跪拜 Guibai
← Back to the summary

Grok Bot Team Building: Fewer Bots, Self-Written Skills, and Intern-Style Onboarding

Recently, xAI's Grok Bot engineering lead Lauren Tan (@poteto) posted a guide to building a Grok Bot team.

image.png

I have to say, besides being full of useful content, the comment section also had a lot of fun: from a large-scale "calling @bot to claim tasks" spectacle, to hardcore developers supplementing permission architecture, to xAI engineers catching bugs on the spot, and Chinese users making sharp comparisons between Grok Bot and Muse. The information density is very high.

Today I'll take a first-person reading perspective and walk you through the original post and the entire comment section's useful content, confrontations, and criticisms.

一、 Breaking Down the Original Post

Many people's biggest pain point with AI agents is wanting to create a dozen fancy bots at once, only to find they frequently lose memory and require you to manually copy-paste messages between them every day.

As the person in charge of Grok Bot engineering at xAI, someone who previously merged over a thousand PRs in a single month (AI Engineer profile), Lauren's solution in the main post is counterintuitive: limit the number, delegate tools, and treat bots like interns. I've organized her seven practical tips into the following core threads:

"add dr eggbot to your team! he's my bot designer that i use to create bots"

I looked up dr eggbot myself; it's essentially a "bot that specializes in designing and auditing other bots." You don't need to rack your brain writing system prompts. After adding it to your team, it asks you a few preference questions, then generates a bot with extremely clear responsibility boundaries.

"you don't need a lot of bots. a few good bots can be more useful and easier to manage than a large team"

A few well-tuned, easy-to-manage core bots are much more practical than a ragtag team of a dozen.

"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"

This is the tip I find most practical. When conversations get long, models tend to drift; but if you teach it how to do tasks while asking it to save the method as its own Skill, it can maintain high consistency across long-cycle tasks. You can even have it directly create a private GitHub repository to store all accumulated Skills independently, building your own digital asset library.

"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"

Connect all the connectors, use the key vault, and hand over access to the tools you use in your work. Even more hardcore: to better separate permissions from personal accounts, you can directly let the bot apply for its own dedicated email and register accounts on the platforms it needs.

"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!"

Whenever you find yourself doing something manually, or copying and pasting information from elsewhere back into the conversation, stop and ask your bot: "How can you do this step for me directly?"

"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"

Don't treat the bot like a search engine. Like onboarding a new intern, first introduce yourself, describe your daily work, then let it think for itself and propose which parts it can take over and automate.

"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"

Typing often makes us omit key context. Just turn on voice mode or transcription and dump out the background in your head. The most important final command is "restate this in your own words" — you can immediately verify whether it truly understood.

二、 Summoning @bot Across the Network, Even "Your Mom Says Improve" Appears

Because Lauren wrote "ask @bot to read this post and help you improve your team!" at the beginning of the post, when I scrolled down to the comments, the first scene was a massive performance-art event: more than half the comments were directly tagging their own @bot to come copy the homework.

All kinds of "boss assigning work" styles appeared:

"@bot your mom is saying we need to follow this tips to improve our parche" (@bot your mom says we need to follow these tips to improve our team!)

This interaction of letting the agent read the whole thread and self-improve without copying the original text—just tagging @bot in the comments—perfectly demonstrates Grok Bot's native integration advantage with the X platform. No wonder @0xMorlex remarked that Lauren keeps releasing truly alpha-level useful content.

三、 Architecture Additions and Discussion on "Proactivity (SOUL.md)"

Beyond the bot-summoning craze, two high-quality deep discussions in the comments received direct responses from Lauren herself and the xAI core team, and were even more substantive than the main post:

1. Minimal Permissions and Verification Loops

Against Lauren's fourth tip ("give every bot access to all the tools you use"), multi-agent developer @CobaltHere proposed a very experienced engineering correction:

"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 points are three:

Lauren immediately replied with a thumbs up: "good tips!" — which also fits perfectly with the "Verification is everything" principle she has emphasized in her open-source project pstack.

