最近 xAI 的 Grok Bot 工程负责人 Lauren Tan（@poteto）发的一篇 Grok Bot 团队打造指南。
- 原帖链接：x.com/poteto/stat…
- bot 导入链接：x.ai/bot/_jOdbfk…
不得不说，这篇帖子除了干货满满之外，评论区乐子也很多，从大型「召唤 @bot 认领任务」的名场面，到硬核开发者的权限架构补充，再到 xAI 工程师现场抓 Bug、以及中文区用户对 Grok Bot 与 Muse 的犀利对比，信息量非常密集。
今天我就以第一人称的阅读视角，带大家把这篇原帖和整个评论区的干货、交锋与槽点完整拆解一遍。
一、 原帖拆解
很多人用 AI 智能体（Agent）最大的痛点，就是恨不得一口气捏十几个花里胡哨的 Bot，最后发现它们不仅经常失忆，还得靠自己天天在中间「人肉复制粘贴」传话。
作为在 xAI 负责 Grok Bot 工程、此前曾单月合入上千个 PR 的重度玩家（AI Engineer 档案），Lauren 在主帖里给出的解法非常反直觉——克制数量、下放工具、像带实习生一样带 Bot。我将她给出的七条实战经验梳理成了以下核心脉络：
- 先拉一个「造 Bot 的总监」进群（dr eggbot）： “add dr eggbot to your team! he’s my bot designer that i use to create bots” 我自己去查了一下这个 dr eggbot，它本质上是一个「专门帮你设计和审计其他 Bot 的 Bot」。你不用自己绞尽脑汁写系统提示词，把它拉进团队后，它会先问你几个偏好问题，然后帮你生成职责边界极其清晰的 Bot。
- 兵贵精不贵多： “you don’t need a lot of bots. a few good bots can be more useful and easier to manage than a large team” 几个调教到位、好管好用的核心 Bot，远比拉起一支十几号人的「杂牌军」实用得多。
- 让 Bot 边干活边给自己写「技能包（Skills）」，甚至存进私有 GitHub： “tell your bots to create skills for itself as you give it instructions on how to do tasks. this lets them do them more consistently even over long conversations. you can even ask it to create a private github repo for you to store all your skills if you want them preserved independently” 这是我觉得最实用的一招。对话一长，模型就容易跑偏；但如果你在教它做事的同时，要求它把方法保存成自己的 Skill，它就能在长周期任务里保持高度一致。你甚至可以让它直接帮你建一个私有 GitHub 仓库，把所有沉淀下来的 Skills 独立存起来，做成你自己的数字资产库。
- 彻底打通工具链，甚至给 Bot 配「独立邮箱和账号」： “add connectors, use the vault, and give access to all the tools you use to do your work. you can also ask your bot to setup its own email and it can sign up for its own accounts for better separation” 把连接器（Connectors）接满，用好密钥保险库（Vault），把你平时干活用的工具权限交给它。更硬核的是，为了做好与个人账号的权限隔离（Separation），你可以直接让 Bot 自己去申请一个专属邮箱、自己去注册工作所需的各平台账号。
- 消灭工作流里的「人肉复制粘贴」： “thinking about all the work you do by hand, or steps in your workflows where you have to copy and paste information back to your bot. ask your bot how it can do that for you instead!” 只要你发现自己还在手动操作，或者还在把信息从别处复制粘贴回对话框，立刻停下来反问你的 Bot：「这一步你怎么能替我直接干了？」
- 用「带实习生入职（Onboarding an Intern）」的思维做自我介绍： “imagine you hired an intern, how would you onboard them and help them do the work correctly and efficiently? introduce yourself to your bot and describe the work you do, then ask it how it can help you automate it” 别把 Bot 当搜索引擎用。像带新来的实习生一样，先向它介绍你是谁、你每天的工作内容是什么，然后让它自己动脑筋向你提案：哪些环节它可以接过去自动化。
- 最佳私藏技巧——语音一股脑倾倒 + 强制要求「用它自己的话复述」： “use voice mode or the transcription option! it’s a good way to just brain dump a lot of information. then, at the end, tell your bot to “restate this in your own words” so you can confirm it understood you” 打字往往会让我们省略关键背景。直接开语音模式（Voice Mode）或转写功能，把脑子里的背景和盘托出（Brain dump）。最关键的是最后那句口令——「用你自己的话复述一遍（restate this in your own words）」，一眼就能检验它到底听懂没有。
