Building an AI Influencer
From Scratch

I only teach what I've actually done. If I haven't built it, broken it, and worked out why it broke, I'm not standing in front of a room talking about it.

So when questions started coming in about AI-generated content creators — can you build one with real aesthetic quality? — I did what I always do. I built one.

There are plenty of AI influencers out there. Most of them are poor quality: uncanny faces, generic settings, low-effort content. Genuinely cinematic, editorially beautiful AI content is rare. I wanted to find out if it was possible, and what it actually takes to produce it.

What Blythe Is

Blythe Sterling is a fully AI-generated social media creator. She has a face, a voice, a perspective, a dog called Cosmo, and a flat in London. She's live on TikTok and Instagram.

She doesn't sell anything. She doesn't make tips content. She exists to answer one question:

What does AI look like at its most beautiful?

Art project first. Case study second.

How the Strategy Was Built

The very first brief, word for word:

"I want to set up social media accounts completely AI managed, and I'm not sure about the exact content. The goal is to monetize and make the most possible money with fully AI-generated content and AI automation. Run research on how that would be possible and which type of social media channels I could monetize. Help me set up the full workflow. Also, let me know what connectors or what tools I would need to set this up. The main goal is maximum monetization."

Blythe Sterling has nothing to do with any of that.

The first instinct was high-automation, high-volume, maximum reach — with an old money aesthetic as the packaging. That framing collapsed within the first hour. Wrapping beautiful visuals around a message I don't believe in isn't interesting. It's just a shell. The breakthrough was the opposite: the aesthetic is not the packaging. The aesthetic is the content. That one shift changed everything.

From there, a character needed to be built from scratch. A name, a face, a voice, a dog, a world. Most of the interesting details didn't come from prompting an AI to invent a personality — they came from asking what I actually think and feel, then translating that into a fictional person. The character that emerged is specific enough to generate consistent content, and layered enough that people want to follow her.

The gap between that first brief and the finished character is the case study.

All video generation runs through Higgsfield — specifically Cinema Studio 3.5 with a trained character model built on Blythe's likeness. Each video takes multiple generation attempts, reviewed manually before anything is approved.

The Trade-Off Nobody Mentions

This is the most important thing I learned, and also the most inconvenient:

You can automate, or you can have quality. You cannot have both.

Here is what that actually means in practice.

Automation removes quality control. Automated pipelines generate content without you in the loop. When you're not in the loop, things go wrong in ways you don't catch until they're already published.

AI ideas are still too cliché. Left to its own devices, AI defaults to safe, predictable, generic outputs. Every piece of content needs a human pushing it somewhere more specific, more interesting, more unexpected. That creative direction cannot be automated out of the process.

Getting it right takes multiple generations. A single video might require five or six generation attempts before it's actually good — wrong movement, wrong framing, wrong mood. You watch each one, decide what's wrong, and regenerate with adjusted parameters. There is no shortcut.

Automation only works when you're okay with mistakes. If the content is functional and faceless — informational, text-heavy, no character — then automation is fine. You can tolerate imperfection.

But the moment you want a face and a consistent character, automation breaks. A character requires visual consistency across every single piece of content. Every generation needs to be reviewed and approved by a human. You cannot delegate that to a pipeline.

The practical result: Claude handles the thinking side — content planning, captions, hooks, creative direction, character decisions. The actual generation is done manually in Higgsfield, frame by frame, with review at every step. That is not a limitation. That is what makes Blythe look the way she looks.

What I Got Wrong

Wrong aspect ratio, multiple times. Generated several videos in landscape (16:9). TikTok requires portrait (9:16). Wasted credits, regenerated from scratch. This happened more than once.

Wrong model for the first portraits. Used the wrong AI model and the results were soulless — obviously AI, uncanny, unusable. Scrapped everything and started again with the right tool. The quality difference was immediate and dramatic. In AI generation, model selection matters more than prompt quality.

Third person text overlays. Wrote Blythe's first caption as "Her flat is quiet in a way she chose." Immediately wrong. She speaks in first person. Always. Now a hard rule.

Estimated costs at four times the actual price. Made early decisions more conservative than they needed to be. Check real costs. Do not estimate.

Changed the dog's breed twice. A Dalmatian's spots are unique in every AI generation — no two frames produce the same pattern. Visual consistency is a technical constraint, not just an aesthetic preference. A uniform coat is the right choice for an AI character.

Where This Is Going

Blythe is in Phase 1: building on TikTok and Instagram, learning what performs, developing the character further. She now has a voice, generated with ElevenLabs.

What comes next is still being figured out. More automation and more reach are both possible — but only if they don't require compromising the quality that makes the account worth following. For now, quality first.

This is a live experiment, not a finished case study. I will update this page as it develops.

Last updated: June 2026.