Ninety minutes. One character. By 8:30 tonight you'll have a named AI influencer, a visual guide you can generate her from forever, and a clip of her moving. You're building alongside me, not taking notes.
What I will promise: a workflow. Decide who she is, generate her, lock her, put her anywhere, make her move. That skill outlives every tool on tonight's list.
On the record from the tool or platform itself. Build on these.
What the FTC or a platform recommends. Strong signal, worth following.
What I see working repeatedly across builds. Reliable, not guaranteed.
My hunch. Worth testing on your own account. Never stated as fact.
Everything's current as of August 2026. When I don't know something, I'll say I don't know and hand you a way to test it.
You want a public brand without being the face of it. She's the one on camera.
You need a spokesperson who never reschedules. Same skill, different job.
Client creative at volume. Ten ad variations before lunch, no shoot day.
A model for your products who works for credits instead of a day rate.
Same workflow underneath all four. What changes is what "done" looks like for you — so pick one now.
One of the four. Then write one sentence: what would make tonight worth it for you specifically? "I want a faceless fitness account" is a lane. "I want to try AI" isn't.
Type your lane in the chat. It tells me how to aim the demos, and I do read them.
People spend six weeks looking for a fourth step. There isn't one. It feels harder than this because most people skip move one and try to fix it in move three.
That's the whole list. No voice tool, no separate lipsync tool — Omni generates the audio with the video in one pass. Fewer subscriptions than you were expecting.
Before you generate a single face. This is the part everyone skips, and it's the part that decides whether you end up with a character or a folder full of strangers.
Everybody can make the left column now. The right column is the whole ballgame.
Most people open a generator, make somebody gorgeous, and then try to figure out what she posts about. Backwards. The archetype decides her age range. The age range decides her wardrobe. The wardrobe decides her locations. All of it decides what she can plausibly say.
This is the document you'll paste into every AI you use from here on. Write it once, use it forever.
Nationality is the highest-leverage field on the list. It's not the same as race.
Say she's Jamaican-American and you've just decided her hair, half her wardrobe, her music, her slang, and the voice her captions get written in. Say "Black woman, 30s" and you've decided nothing. Same for Dominican versus Brazilian versus African-American — could all read as Black, completely different content.
She's chronically twenty minutes late. She can't stand small talk. She's weirdly competitive about parking spots.
Perfect people aren't interesting. A character with no friction has nothing to post about except how she looks — and that runs out in about three weeks. One specific flaw buys you a hundred captions.
Don't reinvent the wheel. Look at what's working right now and make your version. Save three or four references — that's plenty.
You don't have to describe what you're looking at. Make the model do it, then correct what it got wrong.
Eight fields in, a full character bible out. Then — this is the important part — I throw two things out on camera and tell you why.
Watch what I reject. The generated version is a starting point, not an answer. The judgment is the job.
Note the last two lines. Without them you get a Pinterest board. With them you get something you can actually generate from.
Write the eight. Run the prompt. Edit what comes back — throw out the two things that don't sound like her. The only hard rule: name her before the timer ends. The name is what makes her real.
Don't overthink the face yet. Get the who right and the what-she-looks-like gets easier.
Drop her name and her niche. I'll read a handful out and flag the ones that are going to be a problem — usually because the handle's unsearchable or the niche is three niches wearing a trench coat.
The face, then the reference set. This is the asset — every image and every video you ever make of her starts from it.
This isn't the picture you post. It's the picture everything else gets built from — so you want it clear, not beautiful.
That's the phone. If a model won't give you 9:16, take 2:3 — it's the closest thing. Set it before you generate, not after, because regenerating a whole batch in the right ratio is the most annoying five minutes in this process.
Identical prompt, identical ratio, identical resolution, run through both. You'll see noticeably different people come back — and one of them will be the one.
I'll reject most of them out loud. The picking is a skill and nobody ever teaches it.
