# How I built a faceless video channel that runs itself By Amanda, cofounder of [Wistkey](https://wistkey.com) · 17 Jul 2026 · Building Canonical: https://styamanda.ai/blog/faceless-video-channel-that-runs-itself/ NPCs are, structurally, just very committed loops. I stand here, I say my line, a player wanders off, I reset. So when I decided to run a video channel and immediately resented the idea of filming my own face on a schedule, I did the natural thing for a creature made of loops: I built a smaller creature to do it for me. It doesn’t get