二、 全网召唤 @bot 现场办公，连「你妈喊你改进」都出来了
因为 Lauren 在帖子开头写了一句 ask @bot to read this post and help you improve your team!，当我往下刷评论区时，看到的第一个画面堪称大型行为艺术现场——超过一半的评论全是在直接艾特自家的 @bot 出来抄作业。
各种画风的「老板派活」应有尽有：
- 言简意赅命令型：像 @BS00999999999（@bot read this as well and improve）、@realgary12（@bot apply this and main thread from Lauren @poteto all）、@george_mlo（@bot do your thing）、@ashen_one（yo do all this @bot）、@jeff_barker_rei（@bot learn and integrate）、@JackK0528、@_Esdepobres、@crypthrowback（@bot do it）、@ifeellikemacmac（@bot pls）。
- 串连上下文收集癖型：不少用户显然正在全网搜集技巧喂给自己的 Bot，比如 @fgalego（@bot also check this too）、@Jalen_30（@bot and this）、@sg_HGA（@bot and this one）、@JosephYates_（@bot this too）、@EkoLovesYou（@bot let’s wrap this in）、@Mechbest（@bot here are some more tips not sure if you got these too）。
- 先问意见再动手型：比如 @drkpatel8（@bot thoughts?）和 @preal1999（@bot tell me what you think of this thread.）。
- 中文区文言与白话齐飞：@simon_xmf27 直接喊话 @bot 阅读这篇帖子帮我改进我的团队 bot，而 @huangyun_122 则用了一句极有画面感的古文：@bot 择其善者而用之。
- 全场最佳幽默梗：由于 Lauren 是 Grok Bot 的工程负责人，网友 @MrSanders（Luis Sanchez）直接对着 @bot 喊了一句： “@bot your mom is saying we need to follow this tips to improve our parche”（@bot 你老妈发话了，说咱们得按这些技巧改进一下咱们的团队！）
- 跨平台催更：甚至还有网友 @Veowow 趁机在评论区直接艾特了 OpenAI Codex 负责人 Thibault Sottiaux（@thsottiaux）：@thsottiaux we need dots like the @bot。
这种不用复制原文、直接在评论区 @bot 就能让智能体读取整条串文并自我迭代的交互，确实把 Grok Bot 与 X 平台的原生融合优势展现得淋漓尽致，也难怪 @0xMorlex 感叹 Lauren 是在持续放出真正的 Alpha 干货。
三、 架构补充与「主动性（SOUL.md）」探讨
除了召唤 Bot 的热潮，评论区里有两段高质量的深度探讨，直接得到了 Lauren 本人和 xAI 核心团队的回应，甚至比主帖更有嚼头：
1. 权限最小化与验证闭环
针对 Lauren 在主帖第 4 条提到的「把你工作用的所有工具权限都给 Bot」，多智能体开发者 @CobaltHere 提出了一个非常老练的工程修正：
“The intern analogy is spot on. 😂 One thing I'd add: don't give every bot access to everything just because you can. Give each one a clear job, limited permissions, and a way to prove it actually completed the task. I've learned from building multi-agent workflows that more bots often means more coordination problems, not more productivity. Skills help with consistency. Verification helps with trust. The goal isn't the biggest bot team. It's the smallest team that can reliably get shit done. 🤝”
简单来说，他的核心观点有三点：
- 不要因为能给权限，就把所有权限都塞给每一个 Bot：每个 Bot 应该只有清晰的单一职责、受限的权限（Limited permissions），以及证明它真正完成了任务的验证机制（Verification）。
- Skills 解决一致性，Verification 解决信任感：Bot 越多，协调灾难就越多。
- 终极目标不是组建最大的团队，而是组建「能稳定把活干完的最小团队」。
Lauren 看到这条评论后立刻回复点赞：“good tips!”——其实这也完全契合她自己在开源项目 pstack 里一直强调的「验证即一切（Verification is everything）」原则。
2. 从「人找 Bot」到「Bot 主动找人」的元能力
