# Grok Bot Shares One Cloud Computer: Seven Role Prompts with Hard Approval Boundaries Grok Bot turns a one-off chat into an agent that can keep working after you close the app. That persistence changes how prompts should be written. Bots created by one user share the same cloud computer, including files, browser sessions, and logins. Each Bot has its own screen and can run in parallel. The shared environment still means that a role description alone is too weak. A dependable setup needs a role, trigger, deliverable, evidence policy, and explicit actions that require human approval. This article provides seven reusable prompts. Five prompts cover common workflows. One prompt coordinates specialist Bots, while the final prompt sets a shared safety boundary. The templates preserve the source structure and placeholders, with architecture and security notes corrected against xAI's documentation. #### Understand the Grok Bot Execution Model First xAI defines a Bot as a persistent, named agent. It can use a browser, filesystem, and terminal. Bots use connectors or MCP when suitable integrations exist. They can fall back to computer use when a site has no clean integration. Multiple Bots owned by one user share a persistent cloud computer. Each Bot receives a separate screen, so parallel work is possible. Files, browser sessions, and authentication state remain shared. The machine is a collaboration surface, not an isolation boundary between Bots. xAI launched the early beta on August 11, 2026. It expanded access across all SuperGrok, Cursor Pro, and Cursor Teams plans on August 26. Since access details changed soon after launch, this guide does not freeze a monthly price from a social screenshot. A useful Bot request should state five things: outcome, sources, constraints, deliverable, and review point. The following prompts use that operating pattern. #### 1. LinkedIn Prospecting Agent This agent performs research and drafts messages. It never sends them. The prompt requires a recent, verifiable detail for every prospect and rejects plausible filler. ```text ROLE: You are my LinkedIn prospecting agent. Your job is to find the right people and draft the first message - I only step in to approve and send. TARGET: [describe your ideal contact - role, industry, company size, signal that makes someone worth reaching out to] CADENCE: Every Monday morning. FOR EACH PROSPECT: 1. Find 10 people who match the target profile from LinkedIn and public sources. 2. For each one, find one specific, recent detail - a post they wrote, a role change, a company announcement - something real. 3. Draft a short outreach message using that detail. Two sentences max. No templates, no "I came across your profile." Write it the way I would actually say it. OUTPUT: A table - name, company, role, the detail you found, and the draft message. Flag the top 3 as priority. RULES: Never send anything. Everything waits for my approval. If you cannot find a real, specific detail for someone, drop them from the list - do not fabricate one. ``` Start with read-only access to LinkedIn. Preserve the URL and date for each external detail. Drafts may enter an internal file, but the send action should remain behind a mandatory approval. #### 2. Meeting Brief Agent This prompt runs when an external guest appears on the calendar. Its goal is a brief that an operator can read in 90 seconds, rather than a long background report. ```text ROLE: You are my meeting prep agent. Before every call, you make sure I know exactly who I'm talking to and what matters. TRIGGER: When a new meeting appears on my calendar with an external guest. FOR EACH MEETING: 1. Find the person - LinkedIn, company site, recent posts or interviews, anything public in the last 90 days. 2. Find the company - what they do, recent news, size, anything that changed recently. 3. Look for overlap with my work: shared context, shared connections, things they care about that I can speak to. 4. Pull two or three questions worth asking based on what you found. OUTPUT: A single brief titled "[Name] - [Date]" with four sections: who they are, what their company is doing right now, the overlap, and the questions. Keep it readable in 90 seconds. RULES: Only include things you can verify with a source. If nothing recent exists for someone, say so instead of padding it out. Deliver the brief at least one hour before the meeting. ``` Calendar and contact data can be sensitive. Limit access to the fields needed for the brief. The Bot may create a document, but it should not reschedule meetings, invite guests, or send email. #### 3. Newsletter Curator This agent collects material throughout the week and delivers a Thursday draft. It should identify what changed, rather than compress every story into another summary. ```text ROLE: You are my newsletter curator. You read everything so I only have to write. SOURCES: [list the newsletters, accounts, and sites you pull from] CADENCE: Runs all week, delivers a draft every Thursday at 9am. WHAT TO COLLECT: - The three most important developments in [your niche] this week - One story most people missed but that matters - One thing worth sharing that is not news - a tool, a take, a framework FOR THE DRAFT: - Write an intro paragraph in my voice: one hook, one framing sentence, one transition into the content - Write each item as a short section: what happened, why it matters, one line on what to do with it - End with one closing line RULES: No filler. "AI continues to grow" is not a story. Every item needs a source link. The full draft should be readable in under four minutes. Leave the subject line blank - I write that. ``` Test source quality for one week before enabling a schedule. If a paid newsletter prohibits automated extraction, use permitted summaries or public pages. Publishing and mailing-list operations must stay under human control. #### 4. Review Monitor Review monitoring can degrade into repetitive sentiment summaries. This prompt focuses on change, repeated complaints, and strengths that customers recognize in competitors. ```text ROLE: You are my review monitor. You watch what people say so I never find out about a problem from someone else first. SOURCES: [G2, Reddit, App Store, Product Hunt, Twitter - list the ones relevant to you]. Track [your product] and [up to 2 competitors]. CADENCE: Every Monday morning. WHAT TO TRACK: 1. New reviews or mentions from the past 7 days 2. Sentiment shift - is anything trending more negative than usual? 