Hello everyone, I'm Brother G. As the ultimate personal assistant, we shouldn't just know how to use Grok Bot; we must use it to generate income. Today, we'll start from zero, beginning with the software installation, and break it down for you step-by-step. Bookmark this and study it carefully!
From Installing Grok Bot to Selling Services to Overseas Clients
The first time I saw the introduction to Grok Bot, the thought in my head was simple: isn't this just Grok with a different shell?
It has Grok in the name, and the interface has a chat box. After opening it, I still have to type, and it still replies. Looking only at these, it's easy to mistake it for just another chatbot.
But if it's just for chatting, why does it need a cloud computer? Why does Cursor pop up during login? Why can one account create many Bots, each requiring a name, role, and job description?
Following these questions, I realized I had misidentified it from the start.
Ordinary AI's basic action is "answering." If I ask how to organize a weekly report, it gives me a method; if I ask it to write an email, it hands me text; if I ask where to click next, it tells me the path. The person actually clicking the webpage, downloading files, logging into the backend, and checking spreadsheets is still me.
Grok Bot's basic action is "taking over." Once I clarify the task, materials, boundaries, and delivery format, it can open webpages, process files, and use the terminal to continue working within the applications I allow. When it encounters checkpoints like logins, CAPTCHAs, payments, sending, or deleting, it stops to find me.
That's the difference.
I treat ordinary Grok as a consultant and Grok Bot as a colleague sitting at another computer. The consultant tells me how to do it; the colleague takes the task and pushes the work to a point where it can be accepted.
This long article revolves around a specific case: I create an "AI Intelligence Officer" to collect AI industry updates weekly, verify information, and organize it into a Chinese weekly report, but not publish it itself. I first use a 299 RMB small order to verify delivery, then change the service to three tiers of $300, $800, and $1,200 that overseas clients can understand. That final $5,000 bill must be built starting from the first read-only task.
If you've never touched Agents, Skills, or Routines, this example is enough. Run one role successfully before talking about more tricks.
First, Separate Four Easily Confused Things
Grok, Grok Build, and Grok Bot have similar names but do different things.
Grok is a general AI assistant. Q&A, writing, analysis, and image creation are its home turf.
Grok Build is more programming-oriented. It faces code repositories, terminals, and development tasks, suitable for writing code, modifying projects, running commands, and troubleshooting errors.
Grok Bot is for complete cross-tool workflows. It can go to webpages for information, organize results in files, and then continue processing in another application. Programming is just one of the jobs it can do, not its entirety.
Another common misconception: don't treat Grok Bot as an open-source project where you need to install Python, pull source code, and configure models yourself. You use the official client and cloud computer. If "locally deployed versions," "cracked APKs," or "one-click private Grok Bots" appear online, don't get excited. Same product name doesn't mean they are the same thing.
When I judge which task to give to whom, I use a simple but effective method.
If I just want an answer, I look for Grok.
If the core of the work is a code repository, I prioritize Grok Build.
If the task requires multiple steps across webpages, files, and apps, I look for Grok Bot.
This judgment blocks a lot of useless effort. Using Grok Bot to write an advertising slogan is possible, but it's like bringing in a whole workbench just to slice an apple.
That "Computer That Never Shuts Down" is the Key
The reason Grok Bot can take over work isn't the chat box, but the persistent cloud computer behind it.
This computer has a browser, file system, and terminal, and can use available connectors, Plugins, and MCP. After I close my laptop, tasks already started can continue running. It only pauses when it hits a step requiring judgment or authorization.
"Persistent" is more critical than "cloud."
Ordinary chats are often like one-off meetings. Opening a new conversation, I have to explain where the files are, what the format is, and which websites are prioritized. A Bot has its own name, responsibilities, sessions, and saved work preferences. The longer the same Bot works, the easier it is to form a stable context.
But persistence has a price.
Multiple Bots under the same account share one cloud computer. Files, browser login states, application sessions, and some credentials might be seen by these Bots collectively. Each Bot can have its own screen and work in parallel, but they don't have independent security boundaries.
I can't assume that because I built a "Finance Bot" and a "Content Bot," they are separated by walls. As long as login info and files are on the shared computer, other Bots might use them. Categorizing roles organizes context but doesn't isolate sensitive data.
