A language-course studio that writes, renders, and publishes itself
I built Fablingo — a network of six faceless YouTube language-course channels, published as free courses at fablingo.academy — where a multi-model AI pipeline turns a locked curriculum into finished, illustrated, narrated video lessons and publishes them on a schedule. No presenter, no camera, no editor, almost no human in the loop. It’s the clearest proof I have of what disciplined, production-grade AI actually looks like.
The situation
Anyone can prompt an AI to make one video. Almost no one can make a whole course — a coherent curriculum, a recurring cast, lessons that genuinely teach, and quality that holds across hundreds of episodes — without a team of writers, illustrators, editors, and voice talent. I wanted to prove that gap can be closed with engineering, not headcount: a studio that runs itself and whose 3 a.m. output is indistinguishable from what a human team would sweat over.
The approach
The trick isn’t a clever prompt — it’s treating the language, the story, and the pedagogy as structured data, and wiring several models together so they check each other.
Curriculum as data, not vibes
Each channel runs on a locked, CEFR-aligned curriculum — the teacher of the day, the grammar taught "in disguise", the exact phrases, and the story beat for every episode. The AI obeys this plan; it never improvises the syllabus.
An author model and an independent editor
One model writes the lesson as a chapter in an ongoing neighbourhood story. A different model then audits that draft against the locked plan — every required phrase taught, target text exact, length real — and repairs gaps. No model grades its own homework.
Vision-checked illustration & native voice
Every scene gets a house-style illustration and native narration. A vision model validates each image against the lesson topic and regenerates on a corrected prompt before it can ship — so the words and the pictures can never silently drift apart.
Distributed render behind hard safety gates
A fleet of ordinary machines renders the hand-drawn animation in parallel. Hard invariants refuse to ship anything broken — a silent-audio gate, no blank cards, no endings clipped mid-word. The system would rather block a video than publish a flawed one.
Publishing that needs no babysitter
Episodes upload private, then release themselves at each language’s local prime time on a fixed cadence, respecting each channel’s quota. The whole network can be generated and scheduled months ahead, then left to publish with nothing online.
Outcomes
- Six live channels across six languages — each with its own cast, dialect, and script
- A locked ~180-episode curriculum per language, pre-A1 → C1, that escalates coherently
- Generate → validate → render → publish runs end-to-end with almost no human in the loop
- Repeatable and horizontally scalable — add a machine for throughput, add a shape for a new language
Fablingo is the difference between “I prompted an AI to make a video” and “I built a studio that runs itself.” The discipline behind it — treating domain knowledge as data, using models to check one another, and enforcing hard invariants so automation is safe to trust — is exactly what I bring to companies adopting AI.
Want AI that ships real work in your business, not demos? Let's talk.
I read every message and reply personally.
See the live channels & the full build → · Take a course at fablingo.academy → · ← Back to case studies