The Marketing AI Factory
Bye-bye,blank page.
You write one sentence. Eleven AI specialists research it, write it, design it and check it. You press release.
- €7,500, one build
- Built around your brand
- You approve everything
Items in my content system’s tracker
356Scheduled or posted
000I counted on 5 November. It could make content. It couldn’t ship any of it. So I rebuilt it as a factory.
Go on. Give it a sentence.
- Ada
- Vera
- Hook Lab
- Mila + Theo
- Code checks
- Iris
- Kai + Juno
- You
Eight stations, then you. Give it a sentence and press the button.
11
AI specialists
60+
automatic checks
50
metered AI steps
33
editor tools
27
page layouts
66
film styles
786
automated tests
Nine rooms and a basement.
This is the whole factory, in the order work moves through it. Open any room to see what’s in there. Where something is built but not in use yet, the map says so.
Working here: Ada
- Ask the team. One sentence, a file or a link. Ada answers with a priced package of up to six pieces, and one click approves it.
- 90-day plans. A dated plan of pieces for each page, which the team fills in ahead of time.
- Radar. News from the sources you choose. You press Make it or Not for us, and nothing starts until you do.
- Audience listening. Once a week it reads what your audience asks on Reddit and Hacker News and turns that into topic ideas.
- Saved. Posts you share from Instagram become references: make something in this style.
- Repurpose. A finished carousel becomes a reel and a teaser. A reel becomes a recap carousel and a quote card.
- Case-study notes. Your notes on a project become an anonymised case study.
- Product link. A product page becomes a promo video.
- Drops. Drag in a photo or a raw video. Code reads it and Ada proposes what it could become.Built, not used yet
What the night shift learns goes back into every room’s rules.
The basement: what every room stands on
- One engine. 50 labelled steps, each with its own model, effort, dollar cap and clock. A call with no label is refused.
- Cages. An AI call gets a listed set of tools and nothing else. It can’t wander around your computer.
- Memory. Each teammate has one memory that carries from job to job. New lessons go in on trial, and you can dismiss any of them.
- Your brand system. Colours as roles, checked pairs, open-licence fonts, marks. A design that fails the checks doesn’t save.
- One record. The list of every piece and its status. Only code writes to it, never the model.
- 786 tests. Automated tests that run on the app’s own code. Keys stay in your Mac’s Keychain, never in a log.
One click. Four and a half minutes.
The first time a news story came through, I pressed one button. It wasn’t clean, and that’s the part I’d want to know about if I were you: the code refused the first attempt, a slip got through anyway, and now there’s a check for it.
Then I measured it, and it was bad.
The rebuilt factory turned out pieces that passed every check. So I gave two AI judges 58 of them, told nothing about how each was made. Passing every check turned out to mean very little.
51 of 58
read as scripted to one judge. The other said 52.
2 of 34
reels one judge would stop for. The other found 4.
So I tore out how openings get written. Now a Hook Lab mines the real story, writes eight openings, grades them, and a blind reader ranks what’s left. Try it yourself. These are real opening lines my system wrote for reels on my page. One side passed every check, the other is the new draft.
Round 1 of 3 · which one would you stop for?
18 of 18
votes for the new drafts over the pieces that had passed every check
Nine held-out pieces, two AI judges, words only
18 vs 3
times a judge would stop for it, new against old, out of 18
2 vs 18
times it read as scripted, new against old, out of 18
What I can’t show you yet
Judges aren’t an audience. I don’t have a million views to put on this page. Saves, shares and signups from real people are what settle it, and the factory counts them per piece. When I have them, they go here.
Eleven specialists. One job each.
Each has a name, a lane and a memory that carries over from one job to the next. Nobody marks their own work as passed.
- A
Ada
Showrunner
Turns your sentence into a plan and keeps the line moving.
- V
Vera
Researcher
No source, no claim. Every fact goes in a ledger.
- T
Theo
Script writer
Every spoken word, built around the first three seconds.
- M
Mila
Copywriter
Every read word. She can’t add a fact Vera didn’t find.
- C
Cleo
Case studies
Your project’s story, with every name kept out.
- K
Kai
Video editor
Looks at every frame before calling a video done.
- O
Otto
Animator
Films drawn entirely in code, when no footage shows the idea.
- J
Juno
Brand designer
Your colours, type and marks, checked on real layouts.
- R
Rio
YouTube
Decides what gets made and how it’s packaged, then measures it.
- N
Nico
Social
A post written for each platform, and a calendar you release.
- I
Iris
Standards
The independent check. She can fail anyone’s work.
- ?
You
The boss
You pick the topic and you press release. Those two stay human.
What comes off the line.
One request can become up to six pieces, and each one arrives finished: files, caption, alt text, ready to post.
Reels
Vertical, landscape, square or portrait.
A voice, footage or drawings, captions timed to the real audio, a cover, and two or three openings to test.
Carousels
1080×1350, 5 to 10 slides.
A JPEG per slide, a PDF for LinkedIn, a caption, alt text on every slide. Moving slides if you want them.
