The easy part is building the agent
I build AI systems for clients and sell them for a lot of money. Here is what building one live with my 15 year old son taught me about where the real work is.
I do AI systems for my clients and I sell them for a ton of money. Sometimes it does not work, because when you tell the client what it costs they say they do not want it anymore.
The easiest part is creating the agents. The hard part is selling it. A lot of people on YouTube and other social media do not say that. They say make $50,000 selling AI agents. It is not as easy as you think.
6:56
What we were actually building
My son Jeremy is freshly 15. He wants to start his own merch and sell it on TikTok. He is the creator and the model. He has no content creation strategy, it is just me figuring it out. He has no idea what any of it costs, which he says is not his problem, it is dad's problem.
I have worked with over 400 companies in the last 10 years and generated tons of revenue for them. I can market anything. I have a decent social media following, but my interest is not getting more views and more likes. What I do for my clients is help them make more money.
My interest is not getting more views, more likes.
Teach the agent what he likes, then design without him
The first thing we did was have a conversation. Jeremy introduced himself to the agent, said he wanted to build his own t-shirt designs, and asked it to ask him questions. Funny or clean. Brands he likes. Who the audience is, which for him is strangers on TikTok. Front only, back only, or both.
The reason we do this is so it can help me create the designs at home without actually having him there. No going back and forth. I already have his preferences, so I send him a batch and he picks the ones he likes.
15:45
Before you ask an agent to make anything, spend one session making it ask you questions. The answers are the asset. The output comes later.
The name was the hardest part
We asked for 10 to 12 brand names with a point of view attached, then had it save every name to a file so I have it for future reference. Human error was good. Night shift was good. Both were taken three or four times over.
Mujin was the one. It means unmanned, nobody working there. A mujin store in Japan is a real thing, all over the country. You walk in at 3am and there is no cashier. An American brand with a Japanese name, and nobody else is using it.
We figured out the name of the brand. That is the most important part and the hardest part. Everything after that is designing a website, a marketing campaign, and shooting content.
Six designs, print two
Drop one is two designs, one garment color, four sizes. Washed black. But we make six lockups, not two, because the six go into a poll on TikTok and people argue in the comments about which one we should make. That is the thing you actually want.
Then you open pre-orders on the winner and watch how many of the voters actually show up. Other brands have done this before. It also means the audience interacts with you and starts feeling like they belong to something bigger.
33:28
Make more designs than you plan to print. The extras are not waste, they are the content and the vote that tells you what to put money into.
Do not let it hand you a store nobody has been in
The design is a faded, blurry photo of a drinks fridge in a convenience store at night. The agent was clear that stock photos get you a convenience store nobody has ever been in. So Jeremy goes to the 7-Eleven near his house, takes photos of what he thinks looks good, and texts them to me. We use those as the references.
That is how much more original the whole process gets. I am teaching him the new methods, but I am not doing those stupid AI videos and AI photos that brands are doing now.
34:25
Stock photos get you a convenience store nobody has ever been in.
Why I would rather do this than hire it out
It is easier to do this than argue with an employee for 30 minutes. Imagine having a video editor and chasing him. Where is the stuff. You told me Tuesday and today is Thursday.
These tools work 24/7. They never complain and they are never going to ask me for more money. They do make mistakes. On this one it rendered the mockup, found the errors itself, said it was going to fix it and ran it again.
43:39
Watch what the work costs you
Every time we generate an image or a mockup, we pay. Last week I was generating ads for a client and I forgot to tell the image generator to use the cheaper model for drafts. It cost me about $8. For image generation that is a lot. It should be a dollar or less.
Same idea with the model you are in. Switch models. If I am writing a caption, something quick, I use the cheap one. If I am brainstorming or designing, I need the best model. Those things matter.
19:02
Open a conversation with the agent, say what you want to build, and ask it to ask you questions first.
Have it save every name and idea it generates into a file so you can go back to them later.
Go take your own reference photos on your phone instead of accepting stock images.
Set the cheap model for drafts and captions, and the best model for brainstorming and design.
We stopped when we were looking at fonts, because if we go down that path we will be looking at fonts forever. We have the name, the lockup and the direction. Next is the website, the content, the TikTok shop and the traffic.
We nailed Mujin. That was the hardest part.
01
Build the agent around what someone likes before you ask it to make anything.
02
The name and the direction are the hard part. The website and the content come after.
03
Shoot your own reference photos and set the cheap model for drafts, so the work stays original and the credits stay low.
Building something like this?
If you are an operator trying to turn content and AI systems into booked work instead of views, tell me what you are working on.
Send me a text