I built three AI creators in four days. Here’s all of it.

4 min read

From a blank page on Friday to three live AI creators on Monday. The tools, the timeline, what it cost and what broke.

Juno, Grace and Sal, the three AI creators, side by side

Last Friday I had a blank page. No names. No faces. No accounts.

On Monday morning, three AI creators posted their first videos.

Juno reviews AI tools. Grace helps people over 50 spot scams. Sal helps restaurant owners fill seats.

None of them are real people. All three tell you that up front.

This is how they got made. Start to finish. Including the parts that broke.

Why I did this

I’ve been in digital marketing for about 25 years. Most of it with restaurants.

AI creators are coming whether we like it or not. Most of the ones out there are built to pass as human. Perfect skin. Perfect life. Not a word about how they were made.

That bugs me. It feels like a trick, and tricks fall apart.

So I wanted to try it the other way. Openly AI. Actually useful. Still worth a follow.

I didn’t have niches picked out. I started from scratch and narrowed it to three.

The timeline

Friday, Oct 9: Three characters

  • Picked the niches. Six ideas made the shortlist. Three made the cut. The other three are on the bench.
  • Designed each face in Higgsfield, starting from one reference sheet per avatar.
  • Picked a voice for each in Jogg. I listened to every one myself. AI can’t hear, so that part is on me.
  • Wrote a character bible for each. How they talk. What they’d never say.
  • Made the photos and short video clips I’d need for week one.
  • Wrote week one’s scripts.

Saturday, Oct 10: The studio

  • Set up a separate workspace for each avatar, so their voices don’t bleed into each other.
  • Built a checklist for every account: email, handle, bio, photos, banners.
  • Started on accounts. Email addresses first, then Facebook and Instagram.

Sunday, Oct 11: Account day

  • YouTube, TikTok, LinkedIn, X and Threads. Three times over.
  • Made banners sized for every platform.
  • The power went out. The deadline moved from 2 PM to 8 PM to 10 PM.
  • Rendered the three intro videos. Watched them. Approved them.
  • Scheduled Monday’s posts.

Monday, Oct 12: Live

  • Grace posted at 7 AM. Juno at 9 AM. Sal at 2:30 PM.
  • Started building this site the same day.

Tuesday, Oct 13: Week one rolls on

  • First real content posts. Plus this one.

How one video gets made

It’s six steps. Most of it runs through AI. The decisions don’t.

1

The script. I write it with Claude, using that avatar’s character bible. I read every line before it goes anywhere.

2

The face. Every photo and clip starts from the same reference sheet in Higgsfield. That’s what keeps Juno looking like Juno.

3

The talking. I drop the script into Jogg. It comes back as a video in that avatar’s face and voice.

4

The edit. Cutaways, text on screen, captions. I still do this part by hand. I’m working on that.

5

The posting. Postly and Ocoya schedule it everywhere, with each platform’s AI label switched on.

6

The numbers. Every week the results come back and shape the next scripts.

AI does the heavy lifting. I’m the one who says yes or no.

What broke

Plenty. Here’s the short list.

  • Faces drift. Sal kept showing up in a jacket. Or plaid. Now every prompt spells out his shirt, down to the rolled sleeves.
  • Grace looked like a model. Her first version had spiky, runway hair. I had to tell it: retired schoolteacher, not fashion magazine.
  • AI can’t write on screens. Every laptop and phone came out covered in gibberish. So screens stay blank, and any real tool footage is a real screen recording.
  • One bad frame ruins a clip. Sal’s first video drifted because it started from a photo I hadn’t checked. Now nothing gets used until I’ve looked at it.
  • The first voice for Sal was wrong. Too smooth. It sounded like AI trying to sound like a regular guy. I swapped it for a voice that sounds like he’s been on a restaurant floor.
  • Accounts were the hardest part. Not the AI. Phone number checks on YouTube and TikTok. Instagram logins that wouldn’t connect. LinkedIn wants business Pages, not personal profiles. Sal’s TikTok handle got locked for 30 days.
  • The power went out the night before launch.

None of it stopped the launch. All of it slowed me down.

What it cost so far

I said I’d share real numbers. Here’s where it stands after four days.

What

Cost

Faces and photos (Higgsfield)

About 130 credits. Most photos were free in unlimited mode.

Short video clips (Higgsfield)

About 765 credits for 34 clips. 31 were usable.

Three intro videos (Jogg)

3 credits, out of 2,000 a month

21 social accounts

Free. X runs through Ocoya so I’m not paying $30 a month per account.

Monthly tool costs

$240

My time

40 hours over four days

Video is where the credits go. Photos are close to free. That shaped how I plan every week.

The three rules I won’t break

1

Every avatar says it’s AI. In the bio, in the post, with the platform’s AI label on.

2

No fake experience. If Juno says a tool works, someone actually tested it. On screen.

3

Nobody gets torn down without asking first. In writing.

What’s next

Week one runs through Sunday. Next week I’ll post the numbers here. Views, follows, what worked, what flopped.

At day 60, the weakest avatar gets reworked or replaced. You’ll see that call get made here too.

Want the full step-by-step? Read How It’s Made.

If you want to follow along, the build log goes out once a week.

Mo

I’ve spent about 25 years in digital marketing, mostly with restaurants and hospitality brands. I run a small creative agency and build websites for a living.

More from the Build Log