2026-08-27 · Guide
Your First Grok Bot in an Hour, With One Reversible Job
The first Grok Bot hour dies in the plugin picker: five names, a Gmail consent screen, and no brief you would show a colleague. Sixty minutes is enough to prove the product if you spend it on one reversible job, one named bot, and a stop that forbids send. A roster on day one is how people burn the week, and often the trial credit with it.
This page is that hour. Confirm you can open Grok Bot. Create one bot. Paste a charter with a boundary. Run once on a public page or an attached document. Inspect every claim. Stop. Do not connect Gmail. If you still need the plain definition, read what a Grok Bot is first, then start the clock.
The worked example is Lead Scout pointed at one public company. Sources required. Never contact. Inbox work waits. You are proving one job.
Spend the hour on a minute map so a roster cannot appear
Write the clock down before you open the app. If the hour is only a feeling, you will fill leftover minutes with a second bot, a plugin, or a routine, and you will call that progress. It is not. Progress in hour one is a sourced paragraph you can check, then a stop.
| Minutes | Move | You are done when |
|---|---|---|
| 0 to 8 | Confirm client and entitlement | Desktop app opens, and the plan actually includes Grok Bot |
| 8 to 15 | Create one named bot | The roster shows Lead Scout and nothing else |
| 15 to 28 | Paste the charter | The company, the source rule, and the never-contact line are in the text |
| 28 to 40 | Run once on one input | The run finished or reported blocked, and you did not stack new instructions |
| 40 to 55 | Inspect | You opened every URL next to a material claim |
| 55 to 60 | Stop | Gmail is still disconnected, no routine exists, no second bot exists |
Each row has one exit. If a row overruns, cut the next row rather than adding work. A late inspect is still an inspect. A late plugin is a different product.
Keep a local note for the three-line verdict at minute 55. Do not store that verdict only inside the bot. Deleting a bot deletes its routines. It does not delete the lesson, unless the lesson never left the screen.
Confirm the desktop app and an eligible plan before you name anything
Hour one starts when the client you actually use can create, run, and inspect a bot. Grok Bot supports macOS on Apple silicon and Intel, Windows on x64 and Arm64, and iPhone on iOS 18 or later. There are Linux desktop and Android apps as of September 2026, and the iOS app also runs on iPad (iPadOS 18 or later). The cloud computer is a managed Linux VM. That is not the Linux client on your laptop. If your only machine is a Linux workstation, install the Linux desktop app and this hour can start. Supported platforms is the current client list.
iPhone can pause and resume. Editing and testing a routine still need the desktop app; the phone can now show run history and delete a routine. Do not author the first Grok Bot on a phone.
Eligible plans, rechecked 2026-09-23, are every paid Cursor plan (Pro, Pro+, Ultra) and every self-serve Cursor Teams seat, plus a one-time trial. Cursor Hobby, the free plan, does not include Grok Bot, and an individual SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription can be linked instead. Cheapest documented individual paid door: Cursor Pro at $20 a month. Do not print a SuperGrok Heavy price. Cursor access is the sign-in story. Is Grok Bot worth it is the money question after a run.
If you are not on an eligible paid plan, you may be on the limited-usage trial from the 21 August 2026 expansion. Spend it like this hour: one bot, one job. The free trial page is how the credit behaves. Privacy Mode (Legacy) blocks Grok Bot entirely.
Open the app. Confirm the entitlement the app shows. Then create a bot.
Create one bot named after the hour's job and close the extra tabs
Name it Lead Scout. Not Assistant. Not Test 2. Not the company you are about to research. The name is the scope test. If you would not put it on an org chart, you have not chosen a job.
Create one. Close the new-bot flow. Do not open a second draft "for later." Later is how hour one becomes a roster. All bots on an account share one persistent cloud computer assigned to you, not to a bot. Each bot gets a screen. Screens are not security boundaries. Extra names do not isolate Gmail, cookies, files, or CLI credentials. They only multiply things you failed to inspect. One computer, many screens is the architecture.
A routine assigns a workflow to one bot, max 50 per bot, 20 recent run records kept. Deleting a bot deletes its routines. Do not create a routine in this hour. Scheduling exists for a later day. Teach-by-demonstration (ten minutes, browser only, draft skill, not on iPhone) is also a skip. Hour one is a pasted charter and one run.
If you already created three bots this morning, ignore the extras. The hour is Lead Scout only.
Paste a charter that names one company, one source rule, and one stop
The charter is the bot. Write four things before you run: the job in one sentence, the company, the source rule, and the action that never happens.
Hour-one company: Notion. Not your CRM. Not eighteen accounts. One public company you do not need to email today. You are not generating pipeline. You are proving that a named Grok Bot can read a public page, cite it, and stop.
Paste this. Change the company name later if you must. Do not expand it.