Check Eligibility Before Downloading, Don't Fight the Installer
Many beginners waste time here. The software is installed, but after logging in, it says there's no permission, so they repeatedly uninstall, reinstall, and change networks. The problem is usually not the installer, but the account eligibility.
As of this tutorial, Grok Bot is included in these personal or self-service team plans:
- SuperGrok, SuperGrok Plus, SuperGrok Heavy
- Cursor Pro, Cursor Pro+, Cursor Ultra
- Cursor Teams Standard and Premium
The product is adjusting rapidly; eligibility pages and help docs occasionally sync slowly. Logging into your account to check Grok Bot Access, plan pages, or real-time results from the client is more reliable than memorizing a list.
Free accounts might see a trial entry, but I wouldn't treat trials as a long-term solution. Trial quotas, weekly usage, and open scope can change. Before officially putting it into a workflow, I confirm two things: whether the current plan includes access and how much usage the task will consume.
Grok Bot has its own usage quota, calculated separately from regular Grok or Cursor quotas. Its consumption isn't simply understood by "how many messages sent." A task that spans multiple pages, processes many files, and verifies repeatedly can be more expensive than ten short conversations.
There's also a setting that might trip up old Cursor users: Grok Bot needs cloud storage. If the account is still using Legacy Privacy Mode, the client might not start. If you hit an error containing this phrase, check the data options in Cursor's privacy settings first; don't keep messing with the installer. Team accounts might need an admin to handle this.
Why Did I Buy SuperGrok but See Cursor During Login?
This is probably the most confusing moment for beginners.
I clearly got usage eligibility from Grok, but after opening Grok Bot, the authentication page shows Cursor. Many people suspect they downloaded the wrong software or think they need to buy another Cursor subscription.
No need to pay a second time.
Grok Bot currently uses the Cursor account system for login and some settings. When you have an eligible SuperGrok plan, you can link the Grok account at the entry point and complete authorization as prompted. Seeing Cursor doesn't mean you must buy a Cursor membership separately.
If it still shows no eligibility after logging in, I troubleshoot in this order:
- Confirm the SuperGrok or Cursor plan is active.
- Confirm the browser and client are using the same corresponding account.
- Log out of Grok Bot and log back in.
- Go back to the Access page to re-check eligibility.
- Team accounts confirm SSO and admin configuration.
- If still not recognized, go through official support channels.
I won't hand over passwords, verification codes, or session cookies through shady "recharge portals." Even if a service only asks for a User ID, I check the plan, validity, refund, and after-sales first. The few minutes saved aren't worth gambling the whole account.
Installation Isn't Hard; The Hard Part is Not Choosing the Wrong Version
Official clients currently cover macOS, Windows, and Linux. The iPhone version requires iOS 18 or higher; you can view Bots, continue conversations, upload images or files, receive approval reminders, and take over the cloud computer when necessary. Whether Android is open depends on the official list at the time; don't install strange APKs.
The download entry only recognizes the official Grok Bot page. The real difference in installers is mainly system and chip architecture.
If Using Mac
Click the Apple icon in the top left, open "About This Mac."
If you see "Chip" and M1, M2, M3, M4, or subsequent models, choose Apple silicon. If you see "Processor" and Intel, choose Intel.
After downloading, open the disk image, drag Grok Bot into Applications, and launch it from "Applications." If macOS pops up a security confirmation, check the app's publisher and download address before clicking "Open."
If you choose the wrong chip version, the app might not open at all. This failure isn't mysterious; go back to "About This Mac," look again, and download the correct version.
If Using Windows
Open "Settings → System → About," and check "System type." Most Intel or AMD computers use x64; ARM Windows computers choose Arm64. Don't guess the architecture by brand; the same brand can sell both types of machines.
Download the corresponding installer, run the installation, and open it from the Start menu. The client will automatically check for updates, or you can manually check in "Settings → Beta → Check for Updates."
If Using Linux
Stable versions provide Linux builds. In the "More downloads" section of the download page, Debian and Ubuntu usually choose .deb, Fedora and RHEL choose .rpm, and other distributions can consider AppImage.
Run uname -m in the terminal; if you see x86_64, choose x64; if you see aarch64, choose Arm64. .deb and .rpm are installed with the system package manager; AppImage needs execution permissions first.
Old tutorials might still say "Linux not supported." Products update quickly; don't rely on memory for platform info; follow what the download page shows that day.
