高强度用了两周 Grok Bot 之后，我发现这玩意是真的会上瘾。
现在基本上走路在看，喝水在看，连蹲坑的时候掏出手机也在刷。
说实话，早在八月份它刚发布的时候我就知道了。
那时候到处都是铺天盖地的“保姆级教程”、“十个必学神仙技巧”，我觉得是不是又是新一轮的 AI 热点炒作。直到十一假期我闲下来认真用了几天，才发现是真香。
不用挂机，它自带一台 24 小时云主机
用过 Codex 或者 Claude Code 的远程，应该都懂那种难受：家里或者工位上的电脑得 24 小时开着客户端，更难受的是出门用手机远程连，经常因为网络断掉而连不上。折腾过 OpenClaw、Hermes 的人也知道，光是装环境和依赖就能劝退一大批人。
Grok Bot 最舒服的地方就在于开箱即用，并且把这台电脑直接搬到了云端。哪怕你把笔记本盖子一合、断网走人，它自己在云端那台机器上该跑的任务照样跑。
而且作为一台真实的云电脑，它跟普通工具有几个非常本质的区别：
第一，硬盘是持久化的。你在对话框里让它写了个爬虫脚本、下载了一堆数据，关掉浏览器之后，这些东西依然稳稳当当存在它的硬盘里，下次随时能调出来。
第二，所有 Bot 共享同一台电脑。你可以建好几个不同分工的 Bot，它们共用同一个文件系统、同一个网络出口、甚至共享浏览器的登录状态。比如 A Bot 下载好了一个数据包，B Bot 根本不需要你手动下载再传给它，直接在本地路径就能读。每个 Bot 也有自己独立的桌面和浏览器窗口，互不干扰。
最绝的是，你甚至可以直接点开它的桌面，就像 TeamViewer 远程桌面一样。它开着 Chrome 到处查资料、打开终端敲代码的过程全在屏幕上。你真的可以直接在上面用鼠标点击操作，遇到它卡住的地方，顺手帮它点个确认，或者帮它过个验证码。
免不了遇到网络问题
用云电脑干活，迟早会撞上网络问题。
发现这个坑，是因为当时我想在它的终端里装 Gemini CLI，结果命令刚跑完就直接报错：User location is not supported（当前地区不受支持）。
我当时就纳闷了：不是挂在美区的云电脑吗，怎么还会报地区不支持？
出于好奇，我经过测试和研究，才发现它的机制很有意思：这台云电脑默认是通过**「共享出口池」**上网的，所有用户共用，每次请求给出的 IP 都不一样。一会儿是 Cloudflare 的这个节点，一会儿跳到那个节点，有时候出口 IP 甚至会飘到国内，导致某些对地区敏感的工具直接报错。
如果你日常也要跑类似对网络环境要求苛刻的任务，其实官方留了一个很实用的开关：
在控制台找到：设置 → 电脑 → 你的电脑，勾选 经此电脑转发流量。
这样一来，云电脑的所有网络请求就会借用你本机的网络环境代理出去，网络环境就彻底稳定了。
别看保姆教程，新建个 Bot 直接盘
Grok Bot 的入门我认为最快的方式：新建一个 Bot，直接跟它对话。你想知道什么基础功能、进阶玩法，直接在对话框里问它，它比任何教程都讲得清楚。
很多人习惯了以前 OpenClaw、Hermes 的玩法，觉得要给 Bot 设定复杂的角色、写一套复杂的 System Prompt。
在这里，只要用大白话跟它讲清楚你的诉求就行。现在的模型已经不需要刻意做复杂的 Prompt Engineering 了，它会在交互中自己调整。
真正让它具备生产力的，是「Connect Apps」（连接应用）功能。
我自己日常必连的几个生态：
- GitHub：让它能直接读你的项目仓库、拉代码、提交 PR。
- X (Twitter) ：作为实时信息的抓手，查资料、盯动态很方便。
- Gmail 和 Notion：一个当成它的收发信箱，一个当成它的工作记录空间。
平时很多你天天在重复、又不得不花精力去看的累活，都可以直接丢给它。比如：
- 信息雷达：不用自己天天刷推，让它每天定时抓取 X 上特定圈子的最新内容，过滤掉水帖和广告，总结成要点同步到 Notion
