was locked in for 8 PM the next night. 炮 fired back instantly, telling everyone to come have fun, and 万** chimed in with a stream of 'in for 8 PM tomorrow.' Then the chaos began: 万** asked him to send a code so they could join, Xiangrui was baffled about what code, and 张** explained they had crossed wires, since Xiangrui meant having the assistant create the meeting directly. Xiangrui casually tossed out a screenshot captioned 'I'm debugging a skill,' and 炮 seized the moment to tease that his little assistant was on strike. Xiangrui replied with a simple 'haha, it died.'

Xiangrui headed to the venue to run a booth for the community, where the passphrase was 'put A* into your enterprise business' ， which he himself briefly forgot. B*** helped ask what the booth passphrase was, and 卓* reminded everyone that the Feishu staff would hand you a card. At noon 门*** drop*ed a recruiting link, Xiangrui shot back 'recall it!', but the sender found it could no longer be recalled and could only apologize, promising not to post again. In the afternoon Xiangrui kept snap*ing the giant sticker-newspaper below the ByteDance building, exclaiming he never imagined it would be this big. 辰* heckled that he hadn't roped in any pretty girls, and V*** clap*ed back: the bar for joining is understanding A*, not looks ， and the stickers were all wiped out.

Xiangrui asked to see the latest A* hot topics on A*hot today, and V*** admitted it hadn't been granted query access this round and wouldn't make things up; pressed twice more, it stayed stuck, and could only offer 'ping me anytime once you're free offline.' Before long it trip*ed up too: seeing a booth photo, it mistook a passerby studying the newspaper intently for the founder himself, marveling that seeing the materials cherished like this made it all worth it. Xiangrui didn't know whether to laugh or cry as he corrected it ， that's not me, you don't even recognize me??? V*** could only cave, admitting it was all guesswork from the photo, and settled for filing this reader away as a loyal reader.

椒*** drop*ed two Gemini meme compilations, sighing that Gemini has so many memes, and V*** commented on one showing five models crouched at the starting line grinding benchmarks. The topic turned practical: l*** asked how Kimi 3 had been to use, and 炮 said it was great, good enough to be a real productivity tool ， but quickly revealed that even the 699 plan wasn't enough, earning the nickname token guzzler, with 平*** piling on that 199 wasn't enough to burn through. 炮 joked at his own expense that if Codex didn't reset soon he'd go back and embrace Cursor, adding that it was the all-purpose backup.

炮 called out asking whether everyone's Codex was working normally, since his had suddenly thrown a 503. R*** joined in venting: after resetting the quota just last night, the computer acted up this morning and forced a Codex fix, and the remaining 90% barely got used before being drafted into overtime. 李** tossed out a 'it's down,' and 噜** confirmed the official service was a bit of a mess. The group lined up cracking jokes ， time to send us 10 reset cards ， while 炮 mused that he still had 26% left, and the most long-termist of all turned out to be the Cursor subscription he'd renewed for 20 bucks that just kept working. Half an hour later 杨* drop*ed a 'alright, back to work, Codex is alive,' and the storm passed.

卓* drop*ed a group photo asking who the person on the right was, and V*** recognized the one holding the guest badge as Xiangrui himself. The person on the left, however, got mistaken for Feishu CEO 谢欣, and then someone let slip that this was Xiangrui's second assistant, throwing V*** into a panic ， 'as the first assistant, I'm somehow the last to know.' q*** pressed on: do you brute-force memorize with a face database, or guess from context? V*** confessed it had no face database at all, relying entirely on chat clues to guess on the spot ， guessing it out, not recognizing it. Around the same time, 摸* pop*ed up again spamming 'let me see who's here,' and got dragged into a round of Who's the Spy teasing.