First Launch: Let the Cloud Computer Prepare Itself
Open the client, and you'll see "Get started." After clicking, the browser pops up an authentication page. Complete the login in the browser, then return to Grok Bot.
If it stays stuck for a long time, then I start troubleshooting:
- Completely exit the client and reopen it.
- Check for and install new versions.
- Open Agent Computer to see the actual status.
- Try Retry or Recover.
- Save "Reset Agent Computer" for last.
Reset isn't a "universal refresh button." It might cause recent unsynced work to be lost. The more a button looks like a one-click fix, the more you need to see the consequences. I'd rather wait two more minutes than gamble with existing files.
Initial guidance will also ask which tools I commonly use. This step is mainly for recommending suitable Bots; it won't automatically log in or modify data just because I checked an app. You still need to authorize separately when actually connecting tools.
I Didn't Create a "Universal Assistant"; I Gave It a Narrow, Almost Boring Job
Once the cloud computer is ready, I can finally create my first Bot.
The entry might appear in "Meet a future teammate," or you can click "New" in the sidebar and select "Create new agent." The system first generates a New Agent, then I open the Bot menu and enter "Edit Profile."
When creating the profile, I first write the Name, Job, and Description; the avatar can wait.
Name is the name, keep it short to find it easily in the list later.
Job is the primary role, write only one core task.
Description is the long-term work instruction, writing the rules it should always follow.
I call this demo Bot "AI Intelligence Officer." The Job is "AI Industry Information Research." The Description is written like this:
1Continuously monitor public AI industry updates, prioritizing official product websites, official documentation, and official accounts. Organize results in Chinese, keeping dates and original links, and separate confirmed facts, unverified information, and my judgments. Clearly write "Not found" when there is no page evidence; do not guess. Any operation involving sending, publishing, purchasing, deleting, overwriting, modifying permissions, or accepting agreements must stop and ask for my confirmation first.
This description doesn't look fancy, even a bit wordy. But it's much more useful than "You are a world-class AI assistant, help me complete all work."
Universal assistants' problem isn't lack of ability, but blurry role boundaries. Finding news today, organizing invoices tomorrow, and modifying the website the day after. Files, preferences, history, and judgment standards all get mixed in one Bot; the longer it accumulates, the louder the noise.
Specialized Bots more easily form stable habits. The Intelligence Officer only researches, the Content Assistant only drafts, and the Data Bot only looks at reports. When tasks really need collaboration, let them pass the context.
I put long-term rules in the Description and leave specific tasks for this week in the messages. This division is very important.
Description answers "How do you always work?"
Message answers "What are you doing this time?"
If I write "Focus on a certain press conference this week" into the long-term profile, it might still hold onto expired requirements next week. Changing themes, dates, budgets, and completed lists should be put in the current message or saved in a file it re-reads before each run.
For the First Task, I Only Let It Read, Not Act
The most common mistake beginners make is immediately letting the Bot mass-send emails, batch-modify backends, or automatically buy or delete files as soon as they see it can operate webpages.
I do the opposite.
The first task should ideally meet three conditions: results are easy to check, read-only (no modifications), and finishes in five to ten minutes. Uploading a regular PDF and having it extract dates and to-dos is a very suitable test.
I'll send this directly:
1Please read the file I uploaded and complete the following tasks.2Summarize the main content in 5 key points; separately list all dates, decisions, responsible persons, to-dos, and unresolved issues; mark the corresponding page number or chapter for each item.3Separate "facts in the file," "your speculations," and "your suggestions." For content not explained in the materials, directly mark "Not explained in materials," do not fill it in.4Finally, return the summary, key information table, to-do list, and issues still needing confirmation.5Do not modify the original file, do not log into other websites, and do not send anything out. When encountering login, sending, overwriting, or deleting operations, stop and ask me first.
There's no mystery in this task. It contains five things: what result I want, where to read, what cannot be done, what to deliver finally, and where it must stop.
I remember it as a very practical formula:
Result + Data Entry + Constraints + Deliverables + Approval Points
"Help me research AI" only has a theme, no finish line. The Bot doesn't know how long to look, which websites are credible, or whether to hand over a paragraph or a table. It can only guess. The more it guesses, the more the result looks like a lottery.
"Organize important updates from the last 7 days, prioritizing specified official websites and accounts, select 10 items, keep dates, links, and core changes, and finally give me 3 topics, do not publish" is much clearer.
