Design is still make, react, change. What changed is how cheap "make" became. If a bot can put three versions in your day by lunch, you can judge them instead of arguing about them in a doc.
John Bai, writing on the official Grok Bot guides, used Grok Bot to design Grok Bot. The useful sentence is not "AI made the mock." It is: the bots explore, you keep the judgment. This article is the Academy rewrite, plus the briefs on Grok Bot content prompts and Grok Bot skills.
This is the teammate product (cloud computer, memory, tools). Not Grok Imagine on grok.com. Imagine is a different page: Grok image prompts.
Four named bots, not one "design GPT"
Bai's roster is small on purpose:
- Figma production. Repetitive layout work against the real file. Exact positions, not eyeballing.
- Motion. Prototypes how a thing should feel. You talk in comparisons ("hold the look left longer"), not milliseconds you have not seen yet.
- Experiments. Ideas you want tangible before they are on a roadmap.
- Dev. How the idea would actually get built, so the others do not invent APIs.
They huddle. Experiments can hand character motion to Motion and implementation questions to Dev, then bring a prototype back. That only works if each name has a refuse-list. Experiments does not polish the production file. Figma production does not invent a new product direction.
Experiments: make it real enough to hate
The first idea he wanted to feel, not discuss, was ambient access: how do you invoke a bot without another window.
Three versions came back:
- Tucked into the Mac notch. Home without a panel.
- Peeking from a desktop corner on hover. More like a character walking on.
- Following the cursor. Closer, easier to find annoying.
None of them shipped as explored. That was the point. He had three things he could bump into during a normal day and notice: helpful, distracting, or neither.
The brief for an experiments bot is almost anti-roadmap:
You are Experiments on my Grok Bot computer. Take [IDEA] and make [N] versions I can try today. Hand motion to [MOTION BOT] and build questions to [DEV BOT]. Bring back a short note: what I can try, what you faked, what would be needed to ship. Do not add it to a roadmap. Do not polish a production file.
You still decide which version, if any, deserves a real ticket.
Motion: work on the real asset
Historically motion is where a designer hands off. You can describe a bounce. Someone else has to make it real before you can react.
Bai's move was the opposite of "generate a character from a screenshot." He pointed Motion at the production animation spec and asked for a localhost playground around it: timing, springs, distance, scale, easing, running. Then he talked like a person: hold the look down-left longer; the bounce is too much.
That is the pattern to copy. If the bot rebuilds your mascot from a JPEG, you are reviewing an approximation. If it opens the file the product already uses, you are reviewing the thing customers will see.
Keep publish and production deploys off that bot. It can run a local playground. You ship.
Figma production: the work that steals attention
Onboarding needed twenty starter-bot cards from one auto-layout component. It also needed everyday-tool logos placed in a template, in a real order.
Those tasks are not hard. They eat the hour you needed for the next idea. They also multiply small errors: the wrong logo cut, the tenth card spaced differently from the first.
The rule he used: do not let the bot eyeball it. Through a Figma connection it can read x, y, width, height, spacing, type, fills, strokes, and the component. The existing frame is the source of truth.
A brief in that spirit:
You are the Figma production bot. Populate [COMPONENT] for the rows in [LIST]. Use the file as the source of truth. Do not invent a new component. Do not replace a logo with a lookalike. Show me the frame when the set is done. I publish.
If you do not have Figma on the computer yet, the bot should stop at the login and hand you the machine. Plugins prompts cover what this computer is allowed to install.
Keep the loop tight
None of this makes design a one-shot. It makes the loop shorter. You get an idea into the world earlier, you react at higher fidelity, you spend attention on the choice between versions.
Grok Bot does not replace that choice. If a designer on your team is afraid the bot will "ship taste," write it down in the Bot description: no production file writes, no brand-new illustration unless you ask, no sending the deck.
Source
Field story: Designing Grok Bot with Grok Bot by John Bai on x.ai. We kept the three loops (experiments, motion on the real asset, production in Figma). The screenshots and product clips stay on their page.