A long time ago, I helped out with a few YouTube channels and even tried launching my own.
But it always came down to the same thing: either it took way too much time, or, in the case of my own channel, it also cost money I didn't want to spend. Hire an editor, record a voiceover, and then somehow try to promote it all. It was interesting, but my addiction to cheap dopamine and nice-looking numbers kind of pushed back when I saw my videos getting almost no views. Even though they seemed to be decent quality. So the whole thing went on the back burner.
Then recently a thought hit me: "What if all the boring work could be automated and made easier?" That's basically how I came up with the idea of building a pipeline for making YouTube videos.
Of course, I first tried to find something similar online. But all I found were pipelines for churning out AI slop. That's not our way, because it looks pretty meh. And the videos it makes are so cringe that they're embarrassing to watch. Especially in the history niche, which is the one I decided to try.
So we choose the way of the samurai and go write our own pipeline. But first we need a stack, and the pick was Remotion, as the simplest thing for my Claude to work with. At first there was some trial and error with a few video types and figuring out the content. No editing "bibles," no working rules at all. Some videos came out, even decent ones. But it wasn't ready to scale, because I constantly had to explain to Claude what I wanted from it.
So step by step I started tuning the tools, wrote an editing "bible" (partly with Claude, after analyzing competitor channels and their approaches), plus a channel concept and what would actually be on it. And then I got to the technical part.
Experimenting bit by bit, I ended up with a modular system where agents work together with Claude.
1. Concept module. This is where script concepts get written and ideas get brainstormed. I throw them at Claude, it digs up similar stuff on its own and fleshes out the concepts. Then they go to Jev for a rating based on the viewer audiences I've written up, and I get back a ranking of how likely each one is to go viral. All I have to do is pick one and move on to the next stage.
2. Research module. What story can be told without evidence to back it up? Exactly, none at all. So this module combs through public domain internet archives, academic archives, and all kinds of articles, and builds me a document with sources. If it's a book, it converts it to .md and points out where the info I need is. So the output is a complete source document that's easy to work with.
3. Footage module. Honestly, you could call this a continuation of the research module, but nope. Here the agents comb YouTube for similar videos, break them down into contact sheets, and look at what footage was used. In parallel, they search for footage: archival video, photos, documents, and similar material. Some of it was already found during the research stage, so here it just gets a note on what can be used.
That's why this is where tens of gigabytes of video get downloaded, cut into clips, and their indexes written to a file for later use.
4. Script module. Pretty self-explanatory. This is where we write the script. I write the core, then edit it a bit with Claude. Not much to talk about here, it's the simplest and most boring part (technically speaking).
5. Editing module. This is where it gets more interesting. Editing myself is inconvenient right now, because I work on an office laptop with crappy integrated graphics. So the editing is left to the AI.
At this stage, the maps used for the visuals get drawn. There's a neat render-optimization trick that "bakes" only the part of the map that will actually be used. Because the map itself is made of layers: elevation, shading, water, and place names, which are styled with code to match the look I need.
So first the maps get drawn, rendered, and dropped into a separate folder, where Claude and I check them for accuracy. Then comes the plain stitching of the video in parts, and it's rendered in parts. Why in parts? Because the laptop takes 8–9 hours to render a video on integrated graphics, and with power outages things can get awkward. After the render, the voice and music get layered on. For the voice, I went with ElevenLabs and cloned my own.
Once the render is done, I also make 3–4 Shorts from the most interesting moments. That takes another 1–2 hours, purely because of render time.
6. Prep module. Thumbnails and subtitles get made here. Also a routine part, but a pretty important one. The subtitles get translated into 10 languages, after running the video through Whisper first to get the right timings.
And that's it. The video goes up on YouTube, and we wait for the nice numbers.
You can check out the channel here: Switchyard
Top comments (0)