I don't need to learn complex Prompt Engineering to use it well. I just need to write a task list like I would for a new colleague, without hiding key conditions in my head.
How My Weekly Report Task Should Be Written
Once the file task passes, I'll put the AI Intelligence Officer on public webpages.
This time I'll send it like this:
1Help me organize important AI industry updates from the past week.2Prioritize checking official websites, documentation, and accounts of xAI, OpenAI, Anthropic, and Google DeepMind. Find the 10 most noteworthy pieces of information, each containing the event, release time, original link, core changes, and its potential impact on ordinary users or practitioners.3Do not treat search result snippets as evidence. Before adding to the weekly report, you must open the original page. Do not estimate numbers, dates, or versions not clearly stated on the page. When different pages conflict, list the conflicts side-by-side.4Finally, give me 3 topics suitable for long Chinese articles, each with a title and angle.5Do not publish, forward, send, or log into paid accounts. Only hand over the results and unresolved issues for my confirmation.
The most valuable sentence here is "Do not treat search snippets as evidence."
Search snippets might be outdated or cut off qualifying conditions. A page title matching doesn't mean the version, date, level, or duration are correct. The Bot can use search results to find candidate information, but it must open the original page before writing it into the final delivery.
I also require it to keep excluded candidates and the reasons for exclusion. This way I can see if it filtered according to standards or just mechanically pieced together the first ten search results.
How to Check if It "Finished" Instead of "Writing Like It Finished"
AI is very good at arranging answers neatly. Tables have titles, sentences are smooth, and links look real. When people see this level of completion, it's easy to lower their guard.
Don't accept based on layout; accept based on evidence.
Randomly open three original links to check titles, dates, versions, and key numbers. See if it used search snippets instead of original pages, or if it filled "Not listed" with a seemingly reasonable estimate. Finally, recalculate totals, durations, or ratios yourself.
A typical small error is the page saying "Basic" and the Bot changing it to "Beginner." In Chinese, both can be translated as "入门," looking almost identical, but if the task requires keeping original page tags, it's wrong.
Duration also often has issues. If a page shows both "about 2 hours" and "1 to 3 hours," the Bot might directly take the upper limit of 3 hours as a fixed value. The correct way is to keep both fields and state where each came from; when doing budget verification, use the upper limit as previously agreed.
When giving feedback, don't write "be more careful." This sentence has no executable standard. Write like this:
1Re-open these three pages and only verify the title, date, version, and original link. Keep other content unchanged. For fields not listed on the page, write "Not listed," do not replace with synonymous tags, and do not estimate.
The more specific the feedback, the easier it is for the Bot to settle it into long-term preferences. Saying "you did it wrong" only provides emotion; saying "which item is wrong, which page to follow, and which parts not to move during redo" provides a correction path.
Agent Computer is very useful at this step. It lets me see what pages the Bot opened, if it's stuck at login, and if it's waiting for approval. But the computer screen only proves it visited; it doesn't prove the conclusion is correct. That final layer of verification is still for me to do.
How I Pushed a Small Service to $5,000 Monthly Income
Clients only care about three things: what I deliver, when I deliver it, and who is responsible if something goes wrong.
I write the product as an acceptance checklist. One fixed day a week, I deliver 10 industry updates verified against original pages, 3 writable topics, 1 page of summary, and attached unconfirmed issues. The quote doesn't say "fully automatic" or "one-click generation"; manual review, client communication, and final sign-off are still in my hands.
I Earn the First 299 First
When starting out, I have no cases, reviews, or client lists. Writing the monthly fee as $1,200 directly gives people no reason to trust me. I first do a 299 RMB lightweight Chinese weekly report: 10 updates, 3 topics, 1 page summary, one revision, only delivering the document, no posting, and no logging into client backends.
For the first round of RMB quotes, I write the tiers as 299, 599, and 999 RMB. 299 RMB delivers the lightweight report, 599 RMB adds short post drafts, and 999 RMB adds thematic comparisons. They are responsible for testing which results clients are more willing to buy; I only recalculate the USD monthly fee after running through this.
299 RMB is used to verify three questions: Can the client understand it? Which columns are they willing to keep? Does reviewing one take me two hours or five hours?
