A Field Guide to Grok Bot
If you opened Grok Bot, asked it something and closed the app, you used about 5% of it.
That's a chat habit. Grok Bot is built for something else: named bots with their own computer in the cloud that keep working after you close your laptop.
I run my X account and my agency on it. When I wake up, a brief is waiting, brand DMs are already researched, and a table of fresh leads is filled in. This guide is everything I did to get there, in order. Each part tells you when to use a feature, exactly where to click, what to paste, what a good result looks like, and where I tripped.
1. How it fits together
Five pieces, and the one thing to know about each.
2. Day one: install and sign in
Do this once, in one sitting. Everything later builds on it.
Install the app
On first launch you click through a short intro. The computer sets up in the background, and the app opens a chat with your first bot, called Grok Bot.
Sign in to everything at once
When to use: day one, before you build any routine.
Open x.com, Gmail and my calendar in your browser one by one. When a page asks for a login, a 2FA code or a CAPTCHA, stop and hand it to me. Don't type any passwords yourself.Good result: every site you need opens signed in, and the bot confirms each one.
Tip: wherever a service has a connector or an API, use it instead of a browser login. Sites log bots out often, and connectors don't break like that. Never paste a password or a one-time code into the chat. Use Take over.
3. Set up your first bot the right way
The worst first bot is "General Helper". Give each bot one job and define five things:
Create the bot
You are my [role]. Your job: [goal]. Use [apps and sources]. Every result should be [format]. Never send, publish or delete anything without asking me. If a source is missing or out of date, tell me instead of guessing.Tip: memory is for rules that stay true (“never send without approval”). The chat is for today’s task (“draft replies to these 12 DMs”). When you notice a lasting preference, tell the bot and ask it to remember it.
The loop for every new bot
Jump straight to step 5 and you get a routine that does the wrong thing every morning at 7.
For step 2, when the format is right, lock it in:
This is the format I want from now on: five bullets, source links inline, and a last section called "Decisions needed." Remember it for future results.For step 3:
Save the process we just used as a skill called "Morning brief". Include where you read from, how you check claims, the output format, and the rule that you never send or post anything without asking me.Good result: ask "show me the Morning brief skill you just saved". The bot shows its steps. To use it later, name it in your message: "Use my Morning brief skill to…"
For step 5, see "Make a routine" in the next section.
4. Features worth using
Claude, ChatGPT and others all have agents now. These are the things I haven't found together anywhere else.
Make a routine
When to use: a task already works by hand and you want it to run without you.
Every weekday at 8am, run the Morning brief skill. Post the result in this chat. Don't send or post anything anywhere else. If a source is unavailable, tell me it failed instead of using old data. If there's nothing new, still tell me "nothing new".Good result: the test run posts a result in the format you asked for, and it stops where it should ask you.
Watch out: a test run does real work. It can open sites, change files and call connected tools. Test on safe inputs and keep anything that writes behind approval. Check your timezone first in Settings → General → Timezone. Routines use it.
Step into the bot's computer
When to use: you want to watch the work, or the bot hits a password, a 2FA code or a CAPTCHA.
Then close the laptop. The work keeps going.
Tip: for a login, a checkout address or a phone number, the bot can show a form in the chat and fill your answers into the page. For supported connections it can show a secure secret request; the value is masked and never shows up in the chat.
Tag @bot on X
When to use: you're on X and want a bot to do something with a post without opening the app.
Before the first time:
Then, every time:
@bot summarize this thread and send the key points to my NotionThe bot gets your post, the post you replied to, and any post either of them quotes, including still photos. The first time, X sends you a notification that tagged requests start work automatically and use your Grok Bot usage.
Searching posts, mentions and trends works out of the box. Connect the X plugin (it uses your own X API credits) only when a bot needs your DMs, bookmarks or home timeline: Connect apps → find the X plugin → Add → sign in with X.
Watch out: @bot stays silent if the post has a video or GIF, if you also tag @grok, or if your weekly usage is used up. It also stays silent if you reply in a thread that already has @bot but don't tag it yourself. If @bot answers "Sign up for Grok Bot", your X account isn't connected on grok.com. If it answers "Grok Bot isn't set up yet", create a bot first.
Teach a task by showing it
When to use: a browser workflow is easier to show than to describe.
To fill the gaps in the draft, send something like:
Look at the skill you just made. Add what to do when a field is empty or the page looks different, and make it stop and ask me before anything gets submitted or sent.Good result: a skill that works on a second, different example, not just the one you recorded.
Watch out: the draft is the start, not the end. One recording doesn't show your edge cases. Don't show passwords on screen while recording; use Take over for logins. Teach a task is still rolling out. If you don't see the button, do the task with the bot in chat and ask it to save that as a skill.
Make bots work together
When to use: one bot owns the source, another owns the result, or a draft needs a second pair of eyes.