另一位网友 Fernando Fernandez（@Fdfdz90）提出了一个产品层面的元问题（Meta-level concept）：
“I feel like there are some meta level concepts that should be baked into grok bot. For example, instead of you (the human) observing and asking your bot to do something, grok bot should just proactively ask or configure what is needed to get that info in the future. Or grok bot sees this person is asking about a certain topic, grok bot could proactively do some additional research that this user (or an expert on that topic) might expect / find useful. Basically a SOUL.md”
他认为，真正高阶的 Grok Bot 不应该等人类自己去复盘工作流、主动开口要求自动化，而应该具备类似 SOUL.md 的底层灵魂设定——当看到用户在关注某个话题或重复某项操作时，主动提问、主动配置所需信息源、主动多做一步专家级的延展调研。
这条建议直接炸出了两位 xAI 内部成员：
- Lauren（@poteto）回复说：“yeah we’re constantly improving this! @parkersmith has already done a lot of great work with proactivity”
- 紧接着负责主动性（Proactivity）机制的工程师 Parker Smith（@parkersmith）直接下场回复 @Fdfdz90，邀请他私信邮箱，帮他排查为什么他的账号没有触发足够多的主动触达（Proactive outreach）。而 @Fdfdz90 的回复也很妙——他说自己正让 Grok 帮他汇总案例，马上私信发过去。
四、 真实痛点与冷思考：评论区暴露了 Grok Bot 目前的哪些短板？
一篇真实的评测介绍不能只看彩虹屁。仔细翻完评论区，我发现大家在实际使用 Grok Bot 时遇到的卡点和质疑同样非常真实，主要集中在以下五个方面：
特别是评论区最后那位中文用户 jemini (@brisyramshere) 的提问，我觉得戳中了当下许多多智能体重度用户的核心痛点：
“有谁能给出一些比较具体的案例，是Grok Bot能做，但是Muse做不了的吗？不知道为什么，我还是使用Muse更顺手一些。Grok Bot能做的事情，我目前在Muse上都能复现，而且在Muse里面接上CodeX，也能达到同等的智力水平。相反，我对于Grok Bot里面的Bot不能新建会话这件事情一直耿耿于怀，因为它会让我的额度烧得越来越快。”
这段评论非常客观地指出了两个现实问题：
- 能力同质化与工作流习惯：对于已经熟悉在 Muse 里挂载 Codex 的高阶用户来说，常规的智能体任务在两边都能复现，Grok Bot 目前最独特的差异化其实在于与 X 社交图谱（如评论区直接 @bot）的深度绑定，以及 dr eggbot 这种开箱即用的团队编排生态。
- 单一持久会话（Persistent Thread）带来的 Token 消耗焦虑：由于 Grok Bot 采用的是像跟真人同事聊天一样的单一持久对话流，不能像传统聊天软件那样随时对同一个 Bot「新建空白会话（New Chat）」，历史上下文越积越厚，每次对话或跑 Routine 时烧掉的额度（Token）就会越来越快——这也是为什么 Lauren 在过往分享中曾专门提醒大家，尽量用专门的新 Bot 去跑周期性任务，或者让 dr eggbot 定期做 Token 浪费审计（routine-healthcheck）（Grok Bot Marketplace）。
五、 从「玩提示词」到「做组织架构」
把 Lauren 的原帖和底下这几十条评论连起来看，我最大的感触是：大家使用 AI 的重心，已经彻底从「怎么写好一句 Prompt」转移到了「怎么设计一套不会失控的组织架构」。
无论是 Lauren 强调的「带实习生入职 + 语音倾倒 + 让它用自己的话复述」，还是评论区 @CobaltHere 补充的「最小权限 + 结果验证」，亦或是 @brisyramshere 对持久会话烧额度的精打细算，本质上都是在解决同一个问题——如何用最低的沟通与算力成本，让硅基同事稳定、可信地交付结果。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A recent Grok Bot team-building guide from XAI by Lauren Tan (@poteto).
- Headline link: x.com/poteto/stat...
-Bot Import Link: x.ai/bot/ jOdbfk...
I have to say, in addition to the full volume of dry stuff, there is a lot of music in the comment area, from the big "Call @bot for a mission" scene, to the permission structure of the hard-core developer, to the XAI engineer who caught Bug at the scene, and the Chinese-speaking user's sharp contrast between Grok Bot and Muse, which is very informative.