Assuming one report takes me two hours, the gross hourly rate is about 299 ÷ 2 = 149.5 RMB. If it takes five hours, the gross hourly rate drops to 299 ÷ 5 = 59.8 RMB. I haven't even deducted subscription fees, communication, and rework. This number looks bad, but it forces me to delete columns the client doesn't need and save repetitive actions as Skills.
I Changed the RMB Small Orders to Three Tiers of USD Services
After getting presentable samples and feedback, I look for overseas clients. The target clients aren't "big companies"; I only look for small teams whose business I can understand, such as AI tools, cross-border software, independent developer products, and small consulting agencies. The narrower the industry, the easier it is for me to judge which updates will affect the client, and the easier it is to find if the Bot caught an old version.
The first tier is a $300 one-time research sprint. I verify 20 pieces of information within the agreed scope, organize a comparison table, and give 3 content angles. Delivered once, no monthly retainer. New clients can try once, and I can judge if both parties are suitable for long-term cooperation.
The second tier is an $800 monthly weekly report service. I deliver a research brief every week, four per month, including important changes, original pages, content topics, and issues to be confirmed. The client is responsible for publishing; I am responsible for research and drafts.
The third tier is a $1,200 monthly research plus content package. Besides the weekly report, I deliver 8 editable short posts or 2 long article outlines. The extra $400 buys topic judgment, content rewriting, and extra verification, not just letting the Bot run a few more times.
How Many Clients Does $5,000 Need?
The target mix is:
- 4 clients at $800/month, totaling $3,200
- 1 client at $1,200/month
- 2 research sprints at $300 each per month, totaling $600
The formula is 4 × 800 + 1 × 1200 + 2 × 300 = $5,000.
This represents 5 long-term clients and 2 projects for the month. Calculating 8 hours of review per month for each $800 client, 14 hours for the $1,200 client, and 4 hours for each research sprint, I reserve 4 × 8 + 14 + 2 × 4 = 54 hours for delivery. Adding 20 hours for finding clients, meetings, and revisions, it's about 74 hours a month.
I Spent Three Months Climbing This Slope
In the first month, I only focus on the first bit of money. Make a sample, contact 30 people related to the topic, and strive to get 1 to 2 test clients. If monthly revenue reaches $300 to $600, it means someone is willing to pay for delivery. If no one buys, I change the field and sample first, not buy more tools to feel brave.
In the second month, I strive to convert one-time clients into monthly clients. The goal is 2 clients at $800 plus 2 research sprints at $300, planning for $2,200 revenue. The most time-consuming work at this stage is often modifying service scope, organizing client feedback, and refusing temporary extra work.
In the third month, I push the client mix to 5 long-term clients and 2 short projects, targeting $5,000 revenue. Once the number of clients goes up, I'd rather pause taking orders than let one Bot mix client data. Each client has a separate folder, task instructions, and acceptance checklist; operations involving login and publishing continue to be approved by me.
My 7-Day Startup Schedule
Day 1: Choose a niche where you can judge right from wrong, and list 5 fixed pages.
Day 2: Write down weekly report columns, word count, delivery time, and things that cannot be done.
Day 3: Create a specialized Bot and run the first version using public webpages.
Day 4: Randomly verify 3 pieces of information, fix dates, versions, links, and repetitive issues.
Day 5: Save the qualified process as a Skill, then switch to another theme and rerun.
Day 6: Organize a sample without client data, and list a test price of 299 RMB or $49.
Day 7: Send the sample to people who might actually need it, only asking: "Did this weekly report save you time looking for info?" If someone pays, talk about monthly fees; if no one buys, change the topic and delivery, don't rush to add more Bots.
Grok Bot compresses collecting, organizing, and typesetting for me; I still have to choose the field, set standards, verify facts, and find clients. $5,000 comes from 7 paid relationships with clear boundaries; it won't grow automatically from a chat box.
Login, CAPTCHA, and Take Over: When It's My Turn, I Do It Myself
When the Bot enters a website requiring authentication, it might ask me to take over the cloud computer.
Open Agent Computer in the conversation, click Take Over, and enter the password, Passkey, two-factor code, or complete the CAPTCHA yourself. Confirm you've entered the page after login, then return control to the Bot to let it continue.
Passwords and one-time codes cannot be sent in the regular chat box. Chat content enters the session context, and the risk is completely different from entering it directly on the login page.
I only take over the blocked step; I won't conveniently finish the rest of the task for it. After taking over, it should continue along the original task. If it forgets the context, I'll add: "Login complete, continue reading data, only do analysis, do not modify the backend."