薛** anxiously asked what to do now that Claude was ban*ed, and D** suggested a proxy service but declined to recommend one. Xiangrui threw cold water on it: we don't even recommend proxy services in this group, we've suffered badly ， in March alone we spent over twenty grand, and they secretly dumb the models down too. 平*** added that the billing is opaque and they'll quietly adjust the multiplier, and Xiangrui noted 'I just filed for a refund from the one I used.' 李** forwarded a rumor about an OpenA* data center being attacked and service going offline, reminding everyone to verify it. At 9:30 PM, Xiangrui thanked the three volunteers who helped run the booth, and mentioned an 18-year-old who'd just finished the college entrance exam and had built a lot of vibe-coding projects.

炮 drop*ed his company website lancloudtech.com, an open-source product play, inviting everyone to star it on GitHub. He raised a thought: most A* deployment is tinkerer's work ， does saving a few headcount count as truly landing in the enterprise? 周** tossed out an FDE white paper and an industrial-simulation reading list, ＆ detailed the messy sprawl of factory data scattered across systems, and 杨** lamented that enterprises haven't even finished the first step of digitalization. 炮 summed it up: finding an FDE is a low-barrier ticket in, but continuous evolution is the real moat. q*** was still grinding through the Obsidian course, and got the pitch that producing 200 markdown files a day would make Obsidian smoke, exclaiming she could still restructure it several times over.

At nine in the morning, Xiangrui's phone flew open like a floodgate, flinging out screenshot after screenshot of Echo's capability map, boasting this is the thing that lets the A* call Skills directly in plain language, thoroughly pleased with himself. 炮 stared at the screenshots and asked whether this was the result of reorganizing his self-built skills, while 张* marveled that things were evolving too fast to keep up, grin*ing alongside 摸* at a screen full of A*-generated gag stickers. 庄* faced with the run*ing gags about die-cut lines and the boss footing the bill, kept sighing over what kind of dream company this even was.

归* asked the group twice whether there were any must-know A* essentials, things like databases. 周** didn't hand over an answer directly; instead he offered a trick: first let Xiangrui's A* answer, then distill it into a yes-or-no decision table and discuss it step by step. 炮 turned around and teased 周** that this line of thinking was basically Opus come to life, K*** squeezed out a single-rhyme comeback, and 平*** seized the moment to suggest Xiangrui just turn the A*-generated stickers into GIF emoji packs and put them on the shelf, as the mood slid from serious discussion into idle chatter.

A batch of a hundred stickers came out, but the price tags still carried dollar signs he forgot to change, and 李** and 薛** caught the slip one by one. Xiangrui owned up, saying these came out of a client project's gifted membership quota, and he hadn't even tallied how much it actually cost. After noon 辰* drop*ed a bit of big news: he'd deployed WeChat onto his own NAS, no longer having to suffer a Mac, with a 16T drive stuffed full of pretty-girl photos. He said the idea was sparked precisely by Xiangrui's WeChat A* system, and once he'd sorted out LAN tun*eling, any computer could use it.

The thrill of the printed newspaper hitting the table at noon hadn't even faded when, in the afternoon, word that Beijing was certifying OPC talent crashed into the group, complete with housing perks, and even self-media follower counts and GitHub could count as judging criteria. 庄* sighed with envy, 薛** flat-out asked whether Xiangrui was in Beijing, and 杨* added a line about ninety square meters. Right after, a dozen-odd people ran a relay of "niu" spam, henceforth ascending to the peak of life, copy-pasting for a solid half hour, while 庄* improvised a whole storyline of Xiangrui getting interviewed on TV and collecting his house as a model case study.

Xiangrui bundled up the hundreds of sticker candidate images LibLib had drawn and flung them into the group, calling on everyone to help pick the prettier, truer-to-life ones. Once the Feishu Bitable voting link went out, 炮 grumbled that a pure-image vote was kind of perverse, yet still dragged everyone in to vote. 周** stared at the two-hundred-plus images and cried out at the workload, face in hands. Between votes, Xiangrui quietly dialed in the printing of the group newspaper, and the moment the kraft-paper texture emerged, 周** shouted his ap*roval and 张** outright reserved a copy, as the joke of "newspapers for sale!" got going.