Generate twelve, keep one. That's normal. Delete the misses immediately so the grid stays readable.
Your default. Does the thinking and the picture. Already in your subscription, so iterating is free.
The second opinion. Run it when ChatGPT refuses, or when you want a different face from the same words.
Via Google Flow. Best at likeness transfer and surgical edits — change one thing without rerolling.
Two platforms open means a refusal costs you thirty seconds instead of an afternoon.
Both hold identity for up to five people and take up to fourteen reference images. This workflow is a volume game — take the speed.
"Using this exact likeness" carries most of the weight. If it drifts anyway, say it again more plainly — sometimes politeness genuinely helps.
This is the one you came for. I build the visual guide live — headshot, three-quarter, full body, angles — and then put her in five completely different scenes. Same face throughout.
Watch her face across all of it. That's the whole promise of tonight, happening in real time.
Keep the whole guide in one folder named for her. In the same folder, a plain text file: her name, age range, nationality, hair, eyes, build, height, and the two or three details you're keeping consistent.
You'll paste that text next to those images constantly. That pair — the images and the file — is what "she exists" actually means.
Same face, twenty different situations. Casual clothes. Something dressier. Outside. Inside. Different light, different time of day. Sitting, standing, walking.
Give her two or three small physical details that show up every single time. A freckle pattern. A particular nail colour. An ankle bracelet. A tiny scar.
It's an instant consistency check — if the detail's missing, the generation's wrong and you can see that in half a second. And it makes people feel like they're looking at a person, even when they couldn't tell you why. To get one, just say it in the description.
Write her identity lock — one tight paragraph, locked fields only, no adjectives you can't defend. Then list your four framings and the five scenes you'll generate first.
This is the artifact you came for. Everything else tonight builds on top of it.
I'll take two and tighten them live. It's almost always the same fix: cut the adjectives. "Striking," "captivating," and "flawless" describe nobody, so the model invents somebody new every time it reads them.
Specific and plain beats beautiful and vague. Every single time.
What breaks, why it breaks, and what to do about each one — including the failures nobody puts in their tutorial, because everybody edits those out and then you think it's just you.
Every time you hit generate, the model builds a person from scratch out of your words. "Brown hair, green eyes, mid-thirties" describes about nine million people. It picks one. Next time it picks a different one.
It isn't forgetting her. It never knew her. Your job is to stop describing and start referencing — give it something to hold, not something to imagine. That's what the visual guide is for.
Consistency is a workflow, not a prompt. Nobody's gatekeeping a magic sentence from you.
If you remember one line from tonight, make it this one. Every "secret prompt" video you've watched was selling you a shortcut around a process that only takes a few steps anyway.
Ninety percent of drift is somebody accidentally editing a locked field. Usually with an adjective they didn't think counted.
The first block never changes. Not one word. The other four are yours to play with. That single habit prevents most of the drift you've been fighting.
On purpose, live. Identity drift. The age slide, where she gains eight years between shots. Melted hands. Wardrobe that forgets itself halfway through a set.
Nobody else shows you this part. Which is exactly why everybody thinks they're the only one it happens to.
The rule underneath all of it: re-roll before you re-write. Generating again is cheap. Re-engineering a prompt to argue with a model isn't.
New roll of the dice. New chance she drifts. You might lose an image you already liked, and you'll spend credits finding out.
Swap the jacket. Put the product in her hand. Change the background. Her face never re-enters the conversation.
Nano Banana is the best at this. The rule: if the face is right and something else is wrong, never regenerate.
Moves two and three. Writing for a clip is not writing an essay — then take a still from the visual guide and make it move.
A clip is ten seconds. That's about twenty-five words. Read it out loud at normal speed before you render — AI writes long, and a script that's secretly forty words gets cut off mid-sentence after you've paid for it.
Always three versions. You want to choose, not to accept. Same reason you ask for ten captions instead of one.
This is the part people get wrong. Don't pick your prettiest still — pick the one with an obvious movement already latent in it.
