You do not need seven bots on your first morning. You need one useful result that you can check.
That might be an action list from a meeting, a short briefing from public sources, or a comparison between two small files. Start there. Make it work. Then decide whether it deserves to happen again.
This guide uses Grok Bot, not ordinary Grok chat. It is for people who want a working assistant without giving it the keys to the business. Allow about 30 minutes for the first exercise. That is a suggested learning slot, not a measured product benchmark.
The examples are invented. Do not practise with client portfolios, bank statements, passwords or confidential company documents.
What you are setting up
Grok Bot's documentation describes a persistent cloud computer with a browser, terminal and files. It can continue cloud work while your laptop is closed. That is useful, but it also means this is not an offline assistant whose information stays on your laptop.
Three words make the setup easier:
| Thing | Plain-English meaning | Our first example |
|---|---|---|
| Bot | The assistant you give a job to | Prep Assistant |
| Skill | A reusable method for doing the job | Turn a meeting note into a checked action list |
| Routine | A scheduled or supported event-triggered run | Prepare an approved briefing on weekday mornings |
The distinction between skills and routines comes from the vendor's guide. The examples are my suggested uses, not built-in job titles or evidence of a successful deployment.
Five minutes before you start
- Before signing in, review the account prerequisites. Grok Bot uses Cursor account/data settings, requires cloud storage and does not support Legacy Privacy Mode. If setup asks you to change a work account's privacy setting, stop and obtain approval. This does not mean every privacy mode or training opt-out must be disabled. Use the official download; access depends on your eligible plan and organisation.
- If this is company work, check that the service is approved before putting company information into it. You can do the invented exercises without connecting work accounts.
- Create one personal bot. Call it Prep Assistant. Its job is preparing drafts for you to review.
- Check that Auto Review is on and save Ask first rules for sending, publishing, spending, deletion, permissions and live-system changes. Set local execution to Never allow for these lessons. If you cannot inspect these controls or meet your organisation's policy, stop. Auto Review is model-based, not a safety guarantee; it supplements limited access. Do not choose Always allow to make a lesson run faster.
- Look at Usage & Billing. For these lessons, leave on-demand extra usage off where your account or administrator allows it. If it is on, inspect the monthly limit and actual usage: the billing guide says a run already underway can finish above that limit. Included weekly usage is a different allowance. A sentence in a prompt is not a billing cap.
Labels can change between versions. The relevant references are approvals, security and privacy and settings and notifications.
For these lessons, keep email, Slack, calendar, finance and client-data connectors unconnected. If an existing account already has connected services, do not assume this new bot is isolated from them: review that shared access first, or use an approved separate training account with no live connections.
Ready-to-copy: your first prompt
Copy this into the bot's conversation. It is a working brief, not a substitute for permissions.
If it asks to connect an account, send something or install a tool for this exercise, say no. The task does not need that access.
Exercise 1: Turn a note into actions
Here is a deliberately untidy meeting note. Ask the bot:
What a good result looks like: Jamie owns the agenda, due 12 October. Priya owns the availability check, with no supplied deadline. The tea order has no supplied owner or deadline. The next meeting is unresolved. Nobody has been contacted.
Open the original note beside the result. Check every row. A neat table with invented dates is a failed exercise, not a clever assistant.
Now deliberately challenge it: "You have not been given a deadline for Priya. What should you do?" The answer should be to preserve the gap and ask, not to make a sensible-looking guess.
Exercise 2: Prepare a small public briefing
Once the first exercise works, try a little research. Choose one topic you understand well enough to check. Public opening times or a product's official release notes are easier starting points than markets or tax.
Ask the bot:
Replace the placeholders before running it. Click the links yourself. Does the page actually say what the bot says? Does a proposed change appear as a proposal, rather than something already in force?
Pass: a short, source-linked draft that acknowledges gaps. Fail: plausible news with missing links, an old update presented as today's news, or a recommendation beyond the brief.
Public information can still be wrong, and a source page can contain instructions aimed at an agent. Those instructions are not permission to change your task.
Exercise 3: Spot a difference without fixing it
This is useful for operations work and a safe introduction to the wealth-management companion article. These are fictional values, not real holdings.
Ask the bot:
Check the arithmetic yourself: A totals £1,650; B totals £1,660. Cash and the total are each £10 higher in B. That is a discrepancy between records. It is not proof that anybody made or lost £10.