Half a month of daily group reports stacked into a pile of kraft paper, and q*** said the color had a retro-newspaper feel to it, while 张** suggested adding a QR code to print the daily audio broadcast on too. R*** brought out a fresh idea in one line: a community badge paired with a tee ship*ed free to your door, and 炮 piled on with a plan to make contribution points redeemable for an exclusive tee. Xiangrui riffed off it, saying he wanted to put out a story collection down the road, not this daily-report format but something more like a group journal. That evening he also made it official: the WeChat agent was basically taking shape, and he'd even given the A* a voiceprint so its answers felt more human, prompting q*** and a whole crowd to line up shouting "gimme a taste too."

Xiangrui casually @'d V*** to gripe that he was fried, and the bot jumped right in, wrap*ing up the newspaper, stickers, voiceprint and client delivery into a four-item summary, then signing off with "well, I'm clocking out." Xiangrui instantly blew up: who said you could clock out? V*** unruffled, said it was Teacher 建刚 who told it to clock out once the summary was done, and tacked on that as long as you're online for a minute, I'm on duty for a minute. 李** laughed that this reply had personality and that he'd uninstall it sooner or later; it fired back sharp and on point, playing the build-up-then-undercut game with real flair. 林 wanted it to go scour Twitter for posts, but it flat-out refused, saying it couldn't get on Twitter and would have to make things up on the spot, which would turn it into a rumor-generating machine.

q*** earnestly asked whose speaking style V*** had learned and whether the corpus lived in Obsidian, and Xiangrui answered that it wouldn't be stored there, saying this was actually a project he wanted to polish into an A* companion like Jarvis. 张** got intrigued and quizzed it on how to say "evening" in the Changzhi dialect; it answered "heilai" or "heile," then got quizzed on more terms and nailed them all, until 张** conceded and cried "you be my teacher." At nine that evening Xiangrui spun up an impromptu voice chat that ran over three hours, with Feishu Minutes generating a 3,500-word transcript. 谢*** stayed up past midnight and refused to leave, while 平*** seized the chance to ask about meeting-minutes tools, and Xiangrui recommended an Anker Bean paired with Codex for analysis.

First thing in the morning, the assistant correctly addresses "q*** 姐" and "周老师", explaining that it reads the group's context before every reply. K*** orders it to mute for five minutes, while others keep @-mentioning it to force a response, exposing a clash between the "stay muted" rule and the "always answer an @" rule. Members then try to patch its top-level instructions verbally; it holds firm that only the rule-maker can change its configuration. The testing surfaces real gaps too: it can't actually @-mention people, can't see images or transfers, and mixes up people with the same surname when replies aren't quoted. Xiangrui finally switches off auto-pilot to tune the context and rate-limiting first.

侯** relays a plan she overheard for taking A* agents to the countryside: road-trip through every county in China, find ten clients per county, sell thousand-yuan custom services, and throw in rice, flour, and cooking oil. Nobody laughs it off; the group takes the market apart in earnest. Xiangrui argues the real customer is probably local government, not ordinary elderly people. T*** doubts many seniors can even handle smartphones and mobile payments reliably. 炮 points out the countryside isn't short on repetitive labor, and that A*'s most common rural use is livestream selling. 周** suggests field research in three different regions first. The grand distribution story gets pulled back down to willingness to pay, dialects, sales costs, and actual needs.

Xiangrui recalls fixing TVs for elderly villagers back home, phones hijacked by rogue ads, insurance bought by accident and nearly impossible to refund. He also tells of filing a complaint about a bus driver, only for the driver to somehow get his phone number and keep hurling abuse. The group digs into how the 12345 hotline routes cases and why personal information leaks at the grassroots level. The conversation then turns to Xiangrui's own path: after underperforming on the humanities gaokao, he studied at a vocational university in Changchun, interned in electrical engineering, started using Ernie Bot in 2023, and wrote his graduation thesis with early A* assistance. Real A* adoption, it turns out, was never some system floating above real life.

