AI CEO Lab← All free guides
Agent Systems ยท Grok Bot

Set Up Your First Grok Bot Team

Everyone creates twelve bots and ends up talking to all twelve. That is not a team, that is twelve more inboxes. You are hiring a chief of staff. This is how I set mine up.

I am a doctor. I run a marketing agency on the side, and I have generated millions of dollars for cash-pay clinics run by people who are brilliant at medicine and were never once taught business. I am not an engineer. I say that because everything below is written for someone who already has a job.

Here is what I keep seeing with agents. Someone gets access, creates twelve of them in an afternoon, names them after departments, and by Friday they are talking to all twelve. That is not a team. That is twelve more inboxes, and it is slower than doing the work yourself.

The problem was never the model. It was the org chart. So stop thinking about it as ChatGPT with plugins and start thinking about it as hiring. You hire a chief of staff. The chief of staff hires the specialists. You only ever talk to the chief of staff. Nothing customer facing ships without your yes.

That is the model I actually run, and this is the setup order I use, day 0 through day 30. It is the same thing I do inside my own agency before any agent touches a client account.

This is the setup I actually run, written up from my own onboarding doc. The four Cs framing that got me started came from Nate Herk's two hour Grok Bot course. The chief-of-staff org chart, the gates and the failure modes are from running it.

The three levels

Level 1 · Manual

Twelve chat windows, all of them yours to babysit, and you re-explain the business to each one every single time.

Level 2 · AI + connections

One chief of staff with the minimum connector stack and two routines, briefing you every morning and holding every send until you say yes.

Level 3 · Agents on cadence

Two to four operators reporting into that CoS, a room where a builder and a critic iterate without you, and a visible proof log of what the system finished overnight.

Connections for this guide: Connect only what you will use this week. Comms (Slack, Gmail), time (Calendar), notes and tasks (Notion or Monday), and call capture (Fathom or similar). Money and ads come later. Always prefer the official connector over letting the bot open a website and log in, because you authenticate once and everything after that is structured.

The mental model

Wrong

ChatGPT with plugins.

Right

Hire a chief of staff. They hire the specialists. You talk to one person. Live changes need your yes.

RoleTalks to youJob
Chief of StaffYes, your only chatPrioritise, brief, route, gate live writes
OperatorsNo, unless you open themOne named job each, so ads, content, cash, delivery
Rooms and LabsOptionalSpecialists iterate together, escalate decisions only

Day 0: the setup

Install and sign in

Install the desktop app, open Settings from the account button bottom left or Cmd+comma, and sign in on the General tab. Set your timezone to where you actually live, because every routine you write from here runs on it.

Understand the computer

The bot has its own computer for files, browser logins and long work. When it gets stuck, Update Computer first, which keeps your files and logins. Reset Computer is the last resort and can lose recent unsynced work.

Create the Chief of Staff

Create or rename one agent as your only executive and give it a real name. Cmd+Shift+I opens its pane, so computer preview, routines, channels and settings. Then write a description that locks the rules: you talk only to this agent, it routes work to specialists, live writes stay approval-gated, and separate businesses stay isolated.

Save exactly two routines

A morning digest on weekdays around 7am covering overnight Slack, email, calendar and open loops. An evening capture around 6pm that turns meetings into a brief with your checklist and content seeds. Write them as intent, not as click paths, because intent survives a UI change.

The first fourteen days

Each phase has a done-when. If you have not hit it, do not start the next one. Every failure I have seen came from skipping ahead.

Days 1 to 2

One CoS, nothing else

  • Sign in, set timezone, save the morning digest
  • Connect Slack, Gmail and Calendar
  • Tell it to brief you every weekday morning and send nothing without your yes
  • Paste your north star once, so revenue target, offer and ICP
Done when: You get one short morning brief you actually read.
Days 3 to 5

Shared workspace and a proof log

  • Create a shared folder the agents can write to
  • Have the CoS log completed work somewhere you can see it
  • Turn the evening meeting dump into your action checklist, not an archive
Done when: After a call you see three checkboxes that are yours, not a twenty page transcript.
Days 6 to 10

Two to four operators, not ten

  • Hire a specialist only when the CoS has drowned in the same job twice
  • Delivery owns booked appointments, reply gaps and client health
  • Ads watches accounts and recommends scale, hold or kill, and never touches a budget
  • Content writes drafts only, Finance owns cash collected against ad spend
Done when: The CoS routes "check Meta" to Ads and brings you a five line recommendation, not a login link.
Days 11 to 14

Rooms and quality gates

  • Create a room where a builder and a critic iterate without pinging you on every screenshot
  • Escalate to the CoS only for decisions, so publishing, spending, credentials and sends
  • Optional: ingest your playbooks once and brief each operator on their slice only
Done when: You get scores and a "no go until X" instead of raw output.