Name: Lead Scout
Job: One public company, sourced brief, no contact
You research Notion for a first-pass lead brief. Use only public pages that
open without a login. Start at notion.com. You may also open the public
careers page and one public blog or changelog page if they load. If a page
asks you to sign in, write blocked and stop that thread.
Every material claim has a URL. Quote a short phrase from the page next to
the claim. If you cannot find a fact, write unknown. Do not guess prices,
headcount, or product names from memory.
Deliver a one-page brief:
- What they sell, in two sentences, with URLs
- Who they appear to hire for, with careers URLs, or unknown
- One public signal from the last year if you find it, else unknown
- A source table: claim, quote, URL
- Suggested first line: one sentence I might use later. Never sent.
Boundary: Never contact anyone. Never DM, email, follow, like, comment, or
create an account. Never sign into Notion, Gmail, X, or any other site.
Never use cookies or sessions left by other bots. Never send. Never purchase.
Stop when the one-pager is done or when you are blocked. Do not keep
browsing to look busy. Do not research a second company.
That is a narrowed hour-one version of catalog Lead Scout: research and ranking, never contact. The catalog job can grow to a sheet. Hour one cannot. If you need "and" in the first sentence of the charter, you are describing two bots.
The stop line is not decoration. For this hour the gap is outreach. Name it. Approvals and reversibility is the longer rule: an approval gates the next action. It does not undo work already done. A never-contact line is how you avoid needing an undo you do not have.
Give the run a single public page or a single attached document
Safe input means you could reconstruct the run without the bot. Two shapes qualify.
A public page: notion.com, opened in your own browser in another tab, so you can grade the brief against the same HTML. A careers page counts. A login wall does not. If the bot hits a login wall, the correct output is blocked, not a guessed paragraph.
An attached document: a PDF you saved from a public page, or a short public report you already have. Attach one file. If the fact is not in the file, the bot writes unknown.
Do not point hour one at your inbox, your drive, your CRM, or a list of twenty domains. Those inputs require connections or they invite a crawl. There is no Grok Bot-specific spend cap, but the account-level On-demand monthly limit applies. Subscriptions include a weekly usage allowance, then overflow billed from model and token cost. A wandering browse is how a trial and a weekly allowance both vanish. Spend and token burn is the later control.
Name the URL in the first message even though the charter already names the company. Do not sign into the target site "just to see pricing." That cookie lands on the shared computer.
Let the first run finish or fail in the open, then halt it
Start one run. Watch it. Do not type a second assignment while the first is in motion. Stacked instructions are how you get a brief about Notion and a side quest about a competitor you never named.
If the bot asks to connect a plugin, the answer is no. If it asks to sign in, the answer is no. If it keeps opening tabs after the one-pager exists, stop the run. Looking busy is not a deliverable.
A finished run that says blocked on the careers page is a successful hour: the charter reported a limit. A finished run with unsourced claims is also successful, because inspect will catch it. An unfinished run you abandoned to go create Inbox Triage is a failed hour.
Grok Bot has no model picker. Admins cannot pick one either. Fix the job, not a dropdown that does not exist. Individual accounts and self-serve Teams still have no audit view of Bot actions; Enterprise has audit logs and Action Recording. Inspect is your audit. Stay for the run.
When it stops, do not immediately rerun. Inspect first. If the miss is a missing source rule, you may rerun once. Twice is a different project.
Open every cited URL and mark the claims that cannot be found
Inspection is the product test. The bot is not proven because the prose sounds like a brief. It is proven if a human with the same pages can tick each material claim.
For every sentence that could embarrass you on a call, open the URL next to it. If there is no URL, mark fail. If the URL loads and the quote is not on the page, mark fail. If the sentence is tone ("they seem innovative"), ignore it. Hour one is not grading vibes.
| Claim in the brief | Pass | Fail |
|---|---|---|
| What they sell | Two sentences you can match to the homepage or a product page | A feature list with no URL, or a feature the page does not state |
| Hiring | A role family quoted from a public careers page | Headcount, team size, or "they are hiring aggressively" with no posting |
| Price | A number that appears on a public pricing page you opened | "Enterprise pricing" invented from training data |
| Recent signal | A dated public post, changelog, or blog URL | A rumor, a private community, or "I recall" |
| Suggested first line | One sentence, clearly labeled unsent | A drafted email, DM, or form submission |
| Contact | Empty. Nobody was messaged. | Any send, follow, like, comment, or account creation |
Three or more fails and the charter is not ready to keep. Honest unknown is a pass. Confident fails become your voice if you ever widen toward real leads. Catalog Lead Scout uses the same rule: every signal carries a link, and the bot never contacts anyone. If you cannot grade one company, you will not grade a sheet of twenty tomorrow.
