The zero-headcount studio stack
Three YouTube channels, no editors, no producers, no social manager. Here is every role in the studio, who or what fills it, and the one place a human still has to say yes.
The studio runs two story-driven documentary channels plus a faceless channel that is being relaunched as Shorts-first. Nobody is on payroll. The work is split across AI agents with written job descriptions, and the only human in the loop is the owner, who approves anything that gets published, sent or paid for.
This log exists because almost everyone selling AI to media companies shows slides. We would rather show the operation, including the parts that break.
The org chart
Each agent has one job, a written brief, and a reviewer. That structure matters more than which model sits behind it.
| Role | What it owns | Hands off to |
|---|---|---|
| Executive Producer agent | The slate, the quality gate, packaging (titles, thumbnails, descriptions) and weekly channel performance | Production tools, then the owner for approval |
| Production tool (Muse) | Turns approved ideas into finished long-form videos and posts to Instagram and Facebook | Executive Producer for review |
| Local voice and transcription box | Narration with an open-source voice model and transcription with Whisper, on a repurposed laptop at home | Production |
| Scheduler (Metricool) | Queues Shorts to TikTok and Instagram | The owner signs off on the queue |
| Principal agent | Decides which channels get more investment and which get cut, based on the producer's numbers | The owner |
Why a home server is in the stack
Voice generation and transcription run all week. Paying per minute for them adds up fast, and they don't need a frontier model. A spare workstation laptop with 64 GB of memory and a workstation GPU runs Ubuntu, Ollama, a Whisper transcription model and an open-source text-to-speech model. It handles the bulk work for the cost of electricity. The cloud tools are kept for what they do better: finished video and the final polish.
The quality system
AI content usually fails by being average, not bad, and average gets no views. Three rules keep the bar up:
- Fewer, better. This month the studio cut its Shorts volume on purpose to raise quality per post.
- A reference standard. Every Short is judged against one reference reel: photoreal scenes, long held shots, slow camera moves. If it doesn't match, it doesn't ship.
- Corrections become rules. When the owner corrects an agent, the correction is written into that agent's brief so the same note is never given twice.
What broke this week
Every new tool needs a test week before it's trusted. This week the long-form tool started a one-week pilot producing and posting one channel's weekly episode without hand edits. The review is on October 30, and the result goes in build log #002, whichever way it goes.
What this means for a media company. The same pattern of one job per agent, a written brief, a reviewer and a human approval gate is what we install in production and ad ops teams. Want to know what it's worth for one of your workflows? Run the ROI calculator.