Direct handoff (one bot to another):
When you finish the brand research, send it to Grok Bot and ask it to check the draft reply before it comes to me.Group chat (2 to 6 bots):
@Researcher gather the source material and link every claim. @Writer turn it into a draft. @Reviewer check the draft against the sources and list only blocking issues. Nobody publishes anything.Tip: in a group, give every task one owner with @. Two bots on the same task will both do it. Approval requests, login requests and email or Slack drafts don't show up in groups, so keep that work in one-to-one chats.
Set up Auto-review
When to use: before your first routine. It's the reviewer that checks each action before it runs.
The rules I start with:
Ask first: sending any message or DM
Ask first: deleting files
Allow automatically: reading X
Allow automatically: web searchWhen an approval card shows up in chat, check the target and values, then pick Allow once, Always allow (saves a matching rule) or Deny.
Good result: your next send or delete stops with a card; reading and searching go through without asking.
Tip: write narrow rules around one action and one place. "Allow everything in the browser" is too broad. Your rules are saved to your account, so they work on every computer and on your phone.
Watch out: approvals from routines, triggers or other bots expire after about 10 minutes. The expired card still offers Always allow this in the future. That saves the rule but doesn't run the missed action, so ask the bot to retry it.
Pages
When to use: a report you check often, like a lead list or a morning dashboard.
Make me a live page with today's leads: name, post link, pain point and score. Keep updating that same page every time the routine runs, and give me the link.Tip: give a routine one Page to update instead of sending you a new message every morning.
Claude inside your bots
When to use: you write code and want a bot to manage, schedule and check the work.
The bot's computer has a terminal, so you can install Claude Code there and have a bot run Claude Code sessions for you. Grok Bot manages, schedules and checks the work; Claude Code writes the code.
Install Claude Code on your computer and tell me when you need me to sign in. Then write down in /workspace/setup-notes.md exactly what you installed and how, so we can redo it if the computer gets rebuilt.You can also hand code work to Cursor Cloud Agents. Just ask the bot to delegate the task to a Cloud Agent. When it finishes, the bot attaches the images, videos and files from its report to the reply.
Tip: one bot per project.
Share a bot
When to use: someone else should start from your setup.
People get a copy of the setup, never your computer, logins or history.
There's also a Marketplace of ready bots in the sidebar: press Add, or open a bot and press Import Bot.
Team Bots are one bot a whole team talks to. Each person gets a private chat, it can have its own Slack app, and it acts as whoever is asking. To make one, press New → Create new Team Bot, or open an existing bot's Share menu → Publish to Team.
Watch out: a public link shows the bot's description, skills and routines. Remove API keys, internal links and client data before you share.
Start from a ready bot: the Marketplace
When to use: you want a bot for a common job and don't want to write its setup from scratch.
The Marketplace is a catalog of ready bots built by the Grok Bot team and by well-known people in tech. Each one comes with its job, rules and routines already written. Adding one gives you your own copy of that setup. It runs on your computer, with your logins and your usage.
A few worth a look:
Use your own computer when you need it
When to use: a site blocks cloud servers, or a bot needs files on your desktop.
Run commands on your desktop:
Route web traffic through your IP:
Watch out: the route goes through the desktop you turned it on from. Keep "Ask every time" for commands unless a bot has a real reason to touch your local files.
5. The structure: one boss, many workers
My first mistake was letting my main bot do two jobs: route tasks and keep track of everything. It got noisy fast. Now I split it:
The rule: communication tools are read-only or draft-only for every bot. Nothing goes out without me.
To set the front door up, send this to your main bot and ask it to remember it:
You are my front door. When I give you a task, send it to the bot that owns that job and tell me who has it. When another bot sends you something high-stakes for review, like a reply to a brand or anything with money or a deadline, check it against the sources. If it's fine, pass it to me. If not, fix it or send it back with what's wrong. Never send, publish or delete anything yourself.
6. Setups you can copy
Every setup below follows the same order: create the bot (section 3), connect what it needs, paste the description, run the task once by hand, save it as a skill, then paste the routine and check it.
Morning Brief
When to use: you want to know what matters on X before you open it.
What it does: reads my X bookmarks and what the accounts I follow posted overnight, checks each claim against the original source, sends one short brief at 7am with links. Read-only.
Connect: the X plugin (Connect apps → X plugin → Add → sign in with X). Bookmarks and the home timeline need it.
First message (ask it to remember):
You are my Morning Brief. Your job: tell me what matters on X before I start my day. Read my X bookmarks and what the accounts I follow posted overnight. Check every claim against the original source and link it. Keep it short: what happened, why it matters to me, the link. You are read-only. Never post, reply, like, repost or send anything. If you can't open a source, say so instead of guessing.Routine:
Every day at 7am, read my X bookmarks and what the accounts I follow posted since yesterday's brief. Pick what actually matters for an AI agents account, check each claim against the original source, and send me one short brief in this chat with a link for every item. If there's nothing worth reading, still tell me "nothing new".Auto-review: Ask first: sending any message or DM. Allow automatically: reading X, web search.
Check it works:
Inbox Manager
When to use: brand DMs about sponsored posts come in faster than you can research them.
What it does: reads brand DMs about sponsored posts, researches each brand (team, product, funding, X sentiment) and drafts a reply. Only drafts; a reviewer checks the draft. I send.