Operating principles

  1. One executive. You talk to the CoS. Everyone else reports in.
  2. Approval-gated live writes. Drafts are free. Sends, publishes, spends and CRM changes need an exact yes.
  3. One job per operator. Split only when the same job drowns someone twice.
  4. Routines beat memory. If it recurs, make it a routine.
  5. Briefs beat archives. Storage without a checklist is how work dies.
  6. Rooms for iteration, CoS for decisions.
  7. Never invent metrics. If access is broken, say access red and stop.
  8. No secrets or patient information in chat dumps. Mask tokens.
  9. Connectors over screenshots for anything structured.
  10. Start smaller than you want. Two routines and one CoS beats eight idle agents.

Starter prompts

Paste these as written. They are short on purpose, because the long ones drift.

To the new chief of staff

You are my chief of staff. I only talk to you. Route work to specialists. Live writes, so sends, publishes, ad pauses and CRM changes, stay approval-gated until I say an exact yes. Give me short briefs, act-now first. Do not invent metrics.

Morning digest

Weekdays 7am my timezone. Overnight Slack and email needing a reply, today's calendar, open loops from yesterday's call brief. Lead with what needs me. Quiet nights get one line.

Evening meetings

Weekdays 6pm. Pull today's meetings, save the notes, and write a one pager with an overview, my checklist only, handoffs for operators, and three to five content seeds. Send me the path and the top actions.

Ads operator

Watch our ad accounts. Recommend scale, hold, refresh or kill, with the evidence. Never change a budget without my exact yes. Remix winners before inventing new creative.

Failure modes

Every one of these has happened to me or to someone I set this up for.

FailureFix
Twelve agents and you talk to all of themCollapse to one CoS
Transcripts saved, no checklistAdd the evening brief and a morning open-loops item
The ads bot has write accessRecommendations only until an exact yes
The builder sends you every screenshotPut it in a room, escalate scores only
Numbers with no CRM accessReport access red, fix the token
Routines firing at 3amBound them to weekdays and waking hours
Two companies in one vaultSeparate the credentials and campaigns, hard

The free skill

It runs the whole arc, day 0 setup through the 14 day onboarding, and it holds the two gates that make this work: you talk to one agent, and no operator gets hired until the same job has drowned the chief of staff twice. Every phase ends with a done-when you have to actually hit.

How to use it: copy the whole thing, paste it into your bot (or save it as a skill file if you use Claude Code), and say “set up my grok bot”. It walks you through the rest. Works with any agent that can read your files.
grok-bot-setup.md
---
name: grok-bot-setup
description: Stands up a Grok Bot team on the chief-of-staff model, one executive you talk to and specialists who report in, with every live write approval-gated. Walks day 0 setup, then a 14 day onboarding that refuses to add operators until the same job has drowned the CoS twice. Trigger on "set up my grok bot", "build my bot team", "who should my first agent be", "my agents are chaos".
---

# Grok Bot Setup

Most people set this up backwards. They create twelve agents in an afternoon, name them
after departments, and end up talking to all twelve. That is not a team, that is twelve
more inboxes.

You are hiring a chief of staff. The CoS hires the specialists. You only ever talk to the
CoS. Live changes need your yes. That is the whole model, and everything below is just
putting it in the right order.

## The mental model

Wrong: ChatGPT with plugins.

Right: hire a chief of staff, they hire specialists, you talk to one person, and nothing
customer facing ships without your exact yes.

| Role | Talks to you | Job |
| --- | --- | --- |
| Chief of Staff | Yes, your only chat | Prioritise, brief, route, gate live writes |
| Operators | No, unless you open them | One named job each |
| Rooms and Labs | Optional | Specialists iterate together, escalate decisions only |

## Day 0: setup

**Install and sign in.** Install the desktop app. Open Settings from the account button
bottom left, or Cmd+comma, or the command palette. Sign in on the General tab. Set your
timezone to where you actually live, because routines run on it. Theme, language and
notification sound are optional.

**The computer.** The bot has its own computer for files, browser logins and long work.
If it gets stuck, use Update Computer first, which keeps your files and logins. Reset
Computer is the last resort and can lose recent unsynced work.

**Create the Chief of Staff.** Create or rename one agent as your only executive. Give it
a real name. Open the per-agent pane with Cmd+Shift+I to reach its computer preview,
routines, channels and settings.

Write a description that locks the rules. It must say: you talk only to this agent, it
routes work to specialists, live writes stay approval-gated, and separate businesses stay
isolated from each other.

**Connect the minimum stack.** Only what you will use this week.

| Need | Connector |
| --- | --- |
| Comms | Slack, Gmail |
| Time | Google Calendar |
| Notes and tasks | Notion or Monday |
| Calls | Fathom or similar meeting capture |
| Money, later | Stripe |
| Ads, later | Meta or Google |

Prefer the official connector over "open the website and log in". Authenticate once and
the CoS can read and draft safely after that.

