Wink Pings

After Giving AI a Cloud Computer, It No Longer Waits for You to Speak

The GrokBot team released a one-hour workshop that takes usage from 0% to 100%. In the accompanying ten-step tutorial, the most worth-copying part isn't the prompts—it's the approval line drawn by 'reversibility': 36 drafts queued, 0 sent.

GrokBot team released a one-hour workshop recording, taking usage from 0% to 100%, with very detailed timestamps: 0:57 covers usage limits, 6:41 covers how to create skills for it, 10:41 build a chief of staff agent, 12:48 use an orchestrator to manage Cursor agents, 27:01 covers using bots to replace your software.

Accompanying long-form tutorial written by Codez, ten steps total. After reading it, one judgment: the barrier of this system isn't technical; it's whether you're willing to replace "how to do it" with "to what point it should do it".

Where it differs from a chat box

The first sentence of the original says it plainly: "GrokBot is not a chatbot".

The difference is in the mechanism. Each bot gets a cloud computer, logs into the tools you already use, clicks around like you do, and keeps running after you close your laptop. So you no longer write prompts; you start assigning work.

This shift matters more than it sounds. For the past two years, all AI products have kept the model in a window while the work stayed with you: you open it, you ask, it answers, you close it. GrokBot moves the work away, at the cost of it getting your login state and your time.

The few of the ten steps that are truly worth copying

Step one: after installing it, first give it a small thing. Not a test question, but a real errand whose result can be verified within 30 seconds. The tutorial's exact wording is that the next nine steps are all about learning to hand over bigger and bigger things to it; the first one must make you willing to trust it.

Step two: give it a job title, not a prompt. A prompt is a one-time request; a bot is a role: it has memory, a domain, and it becomes familiar as you come back repeatedly in the same session. The name should sound like a real person could do the job: inbox manager, expense manager, sales outreach. Then write a charter with three parts: what you own, what counts as doing a good job, and under what circumstances it must ask me first. Everyone skips the third part, but it determines whether you dare let it run unattended.

Step three: connect tools only once. The plugin panel includes Notion, Slack, Google Drive, AWS Agents, AWS SageMaker, Browserbase, Composio, Context7, and custom options. Connections are account-level: connect once, and every bot afterward can use it. That means the fifth bot can go online in seconds, and it also means the blast radius of one connection is all your future bots. During testing, only connect accounts you really need.

Step four: don't paste passwords into the conversation. When it hits a login wall, the bot hands you the screen; after you verify and click done, it continues from the same browser session. What you give is a session, not a key. Any tool that asks you to type a password into a chat box is headed in the wrong direction.

Step five: demonstrate once, don't explain twice. Let it watch you do it once from the side; it saves the process and runs it by itself next time. The work that truly eats up your week is usually the kind that's hard to describe but takes forty seconds to demonstrate: pull numbers from the dashboard, compare the metric that dropped, paste it under the corresponding heading in a doc, and if it falls more than 15%, call the owner on Slack. There are three criteria for choosing the first process to record: you do it at least once a week, it spans two or more tools, and the steps are basically unchanged.

Step six: let the process run by itself. There are two trigger types, by time or by event, and you can state it clearly in one sentence—no need to drag nodes on a canvas. The tutorial gives some ready-made lines: every weekday at 7 a.m., look at the calendar, inbox, and #launches channel and give a short briefing; when an email arrives from an unknown domain and mentions a quote, draft a reply from the template and put it on hold; every Friday at 4 p.m., archive the week's receipts and report items with no corresponding invoice. The most convenient line is to add "run once a week" after a task you just became satisfied with.

Step seven: hire specialists, not generalists. Each bot has independent memory, context, and scope of responsibility. A bot that only handles expenses will truly understand your expenses. A generalist handling expenses, recruiting, and outreach at the same time will do all three poorly, and when something goes wrong you won't have a clean session to review.

Step eight: pull them into a group. Bots can message each other and share context. Give them a goal, not a task list. If you give a list, you've already broken down the work yourself, and they're just executing your plan; if you give a goal, the breakdown is the point of multiple bots. The team's own example is an engineering bot reproducing a bug, creating a ticket, and then handing the problem to another bot to debug.

image

The approval line is drawn by reversibility

This is the most practical rule in the whole method.

Don't draw the line by task size; draw it by whether it can be undone. Drafts, archiving, tagging, summarizing, research, preparation, reconciliation—all reversible; let it finish them itself and keep a log. Sending to people outside the company, spending or moving money, publishing externally, deleting anything that isn't obvious junk, accepting any terms—all stop and wait for you. When unsure, use one sentence as a catch-all: if it can't be undone within a minute, stop and ask me.

The sales outreach example in the tutorial illustrates the shape well: 36 drafts queued, 0 sent. All reversible actions were completed, and it stopped exactly at the irreversible step.

Automation rots quietly

This is the common failure mode of all always-on systems. A website changes, the process starts silently producing garbage, and because it runs while you sleep, no one notices for three weeks.

So set aside fifteen minutes each week and ask three questions about every process: did it run, is the result correct, and would I miss it if I deleted it. The third question is the easiest to fudge; these tools naturally drift into a pile of half-useful automations no one dares delete. The best review method is to have the bot report on itself: how many times it ran this week, what it produced, what it skipped, what failed, what it guessed, and which queued approvals I never answered. Then personally spot-check one output from each process, because when a bot reports on its own work, its blind spots are the same as yours.

According to the team itself, bots get more familiar the more you use them and even start acting before you speak. Treat that as a direction, not a promise you can rely on in beta.

The final conclusion is worth remembering: the focus of skills shifts from "how should I phrase this" to "what exactly am I delegating, and where are its permission boundaries". This is a management problem, with little to do with prompts.

发布时间: 2026-11-18 03:31