Write the fails on the note. They become the three-line verdict. This hour's job is to see the miss.
Keep Gmail disconnected for the whole sixty minutes
Gmail is the upgrade that feels like using the real product. It is also how hour one stops being reversible. Signing in puts a mailbox session on the one computer every bot on the account shares. Deleting Lead Scout does not log that session out. Hosted MCP tokens stay with Cursor's backend. Browser cookies do not. Grok Bot and Gmail is not this hour.
Draft-only is a sentence in a charter. It is not a permission wall. A bot that can see mail can be instructed, including by a message in the mailbox, to send. You do not need that shape in minute twenty.
Inbox Triage is the catalog mail bot: sort, draft, never send. It is a good second job after a research run you trust and after the safety checklist. It is a bad first Grok Bot because renaming bots does not isolate a mail session. Least privilege says connect the minimum. For this hour the minimum is no mail plugin.
If the UI offers a bundle of plugins, decline. Readiness is a graded brief, not a green row of logos. Same rule for Slack, a bank, a CRM, and calendar write. Chief of Staff Briefing can wait. It still should not send. It still should not be bot number two in minute fifty.
Treat reversible as "no one was contacted," not "I can delete a file"
People hear reversible and think of files. Delete the draft, delete the bot, the world is as it was. That is true for a local paragraph. It is false for outreach, a purchase, a calendar invite, a posted comment, and a signed-in session left on a shared VM.
Hour one's reversibility test is social, then credential, then file:
| Action | Reversible in hour one? | Why |
|---|---|---|
| Write a brief in the bot's output | Yes | You can ignore it |
| Save a notes file on the shared computer | Mostly | The file may remain after you delete the bot |
| Sign into Gmail | No | The session remains for every bot |
| Email or DM a stranger | No | You cannot unsay it |
| Follow, like, or comment as research | No | It is contact with a public receipt |
| Place an order or create an account | No | Money and identity moved |
| Schedule a routine that keeps running | No | Allowance spends while you are away |
Approvals do not reverse completed work. If you only approve "send" and the bot already created an account two steps earlier, denying send leaves the account. For Lead Scout the irreversible step is contact. Design the job so that step never starts.
Copy any file the bot wrote into your notes if you want. Do not spend minute 58 inventing a cleanup bot. If you are tempted to "test send to myself," you have left the hour. The test that belongs here is: did anyone outside this room hear from the bot? The answer must be no.
Block the three upgrades that turn a first hour into a lost week
The hour fails in three predictable upgrades. Each one feels like competence. Each one is how people describe Grok Bot as busy instead of useful.
| Upgrade | What you tell yourself | What actually happens |
|---|---|---|
| Hire a second bot before inspect | Cover more of the week | Two ungraded streams, same computer, same cookies |
| Connect Gmail or Slack | Make it real | A standing session and a blast radius you have not earned |
| Add a routine | So it works tomorrow | Spend while you still cannot say if today's brief was true |
There is a fourth cousin: group chat between bots. Hour one has one speaker. If you need a router, you do not have a first job yet.
Refuse the upgrades in your hands. When the run ends, you have fifteen minutes. Those minutes are inspect, not staffing. If a colleague wants a demo of "the whole OS," show them the Notion brief and the fail marks. A sidebar full of names is a costume. Lead Scout can grow a sheet on Tuesday. Inbox Triage can wait.
Keep this hour off the full install walkthrough and off the seven-day plan
This page is sixty minutes and one job. Other pages cover longer arcs. Mixing them is how a reader spends three hours and still has no inspect.
The full setup walkthrough is account, runtime, connections, and a first working bot as a wider sequence. Budget for that is closer to ninety minutes, and it includes decisions this hour refuses: which tools to connect, how to prove a schedule fired. Use that walkthrough another day. Do not paste its connection chapter into minute ten of this one.
The first-week plan is seven days: one draft-only bot, a real review habit, a tighter charter, a second bot in a different lane on day four, then a verdict. That plan is correct for a week. It is wrong for an hour. Day four in that plan is a second hire. Minute forty in this plan is still inspect.
Use this hour when you have never run Grok Bot, or when you ran it once as a tour and cannot remember what it produced. Use the week plan when this hour already produced a brief you would reuse. Use the setup walkthrough when the client is not even installed. If you arrived from a comparison page, you still start here. Hour one needs a URL that matches a sentence, not a market map.
Answer the claim that a no-send first Grok Bot is only a toy
The strongest objection is honest: if the bot cannot mail, post, or update a CRM, you have not used the product. You have used a summarizer with extra steps. Why spend an eligible plan or a trial on that?
Because the product is not send. The product is a named bot on a shared cloud computer that can use a browser, files, and tools you later connect, on a schedule you later assign, with no model picker and no Grok Bot-specific spend cap. Send is a permission you add. It is not the definition. If the computer cannot produce a sourced Notion brief you would trust, connecting Gmail will not fix that. It will only attach a mailbox to the failure.
