How a Short got made
Each Short started as a row in a table where I wrote the topic. When I started a batch, three agents on OpenAI's Assistants API did the writing and the visual decisions. The Scriptwriter wrote up to 150 words, with a hook at the start and an ending that leads back into it, so the Short plays as a loop. OpenAI's text-to-speech read the script, and when the narration ran past a minute, the script went back to the Scriptwriter to be cut. Whisper transcribed the narration with timings, the Architect chose the moments that needed a new image and gave each one its colour and camera move, and the Photographer wrote the prompts for those images. LemonFox generated the images, with DALL·E 3 as the fallback.
Putting each video together was plain code. Creatomate rendered the first ones, and from 30 December Remotion did, with a slow zoom or pan on every image, captions that appear word by word, and music chosen by the mood the Scriptwriter gave the script. From 21 December Saga uploaded the videos through the YouTube API, and from January it worked on three Shorts in parallel.
My part was writing the topics, running each stage from a terminal menu and making each video public, since the API uploaded them as private. The writing, the voice, the images and the edit ran without me.
The run
Between 14 December 2023 and 16 January 2024 Saga published 205 Shorts, about six a day and 16 on the busiest, each costing under $1 to make. The channel reached 1,000 subscribers on 16 January. That settled the bet in my favour, and it was the channel's last day of uploads.
The first week brought 41 subscribers. Views first jumped on 20 December, to 5,005, and from Christmas Day the channel took off, gaining 788 subscribers between 25 December and 7 January. The best day was 6 January, with 11,450 views and 84 new subscribers.
The 34 days of uploads brought 135,845 views and 1,102 hours of watch time, 94% of the channel's lifetime views. On average, viewers watched 29 seconds of Shorts that ran about 49. I paid for no promotion, and over the channel's life 88% of its views have come from the Shorts feed and most of the rest from YouTube search.
After the bet
What I wanted next was longer stories, and the models of early 2024 couldn't hold one together, in the writing or in the images. In 2024 I rebuilt it as a web studio, where you plan a video in a chat with an agent, pick a voice, generate and animate the images and lay them out in a Remotion editor. A month later I tried a team of CrewAI agents writing four-minute scripts, with a researcher, a writer on Claude 3 Opus and a critic.
Saga's latest version is how I make Cybertantra, my channel about spirituality, with episodes up to fifty minutes long, narrated with Eleven v4, scored with Eleven Music and illustrated with OpenAI's image model.
See The Last Alchemist Watch Cybertantra The first version, TubeSleuth, on GitHub