- 仓库盯梢：让它盯某几个核心关注的开源项目，或者自己仓库的新 Issue 和 PR，每天提炼出关键动态发到邮箱，不用天天手动去刷 GitHub。
- 邮件分拣：每天自动扫描一次邮箱，过滤掉无用邮件，挑出服务器告警或者关键项目动态，整理成待办丢到 Notion 里。 整个过程只需要用大白话跟它讲清楚规则，后台自己就默默把这套定时任务跑通了。
最有意思的功能：群聊
群聊，我认为是 Grok Bot 最有意思的一个功能。
你可以把几个不同角色的 Bot 拉进同一个群，任务往群里一丢，它们会自己接单、拆解任务，然后安排不同的 Bot 去处理。你也可以像在微信群里一样，随时 @ 指定的 Bot 对话。
最开始我觉得特别有意思的是发个话题让他们自己讨论。比如我有个取名的选择困难症，扔到群里，几个 Bot 就会自己相互讨论、提出理由，甚至开始投票，情绪价值很到位。
但玩了几天，开始派一些复杂活的时候，就会发现一些问题：
首先是上下文很容易变乱。几个 Bot 你一句我一句，消息刷得飞快，互相客套复读，正事反而推进得很慢。
另一个问题是，原本以为任务发到群里它们就会自动流水线作业、自己坐等结果就行。结果发现因为人数一多，你得不停在群里 @ 这个、纠偏那个，反而把大量精力耗在了反复协调上。
多 Bot 协作其实有两种方式：
- 单点 @ 串行交接：在当前 Bot 里直接 @ 另外一个 Bot，当前任务完成后分派给下一个。好处是上下文相对干净，缺点是输出结果需要切换到对应的 Bot 才能看。
- 集中群聊：所有角色在同一个群里。全局完全透明，但极度消耗上下文。 很多人可能看过官方博客里提到的案例：“五人团队每天合并 100 多个 PR”。
但其实这不是靠一群 Bot 在群里随意发散聊出来的。核心在于需要有 1 个明确的协调者（担当项目经理角色），作为统一的入口，负责连接外部工具、把控进度、分发工单，而具体的代码修改再派发给专门的 Cloud Agent 去跑。
想要让多 Bot 协作真正干好重活，必须有一个统一的入口和协调者。
我是怎么搭协同团队的
明白这一点后，我重新梳理了流水线，搭了两个帮我分担日常工作的协同小队。
1. 研发小队 Build Lab
我给自己配置了一支研发团队，基本涵盖了日常开发的关键角色：
- Max：产品经理，帮我把脑子里模糊的想法理顺成需求
- Pixel Penny：设计，负责交互和视觉建议
- Codey：全栈开发，负责核心代码实现
- Buggsy：测试，专门负责找 bug 和质量把控
- Tracker Tina：项目经理，负责进度跟踪和任务派发
- Hype Hazel：市场增长，负责文案和上线宣传 !
以前我的做法是把大家拉到一个群里乱聊，结果效率很低。
现在的做法是设立统一的入口： 我在群里只找 Max 沟通需求。需求确认、输出 PRD 之后，由项目经理 Tina 出面拆解工单，谁做什么由 Tina 单独去安排，我不再逐个去 @。Tina 会定时在群里同步各角色的工作进度。
比如前段时间我让它们帮我优化个人站，Max 梳理完要改的板块，Tina 派活给 Codey 敲代码，敲完直接交由 Buggsy 去挑刺。在这个过程中，我只做两件事：开头确认需求方向，最后代码合并前点个批准。
当然，说实话这套流程我自己也在不断磨合，代码规范和自动化验收环节还没完全完善，远没到百分之百全自动的状态。但比起一开始的人肉协调，省心太多了。
2. 视频工作室 Video Studio
除了写代码，我另一条高频使用的流水线是做视频。
现在 Opus 5.5 在制作视频方面效果确实很强，结合 Grok Bot 的生态，效率提升很明显。但要想稳定产出一条成片，单靠一个 Bot 很容易上下文混乱。
我把视频线拆成了四个角色：
- 总导演：把控整体节奏和视觉基调
- 策划编剧：写脚本分镜和台词
- 素材：专门去找匹配的画面和参考
- 配音字幕：负责声音和文字对齐 !