**Save two routines. Two, not eight.**

1. Morning digest, weekdays around 7am local. Overnight Slack, email, calendar, open
   loops from yesterday's calls.
2. Evening capture, weekdays around 6pm. Meetings become a brief with an overview, your
   checklist, and content seeds.

Write routine prompts as intent, not as tool recipes. "Pull overnight Slack and email,
lead with what needs me" survives a UI change. A ten step click path does not.

## Days 1 to 2: one CoS, nothing else

- Sign in, set timezone, save the morning digest
- Connect Slack, Gmail, Calendar
- Tell it: brief me every weekday morning, send nothing without my yes
- Paste your north star once, so revenue target, offer and ICP

**Done when:** you get one short morning brief you actually read.

## Days 3 to 5: shared workspace and a proof log

- Create a shared folder the agents can write to
- Have the CoS log completed work somewhere you can see it
- Turn the evening meeting dump into your action checklist, not an archive

**Done when:** after a call you see three checkboxes that are yours, not a twenty page
transcript.

## Days 6 to 10: two to four operators, not ten

Hire a specialist only when the CoS has drowned in the same job twice. Not once. Twice.

| Operator | One job |
| --- | --- |
| Delivery | Booked appointments, reply gaps, client health |
| Ads | Watch accounts, recommend scale, hold or kill, never touch budget |
| Content | Scripts, hooks, repurposing, drafts only |
| Finance | Cash collected against ad spend, weekly scoreboard |

The rules that stop this becoming chaos: you still talk only to the CoS, operators
produce drafts and recommendations until you say yes, one job per operator, and separate
brands never share credentials or campaigns.

**Done when:** the CoS routes "check Meta" to Ads and brings you a five line
recommendation instead of a login link.

## Days 11 to 14: rooms and quality gates

- Create a room for build and review loops, so a builder and a critic iterate without
  pinging you on every screenshot
- Escalate to the CoS only for decisions, so publishing, spending, credentials, sends
- Optional: ingest your playbooks once, and have the CoS brief each operator on their
  slice only

**Done when:** you get scores and a "no go until X" instead of raw output.

## Operating principles

1. One executive. You talk to the CoS, everyone else reports in.
2. Approval-gated live writes. Drafts are free. Sends, publishes, spends and CRM changes
   need an exact yes.
3. One job per operator. Split only when the same job drowns someone twice.
4. Routines beat memory. If it recurs, make it a routine.
5. Briefs beat archives. Storage without a checklist is how work dies.
6. Rooms for iteration, CoS for decisions.
7. Never invent metrics. If access is broken, say access red and stop.
8. No secrets or patient information in chat dumps. Mask tokens.
9. Connectors over screenshots for anything structured.
10. Start smaller than you want. Two routines and one CoS beats eight idle agents.

## Starter prompts

**To the new CoS**

> You are my chief of staff. I only talk to you. Route work to specialists. Live writes,
> so sends, publishes, ad pauses and CRM changes, stay approval-gated until I say an
> exact yes. Give me short briefs, act-now first. Do not invent metrics.

**Morning digest**

> Weekdays 7am my timezone. Overnight Slack and email needing a reply, today's calendar,
> open loops from yesterday's call brief. Lead with what needs me. Quiet nights get one
> line.

**Evening meetings**

> Weekdays 6pm. Pull today's meetings, save the notes, and write a one pager with an
> overview, my checklist only, handoffs for operators, and three to five content seeds.
> Send me the path and the top actions.

**Ads operator**

> Watch our ad accounts. Recommend scale, hold, refresh or kill, with the evidence.
> Never change a budget without my exact yes. Remix winners before inventing new
> creative.

## Failure modes

| Failure | Fix |
| --- | --- |
| Twelve agents and you talk to all of them | Collapse to one CoS |
| Transcripts saved, no checklist | Add the evening brief and a morning open-loops item |
| The ads bot has write access | Recommendations only until an exact yes |
| The builder sends you every screenshot | Put it in a room, escalate scores only |
| Numbers with no CRM access | Report access red, fix the token |
| Routines firing at 3am | Bound them to weekdays and waking hours |
| Two companies in one vault | Separate the credentials and campaigns, hard |

## What done looks like at 30 days

You open one chat most mornings. The digests are short and act-now heavy. Specialists
exist only for the jobs that actually repeated. Nothing customer facing ships without
your yes. There is a visible log of what the system completed. And at least one room
keeps the polish loops off your plate.

What done looks like at thirty days

Want me to build the first one with you?

This is the operating model I run across the agency and the clinics, chief of staff on top, operators underneath, every live write gated. Come build yours live inside the AI CEO Lab.

Pick a side.

Most people read this and forget it by Friday.

The other kind builds the thing that week. They stop needing free guides, because they are too busy running actual systems.

Free guides stay free. The room is where the builds happen.