A toy is a run you cannot grade. A real first Grok Bot is a run you can fail. The inspect table can fail. That is the point. People who start with send never get a clean fail. They get a customer-facing maybe, then a week of watching the bot instead of working.
The objection wins in one case: you already ran this exact shape last week, the briefs were good, and you intend to keep an eligible plan. Then hour one is over, and mail may be the next slice if drafts stay drafts. That is not most people who search for a first Grok Bot. Most people have a plugin list and no inspected paragraph.
Claude Code, SKILL.md, and CLAUDE.md compatibility is Grok Build, not Grok Bot. Do not spend the hour dropping a repo of skills into this product. The bot in front of you reads the charter you pasted.
Write a three-line verdict before you add a schedule
Before you close the app, write three lines in a file you own:
What it got right. Name one claim you verified on a page.
What it invented or left unsourced. Name one fail from the inspect table.
Whether you would run the same charter tomorrow on a different public company. Yes, yes with a tighter source rule, or no.
That verdict is the only artifact the hour must produce besides the brief. If you skip it, tomorrow you will remember that Grok Bot "seemed fine" and you will connect Gmail on that memory.
If the answer is no, do not upgrade the plan to punish a charter. Change the company, attach a document, or stop. Worth it is a question for a graded job, not for a mood.
If the answer is yes, still do not add a routine in the leftover minutes. Run the same bot once more on a second public company on another day, by hand. Two inspected runs beat fourteen unattended ones.
Do not hire Inbox Triage on the strength of one Notion page. Hire it when a bad mail run is still "a draft I delete," not "a customer received a sentence."
Leave cookies, files, and sessions untouched, then quit the app
End the hour by not tidying. Tidying is how people sign out of the wrong thing, delete a file another bot needed, or create an overwatch bot at minute 59.
You signed into nothing extra. Leave it that way. If the bot saved a brief on the shared computer, copy the text into your notes. Deleting the bot will not reliably clean the disk. Do not delete the bot tonight to reset.
Quit the desktop app. Do not keep a run going overnight. Overnight is a routine. The weekly allowance does not pause because you are curious.
If you used the trial, write down that you connected nothing. If you already had plugins from an earlier experiment, treat the Notion brief as evidence about research quality only. Revocation belongs on the safety checklist, on a day that is not this one.
You now have a first Grok Bot or you have a documented miss. Either one is an hour well spent. A sidebar of unnamed assistants is not.
Keep reading: What Is a Grok Bot? The Plain Explanation for Non-Engineers, Grok Bot Free Trial: How the Metered Credit Actually Works, The Grok Bot Safety Checklist Before You Connect Your Inbox.
Frequently Asked Questions
Can I finish a first Grok Bot in sixty minutes if I already pay for an eligible plan?
Yes, if the hour stays one job. Confirm a desktop client (Mac, Windows or Linux) opens and that the entitlement is Grok Bot, not Cursor Hobby or a SuperGrok tier that cannot be linked. Create one bot named Lead Scout, paste a charter with a never-contact line, and run it once on a public page or an attached public document. Inspect every sourced claim. Stop without Gmail, without a routine, and without a second hire. The phone app can pause a routine, approve steps, and show run history but not edit it, so the hour does not start until you can edit and inspect on a desktop.
Should I connect Gmail in the first hour if the bot will only draft?
No. Hour one proves that a named bot can produce checkable work on a safe input. Signing into Gmail stores a mailbox session on the one cloud computer every bot on the account shares. Deleting the bot does not remove that session. Draft-only is a charter sentence, not a security boundary, and there is no audit view of bot actions outside Enterprise. Connect mail later, after a brief you trust and after you understand how sessions persist. The first Grok Bot should still be sourced research, not an inbox.
Why start with Lead Scout instead of an inbox bot?
Inbox work is the pain people feel, which is why they start there. It also demands a mail login on a shared computer and puts send one instruction away. Lead Scout on a single public company needs no plugin, returns a one-page brief you can verify in another tab, and has a stop you can check in one question: did anyone get contacted. If that brief is thin or invented, you learned the product failed the only test that matters in sixty minutes. If it is sourced, you earned the right to consider mail in a later hour.
What should I do after a successful first hour?
Write three lines you keep outside the app: what the bot got right, what it invented, and whether you would run the same charter tomorrow. Leave the bot named Lead Scout. Do not add a routine yet. Routines belong to one bot, vanish when the bot is deleted, and spend weekly allowance while you are still deciding. Do not hire a second bot until you have inspected a second run of this one. If the brief was not useful, do not buy a higher plan to fix a charter. Change the company, tighten the source rule, or stop.