Autonomous Video Studios
PART I

The systems at a glance

What the two systems do, who they serve, and what they have produced in their first month.

HomePART I · The systems at a glance
CHAPTER 1 · OVERVIEW

What the two systems do

Faceless channels — channels whose videos carry a voice and animation but no presenter — are one of the fastest-growing formats on YouTube Shorts, Instagram Reels and Facebook. They are also expensive to run by hand: a good one needs research, a script, animation, narration, editing, captions, thumbnails, posting at the right hours on several platforms, replies to comments and a constant eye on what works. These systems do all of it, continuously, for dozens of channels at once.

An episode, in one paragraph#

A topic planner proposes a surprising, checkable subject for a channel — what if a Category 6 hurricane hit your city?, a 2,300-tonne ship engine turns slower than your car at a red light — and checks it against everything the channel has already said. An agent (Claude Fable 5.2) then writes the episode in one of two pictures. As code: a WebGL2 scene with real physics, true scale, a moving camera and a narration script, built with the channel's own renderer kit, style rules and the lessons learned from its past failures. Or in Blender: a photoreal 3D scene assembled from free, open-licence models, textures and skies, with real street maps where a place matters. A GPU box renders it; a Chirp 3 HD voice reads the script; a speech recogniser on the same box hears the narration back and times every word, so the captions are the voice, word for word. A critic model, an entertainment reviewer and a series of mechanical gates check facts, readability, pacing, units, the caption sync and the story of the 90-second edit. Only then does it enter the channel's library, ready for its slot.

One episode, end to end — one to two hours, two episodes at a time on each GPU
1Topic and angle

planner picks a surprising, checkable topic per channel; originality vs past episodes

2Channel kit

the channel's renderer, style rules, palette, music and its distilled lessons

3Episode agent

Claude Fable 5.2 writes the scene — as code, or as a Blender build — and the script

4GPU render

WebGL2 in headless Chrome, or Blender 4.5, frame by frame on an NVIDIA L4

5Voice and captions

Chirp 3 HD voice; every word heard back and timed, so the captions are the voice

6Quality gates

critic ≥ 8, entertainment ≥ 7, facts, units, caption sync, the 90-second edit's story

7Library and archive

ready stock per channel; every file versioned, copied hourly to a second bucket

8Publish on slots

YouTube, Facebook, Instagram, LinkedIn, X, TikTok, Threads; guarded, human cadence

↺ a failed gate sends it back with the reviewer's notes
⟲ analytics, R&D findings and the lesson of every failed gate flow back into topics, kits and briefs
Figure 2. How one episode is made. Every stage is automated; a failed gate sends the episode back with the reviewer's notes, and what is learned flows back into the next briefs.

Two systems, one engine#

The original programme
  • its owner's own operation: 65 channels in ten projects
  • 16 live on YouTube, Facebook and Instagram
  • work starts on schedules and the Studio's buttons
  • paid from the owner's model and GPU budgets
ChannelForge
  • any customer's workspaces and teams: 49 channels
  • autopilots, the app, a REST API and an MCP server
  • credits: plans, packs, Stripe and PayPal
  • 166 of 198 features switched on
One engine, in both
Topic
Agent
GPU render
Voice, captions
Quality gates
Publish
Its own estate
  • a control host and the portal on Google Cloud
  • 30 L4 boxes in 13 pools, two AWS accounts
  • the media archive, copied every hour
  • each channel's own platform accounts
— no shared machines, no shared data —
Its own estate
  • an app host and Cloud SQL on Google Cloud
  • PostgreSQL 16 with 313 tables, and Redis
  • its own L4 fleet, every job sandboxed
  • each member's own platform accounts
Figure 3. Two systems, one engine: the same pipeline in two estates that share no machine and no data, so either can be sold on its own.
Original programmeChannelForge
Who it servesIts owner: 65 channels across ten projectsAny customer: workspaces with their own channels (49 today)
Web addresscontent.storeswiki.com (private)forge.storeswiki.com (public)
How work startsSchedules (n8n), a scheduler, the Studio's buttonsAutopilots, schedules, the app, the REST API, MCP
Picturescode-rendered scenes and Blender scenescode-rendered scenes; Blender built in, switched on per workspace
PublishingYouTube, Facebook, Instagram, LinkedIn, X, TikTok, ThreadsThe same plus Bluesky and Pinterest
MoneyThe owner's model and GPU budgets per projectCredits: plans, packs, Stripe and PayPal
GPUs30 NVIDIA L4 boxes in 13 pools, two AWS accountsIts own L4 fleet, isolated from the original
DataPostgres (portal), files and ledgers (control host)PostgreSQL 16 (Cloud SQL, 313 tables), Redis
MediaGoogle Cloud Storage, versioned, copied hourly to a second bucketGoogle Cloud Storage, per workspace

The original programme came first and remains the owner's own operation; ChannelForge was built from a copy of it on 18 October and grew into a product with its own control plane, GPU fleet and database. The two share no machines and no data — deliberately: the original can be sold with its channels, and the SaaS as software, independently.

What makes them different#

  • Original pictures, not stock or generative video. Every frame is computed: from code — simulations, maps, cutaways and true-scale comparisons — or rendered in Blender from free CC0 assets. The result is distinctive, cheap to render, and free of the licensing and likeness risks of stock footage, avatars or text-to-video models; every Blender episode says so in its first seconds: no AI-generated video.
  • Captions that are the voice. Since 13 November every new short shows the narration's own words, each one lit as it is spoken, timed by a speech recogniser that hears the finished soundtrack; a gate measures the sync before the video may be posted.
  • Quality before quantity. A video that scores below 8/10 with the critic, fails a fact, misreads a unit or runs captions too fast is held, never posted. Films have their own floor per act.
  • Self-healing, with a lever for everything else. A sentinel checks the whole pipeline every five minutes and acts: it makes up a missed slot from surplus, repairs a failed post with backoff, refills stock that runs low, moves work to another GPU, zone or cloud account, and restarts what died. Whatever it cannot fix alone — an account to reconnect, a platform to top up — appears in a failure history beside a button that fixes or retries it.
  • Human cadence. What the systems write on the platforms by themselves follows no fixed clock: random sessions in each account's own time zone, uneven gaps and varied shapes, so channels never look like bots.
  • Measured, not guessed. Every post's views, likes, comments and shares are read back; an R&D wing studies what works per channel and feeds it into the next topics and briefs.