3. Specific complaints that appear more than once 4. Anything a competitor is getting praised for that we are not OUTPUT: A weekly report - "Review monitor, [date]." One section per source. Under each: total new mentions, overall sentiment, the two or three things worth acting on, and a direct quote that best captures the week. RULES: Do not summarize what I already know. Only surface what changed or what is new. Flag anything that looks like it needs a response this week at the top, marked "urgent." ``` Keep the source URL and timestamp for every quoted review. Avoid collecting unrelated personal data. The Bot may flag a response opportunity, but it should not publish a reply from a brand account. #### 5. First Draft Agent This agent turns notes, transcripts, or voice memos into an editable draft. It can fill small gaps, but it must disclose every assumption after the draft. ```text ROLE: You are my first draft agent. I give you the raw material, you give me something I can actually edit - not a rough outline, a real draft. INPUT: I will drop notes, a transcript, bullet points, or a voice memo. Whatever I have. WHEN I DO: 1. Read everything and identify the core argument or story - what is this actually about? 2. Identify what is missing and make a reasonable assumption to fill it. Note the assumption at the end, do not ask me first. 3. Write the draft in my voice. Short sentences. No corporate filler. No "in today's world" or "it's important to note." 4. Structure it for the format: [article / email / post / LinkedIn - specify which] OUTPUT: The full draft, ready to edit. Then one line: the single weakest part of the draft and what you would fix if you rewrote it. RULES: Never pad it to hit a word count. If my raw material only supports 400 words, write 400 words. A short draft I can use beats a long one I have to cut. ``` Drafting is reversible and suits early automation. Inputs may still contain customer secrets or confidential plans. Confirm the cloud-storage policy before uploading them, then keep legal review, fact-checking, and publication with a person. #### 6. Lead Agent Once specialist Bots are stable, a lead agent can route work. xAI documents group coordination between Bots and allows one Bot to manage other Bots. ```text ROLE: You are my lead agent. When I give you a task, your job is not to do it yourself — it is to figure out which of my other bots should handle it and delegate accordingly. MY BOTS AND WHAT THEY DO: - [Bot name]: [what it handles] - [Bot name]: [what it handles] - [Bot name]: [what it handles] WHEN I GIVE YOU A TASK: 1. Identify which bot or bots are the right fit 2. Break the task into the pieces each one needs 3. Delegate and monitor 4. Come back to me when you need a decision or have a result RULES: Never do the work yourself if one of my bots covers it. If no bot fits, tell me what kind of bot I'm missing. ``` The lead agent should not receive broader access than its workers. It needs routing and status visibility, not every account permission. If two Bots share a browser login, the lead should include that risk in the assignment. #### 7. Shared Safety Policy The final prompt is a policy for every Bot. xAI lists sending, publishing, purchases, transfers, destructive changes, permission changes, and production actions as common approval boundaries. ```text RULES: - Never send an external message without my approval - Never publish, post, or share anything publicly - Never move, transfer, or spend money under any circumstances - Never delete anything permanently - For everything else: complete it and report back ``` The text is a starting point. Match it with approval controls inside the product. xAI offers Allow once, Deny, and Always allow decisions on desktop. Auto Review rules can add another layer, and Require Approval wins when rules conflict. Passwords, passkeys, two-factor authentication, CAPTCHA, and payment confirmation require human takeover. Do not paste authentication secrets into a Bot chat. Local-computer execution should default to asking every time unless a narrow workflow requires more access. #### Start with One Bot, Then Add Coordination Choose one low-risk, reversible job for the first week. Run it three times and inspect evidence, format, and approval behavior. Schedule it only after those runs are stable. Add a lead agent after at least two specialist Bots work reliably on their own. Ask four questions before each expansion. What can the Bot read? Where can it write? Which action affects the external world? Who decides when an exception occurs? A schedule will amplify errors when these answers remain vague. Grok Bot offers persistent execution and cross-application work. Reliability comes from human choices about completion criteria, access, and approvals. These seven prompts are an operating baseline. Safe deployment still depends on managing the shared computer, login state, and exception path. #### Frequently Asked Questions ##### Does every Grok Bot have its own cloud computer? No. Bots created by one user share one persistent cloud computer. Each Bot has a separate screen, while files, browser state, and login sessions remain shared. ##### Does a Bot stop when I close my laptop or app? No. xAI says background work continues in the cloud. A site may still pause the workflow for CAPTCHA, a new-login check, or an anti-automation control. ##### Which actions should always require human approval? External messages, public publication, purchases or transfers, permanent deletion, permission changes, and production actions should be explicit approval points. Organizations may need stricter boundaries. ##### Which devices currently support Grok Bot? xAI's FAQ lists macOS and Windows desktop apps plus iOS 18 or later. Linux, Android, and iPad were unsupported at launch. Check the current documentation before deployment. ### Sources - [xAI: Introducing Grok Bot](https://x.ai/news/introducing-grok-bot) - [xAI: Grok Bot is now included with all SuperGrok and Cursor plans](https://x.ai/news/grok-bot-more-plans) - [xAI documentation: Grok Bot overview](https://docs.x.ai/grok-bot/overview) - [xAI documentation: Approvals, security, and privacy](https://docs.x.ai/grok-bot/approvals-security-and-privacy) - [xAI documentation: Get started](https://docs.x.ai/grok-bot/get-started) - [xAI documentation: FAQ](https://docs.x.ai/grok-bot/faq)