Story sets
1080×1920, 3 to 7 frames.
A hook, the points, a question, a link. Each frame inside the platform’s safe zones.
Animated explainers
30 to 45 seconds.
Step by step, drawn entirely in code, with a drawn presenter and 16 kinds of picture.
Case studies
A4 PDF, 2 to 4 pages.
Plus its LinkedIn post, a companion carousel and a story set. Anonymised before it’s saved.
Guides
A4 PDF.
A lead magnet in your brand, to give away for an email.
Films
66 styles.
For when no footage shows the idea. Every frame drawn in code, the sound mastered for the platforms.
Posts and quote cards
Six platforms.
A caption and a post written for each platform, inside its limit, and a quote card to go with them.
More than 60 ways to get sent back.
Taste is the part a model skims. So whatever can be a check, is one, and it runs before anyone reads a word. A failure that keeps repeating gets written up as a new check. These are some of them.
Facts
- A number that isn’t in the ledger
- A quote that isn’t on its source
- A recap that changes the wording
Words
- Robot phrasing
- Hype words you banned
- Sing-song even sentences
- A first line a “no” can end
- “Most teams do” with nothing behind it
- A script too long for its slot
- A script too short for its slot
- A phrase from your last five episodes
- A date or a “Source:” line on a slide
Design
- Text that overflows
- Text too small to read
- A colour pair your brand doesn’t allow
- Low contrast
- A missing glyph
- A font that isn’t yours to use
- Outside the safe zone
- A stretched, blurry image
- The wrong slide count
Sound and picture
- A clip with no licence
- A voice you can’t use commercially
- A track you can’t use commercially
- A missing file
- Generic stock footage
- The wrong loudness
- A Short over its limit
Going out
- A client’s name
- A missing caption
- Missing alt text
- A post over the platform’s limit
- Two asks in one piece
- Engagement bait
- A wrong AI label
- A file whose metadata names someone
It only bothers you with what’s yours.
My note to the team, in capitals: “WHY IT NEEDS ME?” Now a stuck piece goes once to the specialist who can fix it. Only money, rights, names, accounts and publishing reach you.
18 of 20
stuck pieces the team solved without me in a week
The other 2 came to me. The whole step cost $0.90.
When you say why you didn’t like a piece, that note goes to the teammate who made it. Edit two pieces the same way and it’s proposed as a rule. You can dismiss any rule, and a dismissed one never comes back.
Every call on the meter.
No guessing what the AI cost you this month. These are my own floor’s numbers from November 2026, measured at API prices. On a Claude subscription they come out of your allowance, not your card.
$2.75
a finished video, on average
Across 158 of them. It once crept to $3.00, and the app raised the alarm itself.
$1.48
a finished carousel or story set, on average
Across 107 of them.
$1.16
the first package: a carousel and a story set
Estimated at $2.64 before it ran.
$35
the most a night shift can use
And $200 a week. A run that would cross a cap stops and says so.
$146.79
all the AI for a week: 1,821 model calls
The week to 6 November, every call on the record. A heavy week costs more: the videos alone came to $435 in the week to 10 November.
A day on the floor.
- 01:30
The night shift starts
It reads the numbers, fills the plan, makes what’s due and unblocks what’s stuck. At its cap it stops and says so.
- Overnight
Iris reviews the day’s lessons
Each proposed lesson gets its evidence checked. The ones that hold up go on trial.
- Morning
You open one screen
Finished pieces wait with their previews and captions. Your decisions are in one short list.
- All day
Radar watches your sources
Every 30 to 60 minutes while the app is open. You press Make it, or Not for us.
- When you post
Send it to your phone
The piece, its caption and its instructions land in a folder. Post it, tick Posted, and it clears.
- Once a week
It listens
It reads what your audience is asking on Reddit and Hacker News, and brings you topics.
It learns from being told off.
A prompt forgets. This keeps a memory per teammate, and every lesson has a status you can see.
- 1ProposedA teammate, or your own note, says what should change.
- 2On trialIris checks the evidence. It reaches the prompts marked as a trial, and you can still dismiss it.
- 3AcceptedIt’s a rule now, and every run from here on follows it.
- 4DismissedA tombstone. It’s never proposed again.
21 of 24
lessons accepted when I went through a waiting pile
What it will never do.
- Name a client in public. The names come from your own notes, and they’re checked again before anything goes out.
- Clone a voice, or generate a real person’s face or voice.
- Put your face or your voice on anything but your own brand.
- Use a clip, a voice or a track whose licence doesn’t allow it.
- Spend past its caps. A run that would cross one stops and says so.
- Touch the record of your pieces. A model can’t write to it. Only code can.
- Skip the AI label. Whether a piece needs one is decided per platform, with the reason written down.
- Release a piece. That’s yours.