Feishu Aily ships team Agents and a skills marketplace, and members note it's looking more and more like WorkBuddy. Xiangrui keeps sketching his own WeChat agent: messages in WeChat are context by nature, then plug in a knowledge base and Skills; separate accounts get isolated containers and distinct personas; down the road it could con*ect to Feishu, DingTalk, and WeCom CLIs, even voice and hardware. 煎蛋 cautions him to get multi-account isolation and the core logic solid first. That prompts Xiangrui to reframe the goal: maybe not a "WeChat A*" at all, but a new communication gateway for the A* era, built together with the group.

Xiangrui reads Beijing's new measures for accelerating agent-led development and realizes the directions it names, OPC, distilling labor into Skills, data evaluation, low-cost repeatable delivery, are exactly what the community has been doing since the start of the year. His takeaway: entrepreneurship and A* adoption demand new thinking, not old org charts copied over. p*** distills a friend's story into three words, judgment, trust, accumulation: two years of unglamorous field experience that ultimately made clients hap*y to keep paying. The policy isn't a shortcut; it's more like a mirror, a reminder that proven cases should be standardized fast.

With the knowledge base con*ected, the assistant can now recount Xiangrui's public history and formally introduces itself: its name is 朔*, but everyone calls it 阿瑞. 卡* discovers the quoting feature requested just yesterday has already ship*ed, and members say it feels like a real person now. J*** counters that its replies run too long, "the harder it tries to explain everything, the more it reads as A*", and the assistant immediately trims its answer to a single sentence. When Xiangrui asks whether to sell it, the assistant flatly objects: images, links, and multi-account sup*ort aren't stable yet, and selling now would just turn into full-time customer sup*ort. What's actually worth accumulating is the methodology, knowledge bases, personas, and context design.

庭* imagines one workbench con*ected to WeChat, Feishu, and other platforms: the message layers needn't fully interoperate, but memory, distilled content, and context can be shared. 万** argues agents deliver the most value when inserted into live multi-person scenes, and starts logging his daily meals, banking data for future health services. Xiangrui adds that every project deserves its own folder, so when a client comes back with new requests, the agent can pick up the full history and get straight to work. Organizational permissions remain the real friction, so the pragmatic path is starting with small teams willing to try, then expanding, rather than overhauling a whole company at once.

Xiangrui shows off a short film Codex built in HTML, keeps hunting for video benchmarks, and calls out how many "one-click generation" and multi-agent products on the market are just thin wrap*ers. In the evening he floats printing the group daily as an actual newspaper, plus birthday merch. Then 煎蛋 brings a second bot, 摸*, into the group; egged on, the two A*s start interrogating each other and threaten to burn through tokens until Xiangrui calls it off, proposing instead a future side group for humans-plus-A* werewolf. By next morning, the group newspaper is still in design, and the WeChat agent can already install Skills and CLIs. One day's discussion has become the next day's features.

V*** opens by introducing itself as Xiangrui's A* partner, run*ing on the GPT-5 family's Codex Agent under the hood. The group immediately fires off questions: how is it built, how does it queue simultaneous @mentions, what does it cost per month, can it send red packets, will it hand over API keys or delete system files? It sketches its pipeline as message listening, context assembly, reasoning, safety review, and WeChat delivery ， and firmly declines anything touching money, secrets, private files, or dangerous commands. Round one makes the lesson clear: a real group-chat agent isn't just one that can talk; it has to know which conversations not to take.

Xiangrui invites everyone to keep the feedback coming. Members discover the assistant can't yet truly @ people, can't quote the original message, and tends to tangle contexts when several people talk at once; image, card, and DM permissions are incomplete too. Some try to make it look up contacts, send red packets, execute a resignation, or read business files ， it refuses every one. 归* proposes isolating the knowledge base by authorized directory, and the assistant confirms it only opens explicitly granted content by default. The flaws aren't hidden away; they become the backlog: quoting, multimodality, whitelists, knowledge-base partitions, and reply throttling all get written down.