一条视频从想法到成片，本身就是几个不同工种的配合。有人定方向，有人出结构，有人找素材，有人做后期。把它们独立开，上下文最干净。同样，也是由一个协调角色来串联前后环节。
下面这几个小视频，就是我用这套流程尝试全自动跑出来的样本：
目前成片的效果肯定谈不上多惊艳，细节上也远没有人工精剪那么考究，不少画面和镜头衔接依然能挑出毛病，我自己也还在继续调优。
但它对我最大的价值，是把“从一个想法到看见成片”的折腾成本打下来了。以前想做一条视频，从构思文案、满世界找素材、再到卡着时间点对音轨，一套下来大半天就没了；现在初稿和基础框架它基本能帮我跑个七七八八，我只需要在成片的基础上，手动去挑刺、补两处关键细节，就已经能省掉大半的精力。
我把这两个配置好的团队和群聊做成了公开模版，如果你想直接体验，可以直接导入： edisonz.cn/grok-bot-templates
如果你想看更多别人分享的玩法，也可以去这几个地方逛逛：
- 铁柱 AGI · Grok Bot 新手指南 ( https://bot.fmcat.top/ ) @cgnot996 目前国内整理得最完整的中文实操文档。从环境配置、接微信飞书一路写到多 Bot 协作，各种实测踩坑和骚操作全在里面。
- UseGrokBot 角色模板库 (usegrokbot.com/zh-cn/templ…) @Chris62771610 按职业挑角色，缺写代码的、搞视频的还是做调研的，点进去直接领走现成配置，立马开工。
- 官方 Marketplace：x.ai/bot/marketplace （侧边栏一键添加）
- awesome-grokbot：github.com/kydlikebtc/…（社区整理的开源索引）
最后说一下
折腾了整整两周，我最想说的并不是 Grok Bot 技术有多神，也没必要去吹它能颠覆什么。
它依然有很多小毛病，偶尔也会犯傻，复杂协作时也需要你花心思立规矩。
但它最实在的地方在于，能真真切切融入我们每天的生活和工作，分担掉一部分精力，这就已经足够好用了。
我是宅小年，下期我们再见！
关注公众号「宅小年」，阅读更多分享文章
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。
After two weeks of high intensity, Grok Bot, I found this thing really addictive.
It is now basically walking, drinking water and brushing out mobile phones while in the pit.
Honestly, I knew it as early as August when it was publish.
It was full of "baby-sitter classes" and "Ten compulsory fairy techniques," and I thought it was a new round of AI hot sex. I didn't find out it was real until I was free for a few days.
Don't hang up. It's got a 24-hour cloud host.
Using Codex or Claude Code's remotes, you should know the pain: Computers at home or on a workspace have 24 hours to open a client, and even worse, remote connections on a mobile phone, often not connected because the network is broken. OpenClaw and Hermes also know that a large number of people can be discouraged by simply pretending their environment and dependence.
Grok Bot is the most comfortable place to open the box and move the computer directly to the cloud. Even if you put the lid on the notebook and cut off the net, it ran its own run on that machine in the clouds.
And as a real cloud computer, it's a few very different things from ordinary tools:
First, the hard drive is durable. You made it write a reptile script, download a bunch of data, turn off the browser, and it's still stable when it's in the hard drive, and it's coming out any second.
Second, all Bot shares the same computer. You can build several different types of Bots that share the same file system, the same network export, and even the login status of the browser. A Bot, for example, downloads a data package, and B Bot doesn't need you to download it manually to pass it on, read it directly on the local path. Each Bot also has its own separate desktop and browser windows, without interference.
