In about two hours, you walk away with a complete production pipeline for your own AI influencer — the same workflow used at MissAI to launch new characters from scratch.
✓A locked AI character that looks identical across every post, forever.
✓A short-form video pipeline for Instagram Reels, TikTok, and YouTube Shorts.
✓A reusable brand template: drop any brand's assets in a folder and your influencer makes on-brand video automatically.
✓A folder architecture that scales — designed so brands can one day upload assets and your influencer auto-advertises for them.
Why this works: the hard part of an AI influencer isn't generating an image. It's stopping her from drifting between posts. This guide is built around that single problem.
Who this is for: operators, founders, marketers, and creators who want to launch a consistent AI persona on Instagram and TikTok and turn her into a brand-deal vehicle. No developer background required.
The Kit
What you need.
Five inputs. Everything else gets built for you inside this workflow.
Claude
Cowork mode (or Claude Code). Writes the system, runs the prompts, and orchestrates Higgsfield. Available during research preview.
Higgsfield
An account with character training enabled (Soul plan or higher). Generates the actual images and videos.
Reference Photos
5–20 photos of your character. Real photos of yourself, AI-generated headshots, or a mix. Front-on, side, three-quarter — variety matters.
One Folder
A normal folder on your computer that Claude can read and write to. That's it.
Time: 1–2 hours, plus ~10 minutes of background time while Higgsfield trains your character. You can walk away during training.
You will NOT need: VS Code, the terminal, or any developer tooling. A coding background. Pre-existing brand deals — the workflow generates pitchable content first, then you sell.
The Build
Seven steps.
Step 01
Set up your workspace
Open Claude with Cowork mode enabled. The job here is to give Claude a folder on your computer it can write into.
Create a new folder on your computer. Name it something obvious like AI-Influencer.
Open Cowork and start a new chat.
Use the folder selector to point Claude at the folder you just made. Approve access.
Open a side window in Finder (or File Explorer) and keep that folder visible. You'll watch Claude write into it.
Mindset: don't overthink the folder. It's just storage. The intelligence lives in your Claude chat and Higgsfield — the folder is the filing cabinet they both reach into.
Step 02
Drop in reference photos
Inside your folder, drop in 7–20 reference photos of your character. These are what Claude will study to write your Character Bible, and what Higgsfield will use to train her permanent face.
What makes a strong reference set:
Front-on face shot — captures eyes and bone structure
Side profile — captures nose and jawline
Three-quarter angle — captures hair shape and styling
Variety of expressions, outfits, lighting — teaches the AI her range
iPhone-quality, not studio — candid is what you want for social. Polished portraits make her feel like a stock model.
What to avoid: heavy beauty filters or distorted angles (the AI bakes those in). Seven copies of the same shot. Only one outfit or lighting condition. Photos older than two years if using real photos of yourself.
Step 03
Send Claude the build prompt
This is the single message that triggers the entire system build. Copy it exactly, paste it into your Claude chat with the reference photos visible in the folder, and let it run.
Paste into Claude Cowork
I'm building an AI influencer for Instagram Reels and TikTok. The
reference photos in this folder are her locked look — and she has
to look identical across every single post forever, because the
whole business depends on her being recognisable.
She'll do two things: her own short-form video content, and paid
brand-deal videos. For the brand deals, I want a system where I
drop a new brand's logo, product shots, and website into a folder,
and she produces on-brand short-form videos for them automatically.
Build me the full system — folder architecture, her locked character
spec so her face never drifts, a master prompt for short-form video,
and a reusable brand-folder template I can copy for every new brand.
Higgsfield MCP is connected for generation.
What Claude does with this:
Studies your reference photos and writes a Character Bible — her locked face, hair, eyes, build, and aesthetic
Builds the folder architecture (Brands/, prompts/, Outputs/, etc)
Writes a master video prompt that bakes in the Character Bible automatically — every video uses the same locked character
Creates a reusable _TEMPLATE/ folder you'll copy for every brand deal
Step 04
Review what Claude built
After about a minute, Claude hands back a folder structure that looks like this:
your-influencer/
├── README.md— control panel / run guide
├── Character-Bible.md— LOCKED identity spec
├── Reference-Image/— your reference photos
├── prompts/
│ └── video-master-prompt.md
├── Brands/
│ ├── _TEMPLATE/— duplicate for new brands
│ └── [your-first-brand]/
└── Outputs/
└── Videos/
Open the Character Bible first. Inside you'll find a Locked Identity Block — a verbatim text description of your character's face, hair, eyes, build, and style. This block is pasted into every single image and video generation, forever. It's what stops her from drifting between posts.
The dual-lock: most AI influencer experiments fail because the character drifts between posts. The Locked Identity Block PLUS a trained Soul character is the dual-lock that prevents drift.
Step 05
Connect Higgsfield + train your Soul
Higgsfield generates the actual images and videos. Two parts: connect the MCP, then train a permanent character.
Confirm a green dot appears next to Higgsfield in your connector list
Part B — Train your Soul character:
Go to higgsfield.ai → Characters → Create character
Drag-drop 5–20 of your reference photos
Name her (e.g. your character's first name)
Click Train. Walk away for ~10 minutes — training runs in the background
When you see a Ready badge and a Soul ID (a UUID), copy that ID
Paste the Soul ID back into your Claude chat
The Soul ID: every future image or video calls this ID and your character appears with the exact same face every single time. This is the foundation of brand-deal-ready output.