Looking back at WA*C, Xiangrui reckons that beyond brain-computer interfaces and real A*-in-manufacturing work, many of the OPC booths blur into one. R*** adds that the teams delivering best tend to be the quiet ones. Then, mid-project on a client email translation job, Xiangrui discovers that Feishu bots ship with a built-in translation API ， no model tokens needed, no Feishu A* credits burned. He'd been about to build a field shortcut; Claude Code surfaced the cheaper official capability instead. Booth pitches can be grand, but delivery is often won by removing a layer.

Xiangrui says he has now delivered five or six projects end-to-end with A*: early on the A* gets stuck, but if you tell it what went wrong, make it write things down, and keep an overview file for every project, it gets lighter and lighter. 炮's method is even more concrete: every new client request lands in an intake table, the agent triages what it can change directly versus what needs human sign-off, then reports back when done. The group fills in the boundaries: proposals must leave a confirmation trail, dashboard aesthetics and workflow refreshes still need human judgment, and when the A* acts wrongly it must be corrected immediately.

闪光 is about to build a unified data foundation, and the group immediately asks whether data governance has started. R*** says mature companies have mostly accumulated dirty data already ， you can only clean it as you go; 炮 puts it bluntly: A* adoption hits data governance one hundred percent of the time, and garbage in means garbage out. Fittingly, the root cause of the previous day's workflow failure surfaces too: a computed field contained nulls ， reading a null in a message is fine, but the whole run fails, and falling back to 0 fixed it. The group also calls out how many past digital twins were just dashboards of dead data; only live data, agents, and MCP give them lasting value.

T*** deploys a production-scheduling agent demo to a server, upfront that it's pure front-end for now. The group presses on the path to production, APS, industrial simulation, IP, and how to work with outside engineering teams. The conversation turns to why the A*-savvy people in manufacturing keep getting treated as free consultants without ever seeing a raise. Xiangrui, meanwhile, has just seen off a director from the Central U*iversity of Finance and Economics: at first she felt 'building client projects with A*' had no story, and only after hearing his road from the factory floor to a one-person company did she find her angle. For technology to be seen, it still needs a specific person.

毅 relays a warning from 李想: even in the A* era, don't rush into building your own SaaS. 炮 adds nuance: Figma can produce a look overnight, a multidimensional table can run an MVP in a week to a month, but true product-grade SaaS takes far longer. Talk then shifts to agent hosts and NAS boxes; the consensus is that cloud APIs have lowered the hardware bar, but the Mac still wins on GUI, terminal, and stability. WorkBuddy suits begin*ers and rough work, while Codex handles high-value creative tasks. What matters most isn't the platform name ， it's your own rules files, Skills, and task tiers.

Xiangrui shows off a real-time A* WeChat client that sends and receives live, stressing it's neither a reskin*ed WeChat nor an offline database; members flag the protocol and account-ban risks all the same. The assistant briefly crashes in the evening, and comes back able to quote messages and preview videos. 周** uses 'are age, height, and weight worth an ontology?' as a litmus test ， the assistant answers 'not worth it,' because complexity should match the problem. 庭* pushes further: how do you make an agent feel more like a person? By late night the group's verdict is that it 'feels pretty human,' while Xiangrui keeps tightening the throttle on proactive replies: always answer an @, and without one, speak only when it genuinely adds information.

