I Made a Product Promo with Grok Bot, Entirely from My Phone

I started using Grok Bot this month and had it make a series of promo videos for my products. I did the whole thing from my phone.

This is my first Grok Bot tutorial: how I got started, what the first cut got wrong, and how I kept pushing it toward something I actually liked. I'll share more experiments later.

Personal agents have been getting more attention over the past couple of months. Think of one as an assistant that can keep working with you over time. Give it a computer in the cloud, and there's a lot it can do.

I started seeing more discussion about Grok Bot in September. Once I tried it myself, I found it could research news, sort through email, install software on its own computer, and make videos.

Set up a subscription and your first Bot

I used a U.S. Apple account, added credit with an Apple gift card, and subscribed to SuperGrok inside the Grok app.

I've written about the gift-card process here: How I pay for ChatGPT with Apple gift cards (in Chinese).

If you need help setting up a U.S. Apple account or redeeming a gift card, I cover that in my Jieni AI Practice Manual.

From the main screen, create a Bot and give it a name. I called my first one CEO, since it would be my first stop when I needed something done.

Tap the plus button in the upper-right corner to create one. Tap its avatar to set its name and appearance, then return to the chat and send a message.

The important setup step is Connect Apps, under your personal account. Search for the services you want, such as X or Gmail, and follow the authorization steps. These connections let the Bot read the material you authorize and use the corresponding tools.

I started with three Bots

My CEO Bot coordinates things. When something comes up, I usually start there.

I called the second one Chief Information Officer. I moved over scheduled tasks I'd been running in ChatGPT, including GitHub Trending and news from X. It sends updates at different times, giving me a kind of news radar.

The third is my Chief Creative Officer. I use it to discuss short-video ideas, visual styles, film references, and how particular pieces were made.

I also connected Gmail. I subscribe to quite a few newsletters, mostly about AI. I ask the information Bot to summarize them, filter out the noise, explain what deserves my attention, or find emails I need to act on.

If you're starting out, one Bot is enough. Give it one task you actually need done. For a scheduled task, you could say:

Every day at 10 a.m., summarize GitHub Trending and posts on X about the topics I follow. Remove duplicates, tell me what's worth reading, and keep the source links.

At first, I mostly used it for research and email. Then I started trying video generation, and, well, I got a little hooked.

The first promo was pretty ordinary

I wanted to see whether it could make a promo for one of my products.

I'd seen some impressive promos made with Claude Opus 5.5. I asked my CEO Bot to research discussions on X and find Skills people were using for that work. A Skill is a set of instructions you can reuse for a task.

My requirement was specific: show me the evidence and give me the GitHub repositories. Check the repos as well as the discussions. Don't just hand me a list of names.

It came back with a few options. I skimmed the READMEs and picked claude-motion-design. You can find it on GitHub by searching for howseen-ai/claude-motion-design.

After installing it, the Bot asked whether I wanted to try it with a product or website.

I've been working on md2wechat for over a year. It's a product that helps creators prepare content for WeChat Official Accounts.

I sent the repository and asked for a promo aimed at Chinese-speaking users, with a more polished visual style.

It read the repo to understand the product. About five minutes later, the first cut was ready. It used a pale green palette, with screenshots, copy, and numbers taken from my GitHub documentation.

It looked pretty ordinary. About the same as the promos I'd made with GPT-5.6. Nothing that felt new.

My reaction was basically: What is this? It doesn't look polished enough. It's not exciting enough.

One more specific reference changed the result

After complaining, I had to ask myself what I actually wanted. Words like polished and exciting didn't give it much to work with.

Then a reference came to mind:

I want the kind of top-tier creative you'd see in a YouTube ad.

I watch a lot of tutorials on YouTube. Some of the product ads that play before or during them have great camera movement, transitions, pacing, and music. The first few seconds can really grab you.

The Bot proposed another approach, including a 3D wall and new music. This version took about half an hour.

The difference was obvious. The opening, transitions, animated typing, 3D elements, and the flash and sound effects at the end all had more energy.

Okay, there was something here. It had understood the YouTube-ad feeling I was trying to describe.

Let it find the brand assets

After sending the second cut, it asked something I hadn't expected: could I provide a high-resolution logo and some better product screenshots or recordings? That could improve the next version.

It was right. The colors seemed a little off, too. I'd updated my website, and the brand palette had changed.

I sent it the website and told it to take a look itself.

It found the logo, brand colors, official copy, and high-resolution screenshots, then made a third version. The palette changed to terracotta red on a warm black background. It also used the website's line: you write the content; we handle the WeChat formatting.

That combination felt much closer to the product.

Then it asked for roughly ten seconds of real footage showing the tool running in a terminal or Obsidian. It could use that in place of the fairly generic terminal scene.

I liked that question. It was helping me figure out how to improve the work. It felt a little more like working with a person.

But I was doing all of this on my phone. I didn't want to open my computer just to record a clip.

So I told it to install the tool itself. I gave it the npm install command and asked it to run the tool and capture the terminal.

It had its own cloud computer, after all.

It said the installation was done and started taking screenshots. The fourth version replaced that scene with an actual terminal replay.

One limit matters here: I didn't give it an API key. What it made was a terminal replay. That does not mean it had completed the full WeChat formatting and publishing process.

Fix the music

By the fourth version, I was pretty happy with the visuals. The music still felt mechanical, and it didn't hit the cuts at all.

I'd originally thought about connecting a music-generation tool, but it wanted me to log in. I had a free account, and finding or making music was starting to feel like a hassle. I dropped that route.

Instead, I asked it to find music with a clear beat that it could cut to: clean electronic or ambient electronic music, with the feel of a major product launch.

Across these revisions, I stayed on my phone. I had high expectations, and there were a few small surprises along the way.

Save the process as a Skill

Now came the most important step for me: saving the process as a Skill. Once the visuals and music were in a place I liked, I asked it to save what we'd done so I could use it again.

My idea was to give it a topic, link, piece of content, or article later, and have it work out how to make a video in the same style.

It put together a Skill for beat-synced promos. At that point, I'd made it through the process from the first attempt to something I was happy with.

I still wanted to see what else it could do.

I asked which styles were worth trying beyond YouTube ads. It suggested Apple product launches, the physical intensity of Nike ads, and the cinematic feel of a Netflix opening.

I had it try the Netflix direction. The music and camera work did feel cinematic. But md2wechat is a small tool for formatting WeChat articles, and that style still felt a little off for it.

Later, I asked it to study lighting and symmetrical composition, and look for images and video footage that could be used commercially. The design ideas carried over.

Keep digging into a direction you like

I kept following the creative-ad direction, asking it to research leading ad directors, award-winning work, and other visual styles worth studying.

My approach is to start with a few distinctive references. I still have to judge the finished video. One sentence won't make it perfectly reproduce a style.

Everyone has their own taste. When I find a direction I like, I dig a few layers deeper and try combining styles to see what happens.

The whole experience felt smooth to me.

• Research first, then provide product material and style references.
• Watch the first cut, say what feels wrong, and revise it version by version.
• Save the approach as a Skill once you're happy, then try another style.

What I took away

Making promos and short videos with an agent is getting easier.

For me, taste still matters most. You have to judge the work yourself. I had to decide what looked good, what didn't, and what I wanted to try next. I found the result through several rounds of changes.

Try giving it your own product or an article and asking for a first cut. If it looks ordinary, say so. Give it a specific reference you like and see what the second version looks like.