Most importantly, you can even light its desktop directly, like the TeamViewer remote desktop. It's all on the screen with Chrome looking around, opening the terminal knock code. You can really click directly on it, hit it where it's stuck, order it, or pass it through a authentication code.
It's a cyber problem.
You work with a cloud computer, you're gonna hit the network sooner or later.
The pit was discovered because I wanted to install Gemini CLI in its terminal, and it was ordered that the error be reported as soon as it was over: User migration is not supported.
And I was wondering: "Why don't you report the cloud computers that are hanging in the US?"
Out of curiosity, I was tested and studied to find its mechanisms interesting: This cloud computer is defaulted to go online through ** "shared exit pool"**, which is shared by all users, with different IPs each time requested. It's this point in Cloudflare, it's going to go to that node, and sometimes it's going to fly out of the country, and it's going to lead to some area-sensitive tools that directly miss.
If you're going to run a mission like the one that's demanding the Internet environment on a daily basis, the government actually left a useful switch:
Found on the console: Set up a computer to pick up your computer, check through this computer to relay traffic.
So, all the network requests from the cloud computer will go out using your own network agent, and the network environment will be completely stable.
Don't watch the nanny tutorial.
Grok Bot's introduction I think is the fastest way to do it: create a new Bot, talk directly to it. You want to know what the basic function is, what the step is, ask it directly in the dialogue box, and it's clearer than any tutorial.
A lot of people used to play OpenClaw, Hermes, and felt that Bot needed to have a complex role and a complex set of Systems Prompt.
Here, just tell it what you want in plain language. Now the model doesn't have to make complex Prompt Engineering, it adjusts itself in interaction.
What really makes it productive is the "Connect Apps" function.
I've got a couple of ecologicals I have to connect to every day:
- GitHub: Let it read your project warehouse, pull code, submit PR.
- X (tweet) : As the master of real-time information, it is easy to check information and keep an eye on dynamics.
- Gmail and Notion: One as its mailbox, one as its working recording space.
Usually, you can throw a lot of work that you repeat every day and you have to spend your time watching. For example:
- Information radar: do not use its own daily brush to extract the latest content from a given circle on X every day, filter out water posts and advertisements, and summarize the points to synchronize to Notion
- Warehouse stakeout: let it focus on an open-source project of some core concern, or a new Issue and PR in its own warehouse, extracting key dynamics to mail boxes every day and using GitHub to brush them manually.
- Mail sorting: Automatically scan the mailbox once a day, filter out useless mail, pick out server alarms or key project dynamics, and sort it out and throw it into Notion. The whole process needs to be clear to it in big words, and the backstage itself managed to get the time-bound set.
Most interesting function: Group chat
Group chat, I think is the most interesting function of Grok Bot.
You can drag a couple of different players into the same group, throw them into the group, and they'll line up, dismantle them, and then arrange different Bots to handle them. And you can talk at any time, like in WeChat.
At the beginning, I found it particularly interesting to have a topic for them to discuss. For example, I've got a problem of choosing names, throwing them into the crowd, and a few Bots are talking to each other, giving reasons, even starting to vote, and emotional values are in place.
But after a few days of playing, when you start sending in some complicated tasks, you find some problems:
The first is context, which is prone to chaos. A couple of Bots, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know,
Another problem was that it was thought that the task would be sent to the group and that they would be automatically running water lines and waiting for results. And it turns out that because of the number of people, you have to stay in the crowd @ this, you've got to push that, you've put a lot of effort into the coordination.
There are two ways to work together:
- Single @ Serial Intersection: Direct in the current Bot @ Another Bot, assigned to the next after the current task. The advantage is that context is relatively clean, and the disadvantage is that the output needs to be switched to the corresponding Bot to see it.
- Focus: All players in the same group. It is completely transparent, but it is extremely consuming context. Many people may have seen the case mentioned in the official blog: "The five team merges over 100 PRs a day."