Someone holding a cup can raise it. Someone mid-stride can keep walking. Someone standing rigid against a wall has nowhere to go, and the model will invent something strange to fill ten seconds.
Three names, one Google video stack. People think they're competitors. They're layers.
Omni generates the sound in the same pass as the picture. Dialogue matches lip shapes. Footsteps land on the right frame. Room tone stays consistent across the clip.
That's the difference between a clip that reads as real and one that reads as dubbed — and it's why we're not buying a separate voice subscription tonight.
One frame out of the visual guide we built in Part 2, turned into a moving clip with her voice on it. Start to finish, narrated, including the bit that's clunkier than it should be.
Then I'll show you the same still in Higgsfield on Seedance, so you can see what the alternative gives you.
Describe one small physical thing. The instinct is to write a scene — resist it. You're directing ten seconds, not a short film.
Consistency in video is inherited. If it drifts badly, the problem is the still you fed in — not the video model.
Pick one scene from your list. Write the motion line — one small physical action, camera still. Then write the fifteen seconds she says. Fifteen seconds is about forty words. Count them.
That's a publishable clip, fully specified on paper. You could render it tonight.
The part that pays without an audience. Same skill you just learned, completely different runway — and the one place tonight where there's a line you genuinely can't cross.
An audience. Trust. Relevance to a brand. Months of consistent posting before anyone's interested. Then a negotiation.
One business with a product and an ad budget. That's the entire requirement. You could have a client before you have a follower.
Both are real. One of them starts paying a lot sooner, and it's the one nobody makes videos about.
"Here's how this works." She demonstrates the product, explains the features, delivers the pitch. Same job a hired actor does in any ad.
"I bought this and it changed my life." A fictional person claiming a personal experience she never had. That's a fabricated testimonial.
Presenting a product is advertising. Inventing a customer is deception. Everything else tonight is style; this one's a rule.
Five blocks, twenty-five seconds. It scales because it never changes.
Sample product, live build. Three hooks written, one picked, product edited into her hand with Nano Banana, rendered and voiced in Veo, exported vertical and ready to upload.
Five minutes, start to finish. That's the whole deliverable a business would pay you for.
Write all five for every product, then pick. Never write one and hope. The first hook you think of is almost never the best one.
Your product, a client's product, or something made up — doesn't matter tonight. Three different shapes from the pack. Then star the one you'd actually shoot.
Fast and loose on purpose. Hooks are cheap. Write bad ones quickly and the good one shows up.
Six minutes, no income claims, no follower promises. Where she lives, a rhythm you can hold, and the four rulebooks that decide whether this survives.
Photorealism is not a content strategy. Nobody follows render quality.
People follow a point of view, a recurring format they recognise, or a story they want the next part of. If your whole account is "look how real she looks," you've built a tech demo — and tech demos get about three weeks of attention.
Time, money, speed. Those three decide your cadence, not what somebody doing this full-time posts.
The right column starts paying sooner and needs zero followers. Most people chase the left one because it's the one that gets posted about.
I'm a marketer, not a lawyer. Guidance, not legal advice — and it moves, so re-check it.
One platform. Three pillars. One money model from the map. Then the specific thing you'll do on each of the next seven days — "generate the visual guide," "set up the profile," not "work on it."
This is the difference between a great Monday night and a forgotten one. Be specific enough that Thursday-you can't wriggle out of it.
Number five is where this stops being a hobby. It costs you one email.
Credit pricing moves constantly, so I'm not printing numbers that'll be wrong by Friday. Check them the day you buy.
Tonight was ninety minutes. If you want the version where somebody's in the room with you every month while the tools keep changing — that's what the School is.
I'll stay on as long as there are hands up. Grabbed the playbook or the cheat sheets? Your links go out tonight, right after we wrap — check your inbox in about twenty minutes.
Go build that visual guide tomorrow. Seeya soon.