A long-form video normally takes a crew: research, script, art, voice, edit and packaging. I wanted to see how much of that one person and a set of AI tools could carry, without losing a consistent look.
The result is a series engine. It works for any channel. The pilot is a finance channel, Mr Market Mood.
Three ideas hold it together.
- One control panel. One settings file picks the series, the models and the voice. A new channel means a new settings file, not a new pipeline.
- A visual constitution. The mascot is locked. The visual rules live in a machine-readable file. One prompt-builder assembles every image prompt, so guards fire by tag and colour carries meaning. The look stays the same across hundreds of frames.
- Gates before spend. A topic must pass a demand check and a hook gate before any image or voice is generated.
The production line:
- Topic, from my content research engine
- Hook gate and retention check
- Script, then beats
- Still images, generated with Gemini
- Voice, with ElevenLabs
- Optional b-roll, with fal.ai
- Music
- Render, in Remotion
- Title and thumbnail scoring
- Upload
Why long-form: a video of 8 minutes or more can carry mid-roll ad slots. That is the main revenue lever.
The growth rules are simple. Study outliers before choosing a topic. Commit only when demand beats supply. Make the title and thumbnail ask one connected question.