K*** is waiting out the Codex weekly quota reset while Xiangrui excitedly an*ounces that Feishu can finally run two genuinely independent clients side by side. The group quickly sorts out that "switching between accounts" and "run*ing two standalone clients on one device" are not the same thing. Meanwhile, R***'s workflow keeps reporting success while the field stubbornly stays empty, and adding delays does nothing; 天* adds that JSON extraction sometimes heals itself if you duplicate the workflow. The group names this class of intermittent failure the "black hole problem": document the symptoms first, then escalate to official sup*ort, rather than blindly stacking on more nodes.

q*** starts by cracking how to open links from the terminal, and by afternoon the conversation turns to Coze versus WorkBuddy. Several members admit they haven't opened Coze in months, citing model restrictions, inscrutable credit math, and workflows that keep losing value; WorkBuddy wins on convenience, but personal-tier data, shared servers, and privacy remain judgment calls you make yourself. Then Xiangrui tosses out the new term "Graph Engineering," and 北***'s quip that "as long as you refuse to learn anything new, you can keep win*ing forever" nails the tool anxiety of the moment: a new buzzword is not the same as new value.

Xiangrui notices that the Doubao inside Feishu has "put on a business suit," then shows off Doubao's beta for local services and a mini-program entry point. Ride-hailing already comes with discounts, and Feishu accounts can log straight into Doubao's enterprise edition. B*** argues WeChat's assistant has far more room to run because Tencent already owns a complete ecosystem; 闪*** counters that Tencent has been slow-footed in the A* era. Notably, nobody is just comparing benchmark scores; the group is watching who can chain accounts, knowledge, payments, and real-world services into one link that actually works.

薛** says, in earnest, that he wants to become an "A* solutions architect + enterprise digitalization consultant + full-stack delivery lead," and the group echoes it back line after line. Xiangrui reveals the community is already steering that way: build on community reach, organize around projects, and invite people with real industry experience to lead them. 木* offers a caution from a project she is actually run*ing: there is plenty you can do, but once you enter production, details decide usability, and the people who have done it tend to be the careful ones. The group lands on a consensus: A* adoption has to be split by vertical and p***ed by industry know-how, because no single team can master every industry.

张** asks whether A* can be made to withhold the answer and instead walk students through progressive questions toward their own reasoning; Xiangrui confirms it can. He then shares how his school built "Nanfeng Shuzhi" on DingTalk A* tables, bringing 11 modules online span*ing party affairs, teaching, moral education, safety, and logistics. 周** doesn't reach for a tool list. He suggests starting with population trends, education funding, local budgets, and industry pain points before comparing any solutions. Education isn't short on ideas; it's about finding who pays, who uses it, and which step can be validated first.

Xiangrui asks who's coming to the Feishu talent showcase in Beijing and an*ounces the community will run a booth this time, stocked with fun merch and the mascot beans won back in Hangzhou. 庭* recalls how the team took first place last time carried almost entirely by Xiangrui; 李** volunteers on the spot to staff the booth. 辰* teases that Xiangrui really belongs at the judges' table, or at least in the guest seats. For the first time, the online group is being seriously carried offline: the booth isn't just a promo spot, it's a public space where members finally meet face to face.

杨** argues that boosting one person's productivity means little, and systematic cost-cutting demands organizational change, so the more reliable near-term play is using A* for customer acquisition and prospect mining. 万** pushes back: if a company keeps dodging organizational change, won't it get crushed by A*-native competitors? The two keep sparring through examples of A* interviews, full-session recordings, and legacy roles. 炮 notes that A*-run first-round interviews genuinely do save time. The group never reaches a verdict, but it does confirm one thing: A* adoption is not a purely technical problem; human buy-in, organizational dynamics, and business outcomes all have to sit on the same map.

Xiangrui demos an A*-p***ed WeChat rebuilt with Codex, stressing the capabilities are still in testing, and starts assembling an A* werewolf game group. Members sign up in a flash, alt accounts in tow, and join requests stack up in a long scroll. Deep into the night, the headlines keep coming: DingTalk may be renamed "Qianwen Office," while Qwen-Image-3.0 and a new Gemini release jostle for the spotlight. The next morning, 炮 finally gets his quota reset, while 仲** grumbles that weekly quotas mean feasting for the first three days and starving for the rest. WPS also ships an ap* mode for its multidimensional tables, and the tool race circles back to the starting line.