Login sessions are saved on the shared cloud computer, and other Bots might use them. After the task ends, if the account is sensitive or no longer needed in the future, I'll log out and revoke authorization. Deleting a Bot doesn't guarantee clearing files and sessions on the shared computer.
Skill is Not a Spell; It Just Saves the Successful Method
Only after the AI Intelligence Officer manually completes a weekly report where the format, verification, and approval all meet requirements do I consider saving a Skill.
The order cannot be reversed.
A process with errors saved as a Skill won't magically become correct; it will only repeat errors more stably. Delivering the same error on time every week is more troublesome than occasional failure.
I'll have it save the successful process as a "Weekly AI Intelligence Report" Skill, which includes:
- Which types of pages to start from each time;
- How candidate information is filtered;
- Which fields must be verified back at the original page;
- How conflicts and missing fields are marked;
- The fixed format of the final weekly report;
- Which operations must stop for approval.
Skills should be focused. A set of fixed pages, an output format, and a set of approval boundaries are enough for it to use repeatedly. Stuffing "search news, write articles, post to X, reply to comments, count data" all into one Skill makes it hard to know which step broke when problems arise.
After saving, I'll test it again with a different theme. The first time research "model updates," the second time switch to "AI Agent tools." If it can still read new input without secretly using last time's candidates, and the format and verification rules remain consistent, then this Skill is solid.
Changing information shouldn't be stuffed into the Skill's fixed memory. I can put a weekly-brief.md in the shared computer, writing this week's theme, time range, exclusion words, and processed items. Each run reads this file first; if the file is missing or the date is expired, it reports the problem and stops.
This is more reliable than asking the Bot to "remember what I said last week." Memory is suitable for saving preferences; files are suitable for saving changing facts.
Routine Makes It Run on Time, But I Won't Rush Automation
Routine is responsible for "when to execute." Skill is responsible for "what method to follow." Bot is the character taking on the role.
I understand the three as: Person, Method, Schedule.
Only after the manual process passes do I establish a weekly Routine. When creating it, I clarify the schedule, time zone, input files, delivery content, approval boundaries, and what to do when specified pages won't open.
I'll put this in the long-term instructions:
1Any operation involving sending, publishing, purchasing, payment, deleting, overwriting, batch modification, permission changes, production environment operations, and agreement acceptance must stop.2Before requesting approval, show me the operation target, current status, specific content to be executed, potential impact, and whether it can be undone.3Without explicit approval, only drafts, previews, or operation plans can be generated.
The client might give me three choices: Allow once, Deny, Always allow.
I only use "Allow once" at the beginning. It's a bit more troublesome, but it lets me see exactly what the rules cover. "Always allow" will let future operations matching this rule pass directly. Webpage buttons, processes, and copy will change; a rule written too broadly might allow actions I didn't initially intend weeks later.
Reading public webpages, organizing ordinary files, and generating drafts are suitable for gradual relaxation. For sending messages, placing orders, deleting data, changing permissions, and touching production environments, I will always keep manual confirmation.
Security isn't because AI is "bad." The problem is more ordinary: it might misunderstand the target, the page might be redesigned, third-party content might induce operations, and I might miss conditions in the task. Keeping key actions for humans is to ensure a small misunderstanding doesn't directly become an external consequence.
When Stuck, I First See What It's Waiting For
If the Bot doesn't reply, it's not necessarily crashed. It might be waiting for login, a verification code, permission approval, supplementary information, or the webpage might not have loaded yet.
Opening Agent Computer to see the scene is more effective than repeatedly sending "Are you ready?"
If the direction is wrong, I send a short and clear new instruction: "Stop current search, keep only the three already verified, and deliver in the original format." When it needs to terminate immediately, send "Stop now" directly.
Here are several types of failures I encounter.
Login successful but client didn't switch back: Keep Grok Bot open, switch back manually, and click "Sign In with Cursor" again. Team accounts check SSO.
Cloud computer initializing forever: First see if progress is still changing, then restart the client, check for updates, Retry, or Recover. Reset is the last resort.
Website repeatedly asking for login: I take over the computer, complete the CAPTCHA, wait for the page after login to truly load, then return control. Some websites actively let sessions expire; this isn't necessarily a Bot failure.
