From official docs and what builders on X repeat. Most of it is about writing down the boundaries.
Tight personas, not fluffy descriptions
The persona should be specific enough that the bot knows exactly what to do, and what itmust not do.
Avoid
"You are a helpful assistant that helps me with everything I need."
Too broad. The bot does not know what it is good at.
Do this
"You handle my inbox and do prioritization. Every morning you read the latest 20 emails, categorize them by priority (Urgent/Normal/Can wait), and list them in a markdown document. You NEVER reply to email for me."
Clear scope. The bot knows its job and its limits.
Define Don'ts clearly
Don'ts are as important as jobs. They protect you from unexpected "helpful" actions, the bot that sends the email it was only supposed to draft.
Role
Sales Bot
Follows up leads and keeps the CRM updated.
Jobs
Follow up leads that have not replied within 3 days
Update the CRM with the latest interaction
Suggest next steps based on the lead's reply
Don'ts
NEVER send email without my approval
NEVER delete leads from the CRM
NEVER offer discounts or special pricing
Make NO decisions that affect pricing
On-demand vs routines
Choose the trigger method based on the task:
On-demand
Manual trigger. Use for:
Tasks that need fresh context ("analyze this competitor's site")
Things you want to double-check before they run
Workflows you are still testing
Routine
Scheduled. Use for:
Repetitive tasks that should always happen (morning inbox triage)
Tasks where "forgetting" is worse than "running once too often"
Avoid over-asking
A common mistake: ask the bot for confirmation constantly. That kills the point of automation.
Over-asking
"Run prioritization on the mail in my inbox. Ask me before you categorize each email. Also ask whether to include mail from the spam folder."
The bot cannot work independently. You end up answering a pile of questions.
Clear rules
"Run prioritization on the mail in my inbox. Ignore the spam folder. Rules: mail from customers = Urgent, newsletters = Can wait, everything else = Normal."
The bot knows the rules and works without asking.
Approvals: two layers
Don'ts in the description are layer one. Layer two is the product's own rules underSettings → General → Auto-review. Set Require Approval for:
send
publish
buy
delete
change production
Require Approval wins over Always Allow.
Least privilege
Connect only the tools the flow needs
Start read-only and drafts
Service accounts where the source system allows it
Review plugins and active routines regularly
Grok Bot can shop online via @link: you approve each purchase; the Bot gets a one-time card. At time of writing: USA, rolling out to mobile.
One bot, one job
Resist the urge to build a superbot that does everything. It gets messy fast.
Why specialists work better
Easier to debug: when something goes wrong, you know which bot it is
Simpler personas: less context means clearer instructions
Scales: you can run several bots in parallel
Faster iteration: change one bot's persona without affecting others
Example fleet
Chief of StaffPinned. The only one you talk to. Delegates.
Inbox / calendarHandles only mail and meetings
Research BotFinds data and summarizes
One narrow specialistFor your actual job
Tell CoS before you create several. Use group chat when the handoff itself should be visible. Duplicate a Bot when the same role needs a new scope (the copy gets profile, skills, and routines, but not history or memory).
State on disk, not in the chat
Docs: /workspace on the cloud computer survives computer updates. Builders on X:"Chat is history. That file is memory."
One Bot writes the board file; the others read
Two Bots writing the same file can lose an edit with no error
New conversation per task; let the file carry the context
Keep usage down
The most documented complaint: the weekly quota runs out faster than people expect. @bot reset quotas for everyone on 5 Sep 2026, and did the same on 26 Aug when access widened. Do not count on that every week.
What burns
Long, chatty CoS threads (the full history is resent)
A catch-all "General Helper"
Inbox triage against years of backlog on day one
Always-on routines
Browser clicks where a plugin would do
Event listeners like "every new Slack message"
What holds
One job per Bot
New conversations per task
Schedules during office hours
Plugin before browser
Grok Bot as the outer loop (research, drafts, delegation); heavy code jobs in Cursor
"If nothing: send nothing" in every scheduled Bot
Check Settings → Usage & Billing before large jobs
All Bots share one computer
The cloud computer is one per account, not per Bot. Logins, cookies, and files are shared. Each Bot gets its own screen so they can work in parallel, but screens are not a security boundary.
Never paste passwords or one-time codes into the chat. Take over Agent Computer and type yourself.
Plugins are account-wide: one login applies to all Bots.
Hide a Bot if you might need it. Delete removes profile, conversation, and routines, not files and sessions on the computer.
Do not show secrets in a Teach a task recording.
Common pitfalls
Multiple Bots ≠ isolation
One computer, all logins. Do not put a password on the computer that another Bot should not be able to use.
X credits run out quietly
A Bot watching X may report "nothing new" instead of an error.
Site layout changes
A scraper can report an empty scan for weeks and sound calm. Ask it to report its inputs, not only findings.
Hiding a Bot does not pause routines
Hide only removes it from the sidebar. Pause routines separately.
Delete leaves traces
Delete removes profile, conversation, and routines, but files and logins on the shared computer may remain.
Older blogs lie about platforms
Linux and Android exist now. iPad still does not.
Crypto-desk posts
Threads promising thousands a week lack verifiable receipts. Treat as ads.
Start small, iterate fast
Build the team week by week:
Week 1Create a Chief of Staff, pin it, write fences. One five-part task against a document, then one in a real tool.
Week 2Find a task you do manually at least 3 times a week. Run it, make it reliable, save as a skill.
Week 3Test run, then a weekday-morning routine. Keep write actions behind approval.
Week 4Add a specialist for what you still do manually. Ask CoS first.
Summary
Write tight personas with clear jobs and Don'ts
Once → reliable → skill → test run → routine
Require Approval for send, publish, buy, delete, change production
One bot per job, state in a file, new conversations per task
Office hours, plugin before browser, "if nothing: send nothing"
All Bots share one computer. Screens are not a security boundary
Roore AB · Stockholm · we build it for you
'Don'ts' is exactly how we build for companies
Deterministic where the answer should be the same every time. Prepared material and your yes where it must not proceed automatically. If you have an invoice, order, or email flow done by hand every week, we can build that first.