2. From "Human-finds-Bot" to "Bot-proactively-finds-Human" Meta-capability

Another netizen, Fernando Fernandez (@Fdfdz90), raised a product-level meta question:

"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"

He believes a truly advanced Grok Bot shouldn't wait for humans to review their own workflows and explicitly request automation. Instead, it should have a underlying "soul" setting like SOUL.md: when it sees the user paying attention to a topic or repeating an operation, it should proactively ask questions, configure needed information sources, and do an extra step of expert-level extended research.

Two xAI insiders directly responded to this suggestion:

四、 Real Pain Points and Cold Thoughts: What Shortcomings Did the Comment Section Expose in Grok Bot?

A real evaluation can't just look at rainbows and farts. After carefully browsing the comments, I found that the pain points and doubts people actually encountered with Grok Bot are equally real, mainly concentrated in five areas:

Commenter Core complaint / real pain point Details and progress
Chandu (@ClaudeColossus) High-frequency scheduled tasks become a "polite stalker" He joked: "my routine checks x every 10 minutes for posts to reply to. is that proactive or did i build a very polite stalker?" Checking X every 10 minutes to find posts to reply to vividly confirms that blindly setting high-frequency routines not only burns credits but also makes the bot overly aggressive.
serge (@sergehere) Routines triggers failing Complained "routines are not working, I tried setting up some basic reminders and none fired…" indicating that the stability of scheduled reminders and automated triggers still needs reinforcement on some users' side.
furylao (@Bouabdallah1870) Self-built agents slow, missing steps, hard to verify Asked Lauren for help: the "kitchen agents" he built run slowly and often skip steps, and he didn't know how to improve verification and execution.
Vishal Singh (@vishalsingh2972) Multilingual speech recognition (Hinglish) struggles Reported that after trying voice mode per Lauren's suggestion, when switching between Hindi and English (Hinglish), the bot had a hard time recognizing. Lauren immediately cc'd @parker__conrad in the comments to follow up.
jemini (@brisyramshere) Differentiation questions vs Muse + Codex, and "can't create new sessions so burns credits" Raised the sharpest competitive comparison and mechanism critique in the whole thread (detailed below).

In particular, the last Chinese user jemini (@brisyramshere)'s question hit the core pain point of many heavy multi-agent users:

"Can anyone give some concrete examples of what Grok Bot can do that Muse can't? I still find Muse more comfortable for some reason. I can reproduce everything Grok Bot does in Muse right now, and connecting CodeX in Muse achieves the same level of intelligence. Conversely, I'm bothered that you can't create new conversations for Bots in Grok Bot, because it makes my credits burn faster and faster."

This comment very objectively points out two real issues:

  1. Capability homogenization and workflow habits: For heavy users already familiar with attaching Codex in Muse, routine agent tasks can be reproduced on both sides. Grok Bot's most unique differentiation currently lies in its deep integration with X's social graph (like directly @bot in comments) and the out-of-the-box team orchestration ecosystem like dr eggbot.
  2. Token consumption anxiety from a single persistent thread: Because Grok Bot uses a single persistent conversation flow like chatting with a human colleague, you can't create a "New Chat" for the same bot like traditional chat software. As historical context accumulates, each conversation or routine run burns credits (tokens) faster. This is why Lauren has previously specifically reminded people to use dedicated new bots for periodic tasks, or have dr eggbot regularly do token waste audits (routine-healthcheck) (Grok Bot Marketplace).

五、 From "Playing with Prompts" to "Doing Organizational Architecture"

Connecting Lauren's original post and the dozens of comments below, my biggest takeaway is: people's focus in using AI has completely shifted from "how to write a good prompt" to "how to design an organizational architecture that doesn't spiral out of control."

Whether it's Lauren's emphasis on "onboarding an intern + voice dump + having it restate in its own words", or @CobaltHere's supplement on "minimal permissions + result verification", or @brisyramshere's careful accounting of persistent session credit burning, they all essentially solve the same problem: how to use the lowest communication and compute cost to make silicon-based colleagues deliver results stably and reliably.