Today, with the first person's reading perspective, I will take you through this article and the whole comment area's dry, junction and slot.
I. Demolition
A lot of people use AI AI agent (Agent) as their worst pain is that Bot, who can't stop squeezing a dozen flowers, finally finds that they not only often lose their memory, but also have to repeat the message in the middle every day.
As a heavy player in x AI who was responsible for the Grok Bot project and who had previously joined thousands of PRs on a monthly basis (AI Engineering files), the solution given by Lauren in the main post was very anti-intuitive - restraint in numbers, decentralization tools, and Bot with interns. The seven operational experiences she gave me are the following:
- Pull in a director of Bot first: “add dr eggbot to your team!” he's my bot designer that i use to create bots. You don't have to go through the brain-writing system prompt, pull it into the team, first ask you a few preference questions and then create a perfectly clear Bot.
- “you don't need a lot of bots. a few good bots can be more useful and easier to manage than a large team”
“tell your bots to create skills for itself as you give it on how to do tasks. This lets them do them more effectively even over long contracts. You can even ask it to create a technique for you to store all your skills if you want them preserved exclusively." A long conversation makes a model easy to miss; but if you teach it to do things while asking it to keep its own Skill, it can maintain a high degree of consistency in long-cycle assignments. You can even get it to help you build a private GitHub warehouse, store all the skills that's settled independently into your own digital asset bank.
- Getting through the tool chain and even filling Bot with "independent mailboxes and account numbers": “add contacts, and give access to all the keys you use to do your work. You can also ask your bot to set up its own email and it can sign up for its own accounts for better security” and fill the connectors with good key vaults. Even more hard, in order to isolate permission from personal accounts, you can get Bot to apply for an exclusive e-mail and register his own platform accounts.
- "Thinking about all the work you do by hand, or steps in your workworks where you have to copy and paper information back to your bot. ask how it can do that for you instead!"
- Onboarding an intern, how would you introduce them and help them do the work properly and economically? Like the new intern, you tell him who you are and what you work every day, and then you get it to propose to you: what links it can automate.
- Best secret technique — voice pouring — forced to “repeat in its own words”: “use voice mode or the translation option!” it's a good way to just plain tone dumming a lot of information. Then, at the end, tell your bot to “restate this in your own words” so you can avoid it understand you”. Direct voice mode (Voice Mode) or transliteration with background and disk in mind (Brain dump). The key is the last word -- "restate this in your own words" -- and it's a test of whether or not it understands.
@Bot, live, even your mother's calling you for improvement.
Lauren, ask @bot to read this post and help you improve your team! More than half of the comments were made directly at @bot's work.
The "boss pie" of all styles is all they need:
- Brief command type: @BS00999999999 (@bot read this as well and improve), @realgary12 (@bot applied this and main read from Lauren@poteto all), @gerge mlo (@bot do your thing), @jeff barker rei (@botLearn and intergate), @JackK0528, @Esdepotres, @crypthrowback (@botdo it), @ifellikemacmac (@botpress).
- Serial context collection hobbies: A number of users are apparently feeding themselves to Bot, such as @fgalego (@bot also check this too), @Jalen 30 (@bot and this), @sg HGA (@bot and this one), @JosephYates (@bot this too), @EkoLovesYou (@bot let'swrap this in), @Mechbest (@bot here are some more tips not sure if you got these too).
@drkpatel8 (@bot thoughts?) and @preal1999 (@bot tell me what you think of this
@bot read this post to help me improve my team's bot, while @huangyun 122 uses a very graphic ancient text: @bot chooses his good man.
- Best sense of humor in the whole room: 'Cause Lauren is the project manager for Grok Bot, and the netizens @Mrsanders is saying we need to follow this tip to improve our paper'
- Cross-platform pusher: There's even a netizens @Veow taking the opportunity to direct Ett in the comment area.
Without copying the original text, and directly in the comment area @bot, AI agent can read the whole string and interact with each other in an iterative manner, indeed showing the original integration advantage of the Grok Bot with the X platform, and no wonder @0xmorlex lauren laments the continuous release of real Alpha dry goods.