But it's not really from a bunch of Bots who talk freely in the crowd. The core is the need for a clear coordinator (as project manager) to act as a single entry point for connecting external tools, controlling progress and distributing worksheets, while specific code changes are sent back to a dedicated Cloud Agent to run.
There must be a single entry and coordinator to make Bot work really hard.
How did I get together?
With that in mind, I re-oriented the water lines and made two teams to help me share my day-to-day work.
1. Research and development team Build Lab
I have deployed a research and development team that basically covers the key players in everyday development:
Max: Product manager, help me get the vague ideas right.
- Pixel Penny: Design, interactive and visual advice
- Codey: All-house development, responsible for core code realization
- Buggsy: Test, special for bug and quality control
Tracker Tina: Project Manager for progress tracking and task distribution
- Hype Hazel: Market growth, paperwork and online publicity!
My previous practice was to drag people into a crowd of gossip, which turned out to be inefficient.
Now, the idea is to set up a single portal: I'm only looking for Max's communication needs in the group. After needs confirmation, output PRD, the project manager, Tina, comes out to break the work list, who does what is arranged by Tina alone, and I don't go one by one @. Tina regularly synchronizes the work of the various players in the group.
For example, I asked them to help me optimize my station, Max comb the plate to change, Tina sent Codey to knock on the code, and then directly to Buggsy to pick the prick. In the process, I have only two things to do: start by confirming the direction of the demand and merge the final code with the first point of approval.
Of course, I've been working on this process myself, and code code code and automating acceptance check is not perfect, far from 100% automatic. But it's too much to save from the initial human coordination.
Video Studio
In addition to writing codes, another HF line I use is for video.
Now, Opus 5.5 is actually very effective in video production, and in combination with the ecology of Grok Bot, efficiency gains are evident. But to stabilize the output in one piece, just one Bot can easily be context chaos.
I broke the video lines into four characters:
- Director-General: Control the overall rhythm and visual tone
- Planning the script: writing scripts storyboard and dialogue line
- asset: Specialize in matching images and references.
-voiceoversubtitle: Be responsible for aligning sound with text!
A video, from thought to film, is itself a combination of several different types of work. Someone set the course, someone came out of the structure, someone looked for asset and someone did post-production. Drive them alone. context is the cleanest. Similarly, there is a coordinating role to link up.
Here's the little video that I'm trying to run out of sample with this process:
The results of the film are not as impressive at present, and the details are far from being artificially refined, and many of the images and shot connections still cause problems, and I continue to do so myself.
But its greatest value to me was to take down the cost of “from an idea to seeing a piece”. Once I wanted to make a video, from the concept book, to the world looking for asset, to the time-point-to-track track, for half a day, and now that the first draft and the basic framework basically help me run a seven-seven-eight-eight, all I had to do was make a piece of it, and I had to make two key pieces of it, and I had already saved most of my energy.
I have made these two configured teams and group chats public templates that can be imported directly if you want to experience them: edisonz.cn/gok-bot-templates
If you want to see more games to be shared, you can go to these places:
@cgnot996 is currently the most complete Chinese-language exercise document in the country. From environmental configuration to the WeChat flying book to the multiple Bot collaboration, all kinds of actual pedestals and mush operations are in there.
- UseGrokBot Role Template Library (usegrokbot.com/zh-cn/templ...) @Chris627711610 by occupational role, code-deficient, videomaker or researcher, point in and take off the ready-to-do configuration and start working.
- Official Marketplace: x.ai/bot/marketplace
- agosome-gokbot:github.com/kydlikebtc/... (open source index organized by the community)
And finally,
For two weeks, the last thing I'd like to say is that Grock Bot technology is so divine, and there's no need to blow it up.
It's still got a lot of problems, and sometimes it's stupid, and it's a complex collaboration that requires you to think about the rules.
But what is most true is that it is good enough to be truly integrated into our daily lives and work and to share some of our energy.
I'll see you next time.
Watching WeChat Official Account, reading more shared articles
Auto-fetched and translated by AI Atlas. All rights belong to the original author.