How I automated my YouTube channel from one desk
# How I automated my YouTube channel from one desk
The video this article pairs with was assembled by a machine on the desk in front of me. I wrote the script and sat at the checkpoints that needed judgment. Everything else, the voicing, the footage selection, the captions, the render, happened without me touching it.
I did not build this because I love automation for its own sake. I built it because consistency is where side channels die. One good video is a fun weekend project. One video every week, for months, around a job and the rest of a life, is the point where almost everyone quits. Ideas were never my bottleneck, and neither was editing. The grind was.
What automated actually means
First, the fantasy has to go. Nobody presses a button at night and wakes up to a finished video and a pile of money. Anything promising that is selling something.
What runs here is a pipeline. Small steps, each doing one job well, chained together, with a human sitting at a few checkpoints. The machine carries the slow, repetitive work. I keep the decisions that shape the channel.
Handing over any of it took longer than I expected. When you care about the output, the instinct is to touch every stage yourself. What finally moved me was looking honestly at which stages I was guarding. Rendering. Caption timing. File wrangling. None of that was creative. It was tedium dressed up as craft. Once I separated work that needs a human from work that just needs doing, the design of the whole system fell out naturally. I never wanted to remove myself from my own channel. I wanted to remove the busywork, so a busy week no longer means a quiet month.
The script is the checkpoint that matters
A faceless video is narration with the right pictures laid on top, so everything downstream inherits the script's quality. I keep a running topic list. It fills from search queries people actually type, from questions I see asked again and again, and from things I worked out myself and wished someone had explained without the mystique. Each script lands around two thousand words, which reads out to roughly twelve minutes.
An AI model helps me draft and tighten. Every line still gets rewritten by hand. A generic script cannot be rescued later by clever assembly, no matter how good the tooling is. The model is genuinely strong on structure and genuinely weak on having a point of view, so the point of view is my job.
My filter is blunt. If I would not say a sentence out loud to someone across a table, it gets cut. Left alone, the model produces smooth, empty filler by the page, and most of my editing time goes to finding that filler and deleting it.
A clone of my voice reads it
Once a script is locked, a text to speech service called fish.audio narrates it with a cloned copy of my voice. The reference came from a few minutes of me reading aloud, recorded once. From then on the script file is the exact input, word for word, so what appears on screen matches what gets spoken.
The clone has flaws. Stress sometimes lands on the wrong word. Long sentences come out slightly relentless, because the pacing does not breathe the way a person does. Few viewers ever notice. In exchange, a twelve minute voiceover takes under a minute instead of an hour of takes and retakes.
One point outranks all the technical detail. This is my own voice, cloned with my own consent, used on my own channel. Aiming the same tool at somebody else's voice without permission is impersonation, and I want no part of it. The software has no opinion about whose voice it copies. The person at the keyboard carries all of the responsibility.
Visuals are where the real work went
A voice over a black screen is a podcast. The stage that decides what sits on screen took more experimentation than everything else combined.
For each paragraph of narration, the system chooses a treatment. Sometimes that is stock footage, pulled from a library by searching for the phrase being spoken. Sometimes it is a diagram generated from the content of the section itself: a comparison drawn as a side by side, a process drawn as a flow with each step appearing in turn. Captions run underneath throughout, taken straight from the script text, so spelling holds up even on odd technical terms.
Timing carries the whole effect. Each visual anchors to a specific phrase, so the screen changes at the moment the narration reaches the idea it illustrates. A paragraph left as plain text on a frozen screen is the fastest way to make a video look cheap, and the pipeline treats that as a hard failure rather than a style choice.
Rhythm took the longest to tune. Too few changes and the video drags. Too many and the screen fights the viewer for attention. A fresh visual every few seconds turned out right, with diagrams held longer so there is actually time to read them.
What it costs
Pocket change is the honest answer, at least in cash.
The drafting model runs a few dollars a month at my writing volume. The metered voice service works out well under a dollar per full narration. Stock footage comes from free tiers I have yet to outgrow, and rendering adds nothing because it runs on hardware I already own.
The real bill was never monthly. Building the pipeline consumed serious time, and the checkpoints still take attention on every single video. The tools have never been cheaper than right now. Knowing what is worth making in the first place is the part that stays expensive.
Rendering happens on a card from 2017
Most people would push the render step to a cloud service and pay per job. Mine runs locally on a 1080 Ti, an old graphics card by any measure, using its built in video encoder to stitch the footage, the diagrams, the captions and the voice into one file. A twelve minute video takes roughly six to ten minutes.
Local rendering means the marginal cost sits at zero, the pipeline works offline, and nothing about the process leaves the house. The tradeoffs are real. The card is slower than fresh cloud hardware, and when a render falls over at two in the morning there is no support line. I decided control and a zero running cost mattered more to me than speed. Someone else could run the same numbers and land on the opposite answer, fairly.
There is a quieter benefit too. A failure inside somebody else's cloud is a support ticket. A failure on the machine under my desk is something I can open up and understand, and that understanding compounds across every later problem.
Two touches that hide the seams
Every video opens on a title card, and a trim step cuts the opening so playback begins on the first spoken word rather than a slide hanging in silence. Nobody consciously notices this. Everyone feels it.
The upload stage writes the title, the description and the tags, then hands the file toward the channel. For this channel that last stretch is still being wired properly, so I check each file myself before anything ships. I would rather publish slowly than wake up to the wrong video sitting on the wrong channel.
Where it still breaks
Often, is the truthful answer. The last twenty percent of any automation eats most of the time.
The voice mispronounces a name until the spelling gets nudged. A stock clip comes back wildly off topic for its phrase. A diagram renders too dense to read in the seconds it holds the screen. Early on, a visual failed silently and a blank green slide played for twenty full seconds, caught only because I watched the whole render back. Another time every caption doubled on itself because the pipeline and the voice disagreed about where one word ended and the next began.
Each of those bugs stayed invisible until I sat through the output, which is the slowest possible way to find anything. So every failure bought a check. The system now inspects its own renders, samples individual frames, and flags what looks wrong before anything goes live. That loop turned a fragile toy into something I trust. A pipeline earns trust by failing loudly and early, so a break surfaces in seconds rather than after a thousand views.
Should you build one
For most people, no. Someone making one video a month who enjoys editing has no problem this solves, and the build only repays itself across dozens of videos.
The answer flips when the editing grind is what keeps breaking your consistency. Moving the repetitive work onto a system is what lets a channel survive real life. Mine did not make me more talented. It made showing up cheap, and showing up turns out to be most of the fight. The machine absorbed the work I was going to skip anyway, which is exactly what freed the work that matters.
More breakdowns of this kind, with the failures left in, live at [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.