Morebusters

Morebusters began as a work sample for an application with Christian Wolf, founder of MORE Nutrition. The brief asked for an ebook based on two of his videos and a web app for reviewing popular videos with false nutrition claims. Both had to sound like him. I spent nine days on the work and added several ideas of my own.

First, his voice

I started with a voice profile. I transcribed 104 Instagram clips and 11 YouTube talks with Whisper, then looked at sentence length, recurring words and the openings of his videos. The corpus contained 18,205 words. A model helped turn the analysis into a style guide for short hooks and longer explanations, with examples quoted directly from the transcripts.

"Datengestützt, nicht erfunden. Frequenzanalyse und echte Zitate erden das Profil, sonst kommt nur ein generischer Influencer-Ton raus."From my plan for the voice profile, written before the first line of the ebook

I used the profile in the prompts for the ebook and the app’s reply scripts. It gave me concrete examples to check the generated writing against.

The ebook

The ebook, “Heißhunger”, runs to 41 pages across ten chapters. I used the two source videos and the wider corpus, labelled scientific claims by strength of evidence, and included PubMed references. I built the layout in HTML with a pagination script and used the app’s design tokens. The chapter images were generated from Christian’s photos, as requested in the brief.

The inbox

The app uses Convex behind a TanStack Start front end. It takes videos from URLs, Instagram mentions or DMs, transcribes them and extracts the claims. An agent checks those against an index of known myths, PubMed and Europe PMC. I separated extraction from assessment and required a counter-argument before each verdict. The reviewer sees the sources and accepts or rejects the result.

"Die Community ist der Filter. Die KI ist der Analyst. Chris ist der Richter."The sentence the whole system is built around

The iOS app

I built the Expo app for reviewing videos on a phone. Each card shows a video, its claims and a priority score. Swipes accept, reject or set it aside; a separate action assigns it to a teammate. Accepted videos move onto a production board. The app is available on TestFlight with a demo login.

Morebusters iOS · swipe triage

Zerup Remix

Christian had announced an Instagram contest for people to share their syrup combinations. I built a working version with a public feed, a combination builder, voting and share cards. It was an extra proposal alongside the application brief.

More Arcade

I also built More Ninja, a Fruit Ninja-style browser game using 102 products from the shop API. It runs as a single canvas app without a backend or build step. The other seven games on the page are visual concepts. I included them as a proposal for a mobile channel where the brand could reach customers directly.

Follow the money

After submitting the required work, I investigated the commercial links between nutrition misinformation accounts. Starting with twelve accounts, the crawl grew to 143 accounts, 169 shops and 539 transcribed videos. I extracted 227 claims and mapped 1,801 links between accounts, shops and narratives. One model extracted claims with verbatim evidence; a stronger model checked them. In this sample, 71 percent of accounts with a personal affiliate ID spread documented falsehoods, compared with 24 percent of accounts where I found no money flow.

"Die erste Messung war falsch, und das war der wertvollste Moment."From the study notes: the first monetization pass missed most shops until every bio link was crawled rendered