Chat windows, an agency, or a factory
- Who checks the factsChatGPT and Canva:You, if you rememberAn agency:Whoever has the accountThe factory:Code, then an independent editor
- Who does the workChatGPT and Canva:YouAn agency:Their teamThe factory:Eleven specialists, you approve
- Learns from your notesChatGPT and Canva:Only what you paste back inAn agency:While the same people stayThe factory:Your notes become rules
- Cost of one pieceChatGPT and Canva:Your afternoonAn agency:Inside the retainerThe factory:Metered, per piece
- A strategist in the roomChatGPT and Canva:NoAn agency:Yes, an agency wins thisThe factory:No, that’s still you
- Working at 2 a.m. on SundayChatGPT and Canva:If you areAn agency:NoThe factory:Yes, under a cost cap
One factory, built around your brand.
The Marketing AI Factory
€7,500
one-time · one brand
What you get
- Your brand system by Juno: colours, type, marks, templates, avatar and banners, checked in code
- Your bible and voice: who it’s for, the promise, the pillars, how you really write
- The eleven specialists, the four gates and the 60+ checks, set up for your business
- A dated plan of pieces for your page
- A first batch of finished reels, carousels and story sets, through every check
- Your desk, the calendar, Send to phone, the Studio and the cost meter, on your Mac
- Training, so you can run it yourself
Not in the price
- Your Claude subscription, the AI it runs on
- Paid voices and AI clips, if you want them. Free voices and stock footage are built in.
- Ad spend, and the accounts it posts to
- A Mac to run it on
How the build goes.
- 1
You tell me what you sell
Who it’s for, what you promise, what you’ve already written. That becomes your bible and your voice.
- 2
Juno builds your brand
Personality first, then colours as roles, type and marks, proven on real layouts.
- 3
The team gets your rules
What you’d never say, the one thing you ask people to do, and which topics are yours.
- 4
The first batch runs
Reels, carousels and story sets through every check. You review them, and your notes become lessons.
- 5
You take over
I train you on your own factory, on a real working day.
After the build
You run it.
It’s on your machine, on your own AI subscription, and you know how it works.
Or
I keep it running.
You send the sentences and approve what lands on your calendar. We agree a monthly fee on the call.
It’s for you if
- You know what you should be saying, and you never have the time to make it.
- You’d rather approve work than make it.
- You have real expertise, real projects or real opinions to draw on.
- Someone on your side can read a post and say whether it’s right.
It’s not for you if
- You want viral numbers guaranteed. Nobody honest can promise those, and I won’t.
- You don’t want to read what goes out under your name.
- You have nothing true to say yet. It checks facts. It can’t invent a business.
Fair questions
What does the €7,500 cover?
One factory built around one brand, on your Mac, with your brand system, your voice, a dated plan, a first batch of finished pieces, and training. The list is above. It’s a one-time build.
What will it cost to run?
You need a Claude subscription, and the work draws on its allowance. Measured at API prices, a finished video averages $2.75 on my floor, and a finished carousel or story set $1.48. Free voices and stock footage are built in. Paid voices and AI clips are optional, and each video has its own spending cap.
What do I need?
A Mac, a Claude subscription, iCloud Drive to send pieces to your phone, and someone on your side who can read a post and say whether it’s right.
One brand, or several?
The price is for one brand. Each brand is its own workspace with its own voice, plan and results, so adding another later is straightforward. Ask me what it would cost.
Will it sound like AI?
Mine did, and that’s why there’s a Hook Lab, a list of robot phrasing the code rejects, and a review that leaves a good opening alone. You also read every piece before it goes out, and when you say why you didn’t like something, it becomes a lesson.
Who checks the facts?
Vera writes down every claim with its source. Code checks that the numbers on the slide are in that ledger. Then Iris reads the finished piece and can fail it. Your review comes after all of that.
What if it gets something wrong anyway?
You catch it at your desk, say what’s wrong, and press Regenerate, which redoes only what you commented on. The note goes into that teammate’s memory. If the same failure keeps coming back, it gets written up as a new code check.
Can I edit things by hand?
Yes. There’s a full video editor and you can change any slide’s words. Your edits are kept, and if you make the same kind of change twice, it’s proposed as a rule.
Does it post for me?
Not by default. A finished piece lands on your calendar with its caption, and you send it to your phone and post it. Automatic publishing is built but tested only on stand-ins, not yet on a live account. It stays off until your accounts are connected and you switch it on.
Do I need to be technical?
No. You write a sentence and press Approve. I set it up and train you on it.
What does it run on?
An app on a Mac and Claude, Anthropic’s AI model. Every call is measured, with a cap per run, so a bad night can’t use up your whole allowance.
Will it use my face or my voice?
Never a cloned voice. Reels use a stock AI voice or none. Your own photos and videos can go in, only on your own brand. You’re asked about the rights once, and “Not sure” blocks every use.
Can it name my customers?
Case studies are anonymised by default. Every write is scanned for the client’s names, products, people and places before it’s saved.
Which languages?
English is what it’s proven on. It can write in Romanian and Hungarian, and the free voices for those two are for drafts only, so a finished video in either needs a licensed voice.
Tomorrow morning, the hard part is one sentence.
Tell me what you sell and where you post. I’ll tell you honestly whether a factory makes sense for you.
Or pick a time for a free intro call