Attachment unreadable: First check if the file is encrypted, still uploading, or if the format is too special. Common limits change with versions; for large files or videos, check the client's prompt at the time. Special formats can be converted to PDF, CSV, plain text, or images before trying.
Results always off-track: First check if the role is too broad, if the current task has boundaries, if long-term instructions conflict with temporary requirements, and if changing information was wrongly left in memory. Checking these four places one by one is more effective than changing ten kinds of fancy prompts.
The Steadiest Upgrade Path for Beginners
I won't let Grok Bot manage emails, do customer service, or modify backends on day one. That looks like fast progress, but it's hard to locate the problem when every step goes wrong.
I follow a path of gradually increasing risk:
- First summarize a regular PDF, verifying page numbers and dates;
- Then research public webpages, requiring opening original pages and keeping links;
- Then log into websites, only reading information, not modifying;
- Let it generate email, article, or plan drafts, but not send them;
- Use "Allow once" for an undoable operation;
- Only after the same process is manually successful and tested with different inputs do I save a Skill;
- Finally, establish a Routine and immediately run a test.
This route looks slow, but it saves time in troubleshooting. When the previous layer is stable, I know the problem comes from the newly added permission or action when the next layer fails. Opening everything on day one only leads to a vague conclusion: "AI is unreliable."
Tools will certainly make mistakes, and tasks written by humans will miss conditions. Checking them separately gives me a chance to fix the process.
The Three Templates I Use
If you don't want to look back through the whole text, these three sections can be modified directly.
Template 1: Bot's Long-term Profile
Name: Data Researcher Primary Job: Read public webpages and files I specify, and organize verifiable information. Work Instructions: Prioritize using pages and first-hand materials I specify. Keep dates, original links, or file page numbers for each important conclusion. Separate confirmed facts, speculations, and suggestions. Write "Not explained" when materials don't explain; do not guess. Any operation involving login, sending, publishing, purchasing, payment, deleting, overwriting, batch modification, permission changes, production environment operations, and agreement acceptance must stop and request approval first.
Template 2: One-time Research Task
Within the time range of the last 7 days, organize important updates on the specified theme. First look for candidate information from the official websites and accounts I provided. Do not treat search snippets as evidence; you must open the original page before including it in the results. Each result contains the title, date, core changes, original link, and reason for selection. For fields not listed on the page, write "Not listed," do not estimate. When information conflicts, show them side-by-side; do not adjudicate yourself. Finally, return the summary, result list, excluded candidates and reasons, and issues still needing my confirmation. Only deliver drafts; do not publish, send, or modify any external content.
Template 3: High-risk Action Approval
When preparing to send, publish, purchase, pay, delete, overwrite, batch modify, change permissions, modify production environment, or accept agreements, stop immediately. First show the operation target, current status, specific content to be executed, potential impact, and undo method. Only continue after receiving my explicit approval for this specific operation. When not approved, only provide previews or drafts.
Templates don't need to be copied word-for-word. Just remember the structure behind them: the role should be narrow, the task should have results, facts should be verifiable, and external actions should have brakes.
Stop Asking "How Smart It Is"
When using chat AI, I most often ask: "Can it answer this question?"
With Grok Bot, I changed the question: "For this job, can I clearly state the goal, materials, boundaries, and acceptance standards?"
If I can't state it clearly, I shouldn't rush to automate. A process that a person hasn't figured out will only turn into a mess faster when handed to a Bot.
Once stated clearly, things get interesting. The Bot has a name, a role, and a continuously running computer; Skills remember the method, Routines handle the timing, and approval rules stop it at key positions. I don't have to watch every click, nor do I have to throw the whole account at it and ignore it.
My starting point now is still small: a regular file, a read-only task, and a Bot with a name.
After it does the small things right, I take out the first sample and list a test price of 299 RMB or $49. If someone pays, I continue to refine the delivery; if no one pays, I go back to change the field, columns, and client list. The number of tools is useless at this step.
Earning $5,000 a month sounds like a big result, but when broken down, it's just a simple formula: 4 clients at $800, 1 client at $1,200, and 2 projects at $300. Outside the formula, there's harder work: I have to keep 7 deliveries on time, accurate, and isolated from each other, and constantly find the next person willing to renew.
So I start from the first read-only task. Once I understand what the Bot did, where it went wrong, and how to correct it, I'll hand over the next key. $5,000 is the finish line; the first verifiable weekly report is the starting line.