Step 06
Make your first video
Videos use the same character (same Soul ID = same face) but a different prompt and a different Higgsfield tool. The pipeline: generate a still for each beat, animate it, stitch the beats into a 15–30 second clip.
How to run it:
Open prompts/video-master-prompt.md from your folder
Copy the code block, paste into a Claude chat with Higgsfield MCP connected
Monologue to camera. Best for opinions, takes, stories.
Highest comments
GRWM
Get Ready With Me. Product appears as a natural step.
Best for beauty / supplement
Montage
Silent b-roll beats with on-screen text overlays.
Best for launches / mood
POV
First-person, hands and objects only. No face.
Highest replay rate
Transformation
Before / after structure across the clip.
Best for fitness / skincare
Day-in-Life
5–7 short clips of her day, one product moment baked in.
Best for premium / lifestyle
The 5-beat video architecture
Every video is structured as 4–6 second beats, stitched together. Hit these five beats and you have a complete short-form post:
0–2s
Beat 1 — Hook
Visual + one spoken or on-screen line. Under 8 words. Stop the scroll.
2–6s
Beat 2 — Promise / Tension
Set up why we should keep watching. One specific question, claim, or stake.
6–18s
Beat 3 — Payoff / Content
The actual value. The tip, the routine, the product moment. Natural integration only.
18–25s
Beat 4 — Reinforce / Twist
One last hit — a punchline, a result, a contrarian close.
25–30s
Beat 5 — Soft CTA
"Save this." "Link in bio if you want to try." Never "follow for more."
Voiceover: if your Higgsfield plan supports VO generation, the prompt passes the script through automatically. If not, Claude hands you the VO script as a separate block — record it in Voice Memos or generate via ElevenLabs.
Budget: each beat costs two generations (image + video), so a 20-second video runs through 8–10 Higgsfield calls. Use the Check Balance and Show Plans tools to know your runway.
Step 07
Onboard your first brand deal
This is where the business model lives. Each brand is one folder duplication + one brief + one Higgsfield Assets upload.
Part A — Duplicate the template:
Inside Brands/, duplicate the _TEMPLATE/ folder
Rename it to the brand slug — lowercase, hyphens (e.g. luxe-resorts, protein-co)
Drop the brand's assets in: logo.png, one clean product image per SKU, plus any lifestyle imagery
Open brand-brief.md and fill in identity, range, claims, audience, voice, visual rules, content angles, and hard NOs. Pull from real packaging and the website — never invent claims.
Part B — Upload assets to Higgsfield:
Most common mistake: skipping this step. Higgsfield has never seen this brand's packaging. Without real product photos in its library, it invents what the product looks like — wrong colours, wrong logos, gibberish text on labels.
Go to higgsfield.ai → Assets (top-right corner)
Drag-drop the brand's logo and product images into the Assets library
Tell Claude the assets are uploaded — it will find them via the MCP and pass them as media references on every slide where the product appears
Part C — Run the video prompt with the brand folder:
Run the video master prompt with Brand folder: pointing at your new brand folder. Your influencer now produces on-brand video — with the same locked face she has on every other post.
Every new brand from here is one folder duplication, one brief, one upload. The system scales as wide as you can sell.
Field Notes
Five mistakes to avoid.
01
Drift
Your character looks slightly different post-to-post.
Fix: never skip the Locked Identity Block. Always pass the trained Soul ID. If a generation comes back off-model, regenerate before you post — don't let drift compound across the feed.
02
Hallucinated products
The AI invents what the product looks like.
Fix: upload real product photos to your Higgsfield Assets library and pass them as media references on every slide. Without a reference image, the model is guessing — and the label, colour, and shape all become a coin flip.
03
The bot schedule
Auto-publishing on a perfectly-timed interval from a server.
Fix: algorithms recognise the pattern instantly, and brand-new accounts get throttled first. Generate the content with AI — but tap publish from your own phone, ideally not at the same minute every day. Anything under 30 days old: keep it fully manual.
04
Landing-page hooks
Your opening line reads like a product page headline.
Fix: if your hook could fit on a billboard, it's too generic to stop a scroll. Write like you're telling a friend something you only just figured out — specifics over slogans, numbers over adjectives, an angle that feels personal rather than corporate.
05
Over-production
Your content looks like an advertisement.
Fix: studio lighting, perfect framing, gimbal-smooth motion — every signal that says "a team produced this" gets demoted by the feed. Bake the opposite into every prompt: soft daylight, hand-held framing, the kind of moment you'd actually film between things.
What's Next
Where you go from here.
Once your first video is out the door, three obvious next moves — in the order we recommend tackling them.
01
Turn the workflow into a Skill
Ask Claude to convert your video master prompt into a reusable Skill. Say "create a video for [brand]" and the Skill handles everything.
02
Schedule it
Claude can run your video Skill on a schedule. Fresh clips land in Outputs/Videos/ daily or weekly. You wake up, review, and post from your phone.
03
Scale to multiple brands
Each new brand = one folder duplication and ten minutes writing the brief. Running 5–10 brand deals takes the same effort as running one.
The final form: an interface where brands upload their logo, products, and website URL, and your influencer advertises for them automatically. The folder system you just built is the contract for that interface.