Wink Pings

Another One Says It's Not a Chatbot—This Time It Actually Has a Computer

The GrokBot team hosted a one-hour workshop on how to turn AI from a Q&A box into a coworker that gets work done for you. Beyond the feature demo, what's more worth watching is how it handles login, where it draws the approval line, and the less-than-polite feedback in the comments.

Another One Says It's Not a Chatbot—This Time It Actually Has a Computer

The GrokBot team released a one-hour workshop with a very ambitious title: from 0% to 100%, teaching you how to hand work off to an AI agent.

The core claim is just one sentence: GrokBot is not a chatbot.

What's the difference? The AI tools you've used before are all waiting for you: you open it, you ask a question, it answers, you close it, and the work is still in your hands. GrokBot gives each bot a cloud computer. It can log in to the tools you commonly use, click through and operate them like you do, and it keeps running after you close your laptop.

In other words, it's a shift from asking questions to delegating.

The workshop is split into five segments by progress: 0:57 covers how to manage usage limits, 6:41 covers how to build skills for the bot, 10:41 builds a Chief of Staff agent, 12:48 covers using GrokBot to orchestrate agents in Cursor, and 27:01 covers replacing some of your software with bots.

The most informative part of the timestamps is the last segment. The earlier parts are all about how to get things done; the last one asks: which software could you actually stop using?

The key isn't the feature list; it's a few mechanisms

In the original 10-step tutorial, there are several design choices more worth looking at than the feature list.

First, give the bot a job title, not a prompt. A prompt is a one-time request; a bot is a role: it can accumulate memory, own a domain, and get better over time within the same thread. The tutorial includes a charter template you can directly adapt, with three parts: what you're responsible for, what counts as good, and when you must ask me first. The third part is the easiest to skip, yet it's exactly what lets the bot run unattended, because you've drawn the boundaries of authority in advance.

Second, how login is handled. The bot operates in a cloud browser; when it hits a login wall, it hands the screen over to you. After you log in and click done, it continues in the same browser instance. What it gets is a session, not a password, so you're always present for the sensitive step. The tutorial puts it bluntly: if a tool asks you to paste your password into a chat, that's the wrong path.

Third, demonstrating once beats explaining twice. Many time-consuming tasks are painful to describe but take only forty seconds to do. You perform it once while it watches, it saves it as a routine, and next time it runs on its own. The criteria the tutorial gives are also practical: done at least once a week, spans two or more tools, and the steps are basically unchanged. If all three conditions are met, that's work you can hand off.

Fourth, draw the approval line based on reversibility, not task size. Drafting, archiving, tagging, summarizing, researching, preparing, reconciling—these can be undone, so let it do them on its own. Sending externally, moving money, publishing publicly, deleting, accepting terms—these cannot be undone, so stop and wait for a human. The tutorial leaves one test: if it can't be undone within a minute, stop and ask. In the sales outreach example, 36 drafts are queued and 0 are sent; the bot completes all the reversible parts and stops exactly at the irreversible step.

The comments are less polite

Below the workshop, besides the applause, there are several less flattering pieces of feedback worth looking at too.

Someone says directly that every AI tool now claims it's not a chatbot, and they're tired of hearing it. A meme follows that comment. It's not an unfair point: the slogan itself doesn't create a difference; whether it can actually click through that whole workflow for you does.

Someone throws cold water from the permissions angle: the '0% to 100%' phrasing is a bit much. The workshop covered limits and skill building, but for an agent to really take over work, the hard part is the permission boundary—what it can read, what it can execute on your behalf, and who calls stop if it executes something wrong. This objection actually lines up with the approval-line section in the tutorial, which shows this is indeed the wall everyone hits first.

Someone else says their bot is still a bit dumb, that it keeps running the original task even after being taught wrong, and that they can't tell how many people have actually gotten a complete workflow running. Another more direct comment: if a product needs a one-hour course to teach you how to use it correctly, then the product itself isn't well made; this should be built into the software.

There's also a reminder about cost: the more bots you have, the faster tokens burn. The product is good, but the brake is still on.

Put together, these voices are more informative than a one-way feature introduction. GrokBot's mechanism design has substance, especially the login handshake and the approval line drawn by reversibility; you can tell the team has thought about loss of control. But it's still far from being something everyone can use to mindlessly hand off an entire business.

What really changes is habits

The closing line of the tutorial is quite accurate: the technical change isn't as big as it sounds; the change in habits is much bigger. The old skill was how to phrase things clearly; the question now is what exactly you're delegating and where its authority ends.

This is no longer a prompt problem; it's a management problem.

If you're going to try it, you can follow the path the tutorial gives: for the first task, choose something small whose result you can verify within thirty seconds; for the first role, choose work you have to do every week, that spans multiple tools, and whose steps are stable; draw the first approval line at actions that can't be undone within a minute. Then spend fifteen minutes each week having the bot report which routines it ran, what it produced, and what it skipped, and spot-check one item. Automation rots quietly, and this step is the only place you'll catch it.

You can watch the workshop and read the tutorial, but don't expect to be at 100% after finishing. Start with one real, small task and see where it stops.

Original tutorial: https://x.com/i/article/2089655442657910784

发布时间: 2026-11-18 07:26