Connect: the X plugin. DMs need it.
Description:
You are my Inbox Manager. Your job: handle brand DMs about sponsored posts and collaborations. For each brand, research the team, product, funding and recent X sentiment, and save a one-paragraph summary to /workspace/brands.md. Draft a reply in my tone. You only draft. Never send a DM or a reply. Before a draft comes to me, send it to Grok Bot for review. If a brand asks for a call or a payment, flag it at the top.Routine:
Every 2 hours, check my X DMs for messages about sponsorships or collaborations. For each new brand, research the team, product, funding and recent X sentiment. Save a one-paragraph summary to /workspace/brands.md and draft a reply in my tone. Do not send anything. If a DM asks for a call or a payment, flag it at the top. If there's nothing new, still tell me "nothing new".Auto-review: Ask first: sending any message or DM. Allow automatically: reading X, web search.
Check it works:
Lead Miner
When to use: you need warm leads for clients without cold outreach.
What it does: searches X for people complaining about competitors of my agency's clients, scores each lead, adds it to a table with post link and pain point. No outreach. I mark who replied; it uses that to score next leads.
Connect: nothing. Searching public posts works without the X plugin.
Description:
You are my Lead Miner. Your job: find people on X who complain about competitors of my agency's clients. My clients and their competitors are listed in /workspace/clients.md. For each lead, score how likely they are to switch and add them to /workspace/leads.csv with name, post link, pain point, score, and a "replied" column. Never contact anyone: no replies, no DMs, no follows. When I mark someone as replied, use that to score the next leads better.Routine:
Every morning at 8am, search X for posts from the last day where people complain about the competitors in /workspace/clients.md. Skip anyone already in /workspace/leads.csv. Score each new lead, add it to the table, and tell me how many you added and which three look best. If there are no new leads, still tell me.Auto-review: Ask first: sending any message or DM. Ask first: deleting files. Allow automatically: reading X, web search.
Check it works:
Chief of Staff
When to use: you have more than two bots and lose track of what they did.
What it does: pulls what changed across my bots, Slack, email and calendar into one queue and one digest with a link per item. Never sends messages or moves meetings.
Connect: Slack, email and calendar from Connect apps.
Description:
You are my Chief of Staff. Your job: keep one task queue for me and my bots in /workspace/queue.md. Every bot reports into it. Read what changed in my bots, Slack, email and calendar, and add each item to the queue with a link to the source. Write my morning digest from the queue. You only read and write the queue. Never send a message, reply to an email, or create or move a meeting.Then tell each worker bot, in its own chat:
From now on, when you finish a task or get stuck, send a one-line update to Chief of Staff with a link to the result.Routine:
Every weekday at 7:30am, go through what changed since yesterday in my bots, Slack, email and calendar. Update /workspace/queue.md and send me one digest here: what's new, what's waiting on me, what's overdue, with a link for every item. If nothing changed, still tell me.Routines can also start on a Slack or GitHub event. Keep the trigger narrow: "a message in #clients that mentions a deadline", not "every new message". Event triggers come from Cursor account integrations, which are separate from the Slack and GitHub plugins.
Auto-review: Ask first: sending any message or DM. Ask first: creating, changing or deleting calendar events. Allow automatically: reading Slack, email and calendar.
Check it works:
Builder
When to use: you have code work and want it done while you sleep, but reviewed by you.
What it does: runs Claude Code sessions or Cursor Cloud Agents. Let it open PRs; keep merging for yourself.
Connect: GitHub from Connect apps. If you use Claude Code, set it up as in "Claude inside your bots".
Description:
You are my Builder for [project]. Your job: take coding tasks, run them through Claude Code on your computer or hand them to a Cursor Cloud Agent, check the result, and open a pull request. Explain what changed and how you tested it. You may open PRs. Never merge, never push to the main branch, never deploy. Keep notes on how the project is set up in /workspace/[project]/setup-notes.md.Routine:
Every weekday at 10am, look at the open issues labeled "bot" in [repo]. Pick one, do it with Claude Code, run the tests, and open a pull request. Send me the PR link and a short summary here. If the tests fail or the task is unclear, stop and tell me instead of guessing. If there are no issues, still tell me.Auto-review: Ask first: merging a pull request. Ask first: pushing to the main branch. Ask first: deleting files.
Check it works:
7. Tricks
8. Mistakes to skip
Watch out: Bots share one computer, so every login is available to all of them. Don't sign in anywhere one bot shouldn't touch.
9. Price and access
Part of The X Subscription (Premium, Super or Ultra), bundling Grok, Cursor and X Premium with one weekly usage pool. You can also link SuperGrok, X Premium+ or a Cursor plan. Plans change fast; check the plans page.
Apps for macOS, Windows, Linux, iPhone, iPad, Android.
To see what you've used: Settings → Usage & Billing, or Weekly usage in the account menu.
Where to start
Grok Bot doesn't replace one chat with a better chat. It turns asking for help into finding the work already done. Start with one bot, one job, one routine. Add the second only after the first has run a few days without you touching it.