Meet Marguerite: We Built an AI Influencer to See If She Can Hit 21,000 Followers by Halloween

I want to know if an AI-generated person can build a real audience, running the same growth mechanics a human creator would use. Not a thought experiment. An actual account, actual posts, actual numbers I can watch climb or stall in public.

So we built one. Her name is Marguerite Vallée, she’s 24, and by day she’s a costume-shop girl who thrift-flips her outfits into something better. She does not exist. Every photo, every video frame, and her voice were generated.

How she gets built is also part of what we teach in weeks two and three of the AI Secret Sauce class at Cowork Chilliwack: character sheets, trained identities, and the accounts that post themselves.

Why Halloween

Timing a launch to a season that already has built-in attention is free distribution. Halloween gives Marguerite sixteen costume looks between now and October 31st, each one a two-day content cycle: a build-up/teaser day, then a full mirror-before-to-reveal-after clip on a trending sound. The looks mix a couple of evergreen classics with 2026’s biggest pop-culture moments, the kind of thing people are already searching for this time of year, plus one Día de los Muertos calavera look, the only one that uses actual face paint. Everything else is thrifted clothing, props, and masks, because not everyone doing this has an SFX artist on call, and that’s kind of the point.

Why 21,000

That number isn’t arbitrary. It’s the benchmark set by a comparable case we studied before starting this: a single AI character, run for 21 days, that pulled 8 million views and 21,000 followers, and turned a $9 digital product into over $2,000 in sales along the way. We wanted a number that came from evidence, not a guess, and then we wanted to see if we could actually hit it, in the open, with a character built from scratch instead of copying someone else’s playbook.

I’m not promising Marguerite gets there. I’m running the experiment and posting what happens either way: the wins, the flat weeks, the costumes that didn’t land.

What’s actually behind her

Marguerite’s face is locked to a trained Higgsfield Soul ID, so every generation, whether it’s the sewing desk, the mirror wall, or the reveal shot, comes back as the same person, from the same angles, with the same small details: a choker, one hoop in each ear, a pin-cushion cuff on her left wrist. That’s what makes her recognizable at a glance instead of looking like sixteen different AI renders wearing sixteen different costumes. Her voice was designed once, cloned, and locked, so every video sounds like her too. None of this is hidden: her accounts carry the platform AI-disclosure labels required in 2026, because an unlabeled AI profile doesn’t get recommended anyway, and because that’s the only way to run this kind of experiment honestly in front of an audience that deserves to know what they’re looking at.

Eventually, if the audience is real, Marguerite sells something: a “16 Halloween Looks for This Season” guide, the same kind of low-priced digital product the case study we benchmarked against used. Not yet, though. First, prove someone’s actually watching.

The posting runs on the same API, not a browser tab

Marguerite’s Facebook Page and her Instagram account both run through the Meta Graph API, on the same system-user token that already covers the AI Secret Sauce and BC Latin Dance Pages. One token, three brands, nothing to log into by hand.

That’s the part that actually removes the busywork. Publishing a reel doesn’t mean opening Meta Business Suite and clicking through an upload flow every single time. It means typing an instruction in plain language into the Claude Code terminal and letting the API call go out from there: what to post, where, with what caption. Follow Marguerite on Facebook or on Instagram and you’re watching accounts that get posted to by a terminal command more often than by a phone in someone’s hand.

Her first two reels

Two are live so far, both built the same way: a script in her voice, the locked face, a generated reveal on a trending sound.

What I’m watching for

Not vanity follower counts in isolation. Saves, shares, and comments first, the same signals that separated the case studies that worked from the ones that didn’t. I’ll post updates on this as the costumes roll out. Follow along, watch the number move, and see if it actually gets there by October 31st.

---

Keep reading

• We Tried Building an AI Team on Grok. Here’s Why We Moved to Self-Hosted Rakazo Instead: the bot platform behind the accounts posting for Marguerite, BC Latin Dance, and AI Secret Sauce
• I Cloned Myself and My Dog With AI Video: the same character-consistency pipeline (Higgsfield, trained identities) Marguerite is built on
• Shipping Fast With AI? Here’s the Server Security Stack You’re Probably Skipping: how a system-user token like the one running Marguerite’s Page stays out of plaintext