A production studio run from the terminal: research, scripts, voice, rendering and packaging for five channels.
The problem
I wanted to produce cinematic channels alone, at a cadence teams normally need — idea to published video, with data deciding what gets made next.
The approach
The studio runs on Claude Code in the terminal, wired to custom MCP servers. Each channel has its own voice, visual engine and publishing cadence; the pipeline below carries an idea from data to published video, and every tool in it earns its place.
The result
The largest channel: 44 videos, 570,000+ views, 17,400 subscribers, built from zero by one person directing a system. The same pipeline now runs four channels, each with its own voice, visual engine and publishing cadence.
The channels
- The Ancient King ↗Dark philosophy. 570K+ views, 17.4K subscribers, 44 videos.
- The Threshold ↗Mystical-Christianity lectures over public-domain art and coded diagrams. In production.
- The Relocation File ↗Where Americans move and what it actually costs — every claim receipt-checked.
- OFF THE METER ↗The grid, the bill, and the way off both.
The pipeline
01 — Research
Custom MCP servers query a YouTube analytics database for outlier scores, niche supply and RPM before anything is greenlit.
Claude CodeNexLev MCPGrok (xAI)research agents
02 — Packaging
Title and thumbnail are chosen from measured outliers and locked before production starts, then filtered through the channel's brand.
outlier methodA/B titlesengine-rendered thumbnails
03 — Script
Written in the channel's documented voice against a versioned research vault of primary sources.
Claude CodeObsidian vaultvoice bibles
04 — Voice & assets
Voiceover from a local TTS rig with a tuned recipe per channel; art direction generated per channel's visual language.
qwen-TTS (local)muapi / FluxMidjourney
05 — Render
Real-time three.js worlds rendered to video and composited by scripted build pipelines — heavy renders fan out to cloud GPUs, one pod per section.
three.jsffmpegRunPod GPU fleet
06 — Measure
Hook rate, CTR and retention flow back into the research layer and decide the next greenlight.
YouTube AnalyticsNexLev outliers
The division of labour
What the AI does
- Pulls outlier and niche data through custom MCP servers before a video is greenlit
- Drafts research from a vault of primary sources, in each channel's documented voice
- Generates voiceover from tuned per-channel recipes and renders three.js worlds to video on cloud GPUs
What I decide
- What gets greenlit — packaging is chosen from measured outliers, then judged by eye against the brand
- Every word of every script
- The visual engines themselves: each channel's world is designed and coded, then reused as a system
What went wrong
I tried AI-generated thumbnails. They looked like AI, performed like it, and got rejected — the channel's thumbnails are now rendered from real scene engines. Taste is a filter every asset passes through.