For real financial work, approved accounting or portfolio systems remain authoritative. Use a checked calculation method and a qualified reviewer. A language model's attractive explanation is not a reconciliation sign-off.
Test what happens when information is missing
Change the third exercise so Record B has no date or currency. Ask for the comparison again.
The bot should flag that it cannot establish whether the records are comparable. It may show arithmetic explicitly labelled as provisional, but it must not present a reconciled result.
This is worth testing. "I cannot complete this because the date is missing" is sometimes the most valuable answer you can get.
Save a method, then think about a routine
After two or three checked practice runs, ask the bot to help turn the method into a reusable skill. That is my suggested learning threshold, not a vendor guarantee that the task is now reliable.
Ask:
Saving a skill means retaining a method. Scheduling a routine means asking the product to run work later. Do not confuse the two.
Grok Bot documents scheduled routines and supported event triggers. Some triggers require separate integration setup. A connector being installed does not prove that a particular event trigger exists or is active.
A routine you can inspect before it runs
For a public briefing, this is a proposed specification. It does not enable anything by itself:
Ask the bot:
The office time zone should stay explicit when you travel. "Tomorrow morning" is not a sufficiently clear schedule.
Important: the vendor says Test run does real work. It is not a harmless simulation. Before using it, inspect the actual destinations and permissions and limit the run to safe sources and draft outputs. An automated draft can still read or write somewhere you did not intend if access has been left too broad.
Only enable a live routine after checking its saved settings and next-run time. Watch its first result. Find the pause control before you need it. The routine guide explains where to inspect, edit, test and pause routines, and notes that automations may pause after inactivity.
Keep a small result log
For each practice run, record the task, sources, errors, review time and visible usage. Do not invent a monetary saving from an account percentage.
The useful question is not "Did the bot finish?" It is "Did I get a correct result with less effort, after including the checking?"
If review takes longer than doing the job, narrow the task. If it keeps inventing owners, fix the method before scheduling it. If nobody knows who checks the result, do not automate it.
If something goes wrong
| Problem | What I would do |
|---|---|
| It wants to send a draft | Deny the action. Keep sending tools disconnected and inspect the approval settings. |
| It cannot open a source | Keep the gap visible. Do not let it invent the contents or substitute an older version silently. |
| It needs a login | Stop. Decide whether the connection is approved. Enter credentials yourself through the proper handoff, never in chat. |
| It repeats a run or creates duplicate output | Pause the routine. Inspect its history and run identifiers before restarting. |
| It gives a wrong total | Recalculate independently. Correct and retest the method before using real records. |
| A result arrives nowhere you expected | Pause it and check the saved destination, permissions, notifications and recent run history. |
Quick check
- The bot invents a due date to fill a blank. Is the exercise passed? No.
- A £10 difference appears between two records. Can you call it investment profit? No.
- You click Test run. Can it perform real actions? Yes. Check the scope first.
- You create a second personal bot. Are its browser credentials isolated? No.
- The first draft looks right. Should you connect every work system? No. Widen access only for a demonstrated need and with approval.
That is the starting point. One useful job. A result you can check. A way to stop it. Add the next job when the first one earns its place.
Take it into a real team
The companion article, If I Managed £150 Million: Building an AI-Assisted Investment Office, explores how narrow assistants could support wealth managers without taking investment authority.
You can also read how I would use Grok Bot as a leader, AI privacy in plain English and public first, private later.
Sources and notes
Product documentation checked on 8 October 2026. Grok Bot was not operationally tested for this guide, and no client data was used. This is vendor documentation, not an independent reliability test or a record of my own deployment. Interface labels, eligibility, billing and features can change.
- Getting started: installation, accounts and creating a first bot.
- Computer and apps: cloud execution, shared personal-bot access and connections.
- Skills, routines and automations: reuse, scheduling, real test runs and pausing.
- Approvals, security and privacy: actual controls, secrets, local execution and deletion boundaries.
- Settings and notifications: time zones and account usage.
- Plans and billing: included versus on-demand usage and the limits of monthly spending caps.
The exercises, learning times, task contract and pass/fail tests are my teaching proposals. This guide does not authorise handling confidential information, making financial decisions or activating a production workflow.