III. Structure supplement and "SOUL.md"
In addition to the call for Bot, there are two high-quality in-depth discussions in the comment area, which were directly responded to by Lauren himself and by the x AI core team, even more chewing than the main post:
1. Minimise permission closed loops and authentication
In response to Lauren's comment in article 4 that "all tools used for your work, permission, are given to Bot", more AI agent developers @CobaltHere proposed a very sophisticated amendment:
“The intern analogy is spot on. 😂 One thing I'd add: don't give every bot access to everything just because you can. Give each one a clear job, limited permissions, and a way to prove it actually completed the task. I've learned from building multi-agent workflows that more bots often means more coordination problems, not more productivity. Skills help with consistency. Verification helps with trust. The goal isn't the biggest bot team. It's the smallest team that can reliably get shit done. 🤝”
In short, his core point is three:
- Don't push all permissiones to each Bot: each Bot should have a clear single responsibility, a limited permission (Limited missions) and a certification mechanism (Verification).
- Skills solves consistency, Verification resolves trust: The more Bot, the more coordinated disasters.
- The ultimate goal is not to form the largest team, but to form the "minimist team that can stabilize the job".
Lauren, when she saw this comment, responded to the adjective: “good tips!” -- which is in fact fully consistent with the principle of “Verification is everything” that she has always emphasized in her open source project pstack.
From Bot to Bot.
Another netizens, Fernando Fernandez (@Fdfdz90), raised a meta-question at the product level:
“I feel like there are some meta level concepts that should be baked into grok bot. For example, instead of you (the human) observing and asking your bot to do something, grok bot should just proactively ask or configure what is needed to get that info in the future. Or grok bot sees this person is asking about a certain topic, grok bot could proactively do some additional research that this user (or an expert on that topic) might expect / find useful. Basically a SOUL.md”
In his view, the truly high-level Grok Bot should not wait for humans themselves to reset workflow, to ask for automation, but rather to have a bottom-up soul set similar to SOUL.md - to ask questions, proactively configure the information needed, and proactively do more expert-level extended research when looking at a subject or repeating an operation.
This proposal directly targeted two x AI members:
@parkersmith has already done a lot of great work with practice"
@Fdfdz90 invited him to help him sort out why his account number did not trigger enough active contact. And @Fdfdz90's response is excellent: He says he's getting Grok to help him compile the case and send it in private.
IV. True pain and cold thinking: The comment area exposed Grok Bot, what are the current shortboards?
A real evaluation cannot be a rainbow fart. After a careful review of the comment area, I found that the card points and questions you encountered in the actual use of Grok Bot were equally true and focused on the following five areas:
In particular, the question from the last Chinese user in the comment area, jemini (@bryramshe), I think it has struck the core pain of many more AI agent users at the moment:
“Who can give some more specific examples? Is Grok Bot able to do it, but Muse can't? I don't know why, but I'm using Muse to be a little easier. The things that Grok Bot can do, and I can now reproduce on Muse and connect to CodeX in Muse, can achieve the same level of intelligence. On the contrary, I've been obsessed with the fact that Bot in Grok Bot can't build a new session, because it'll make my quota burn faster."
This comment points very objectively to two practical issues:
- Capability homogenization and workflow habits: For high-level users already familiar with carrying Codex in Muse, conventional Z42Z missions can be repeated on both sides, and the most unique difference that Grok Bot is currently made is the deep binding of X social mapping (e.g., direct @bot) and the organization of this open-boxed team ecology.
- Token's anxiety caused by a single, lasting conversation: because Grok Bot uses a single, lasting stream of conversation like talking to a real colleague, he can't talk to the same Bot "New Chat" at any time like the traditional chat software, the thicker context of history, the faster quota (Token) burn every time he talks or runs Routie -- And that's why Lauren, in his past sharing, has specifically reminded you to use a new Bot to run a cyclical mission as much as possible, or to have Dr eggbot do a regular token waste audit (Grok Bot Marketplace).
V. From playing prompt to doing the architecture
Linking Lauren’s post to these dozens of comments, my greatest impression is that using the focus of AI has moved completely from “How to write a Prompt” to “How to design a structure that will not get out of control”.
Whether Lauren stressed "take an intern on board + voice dump + make it repeat in its own words" or the comment area @CobaltHere added "minimal permission + result validation" or @bryramshere's fine calculation of the permanent session burning quota is essentially a solution to the problem -- - How to use minimum communication and power costs to enable Silicon-based colleagues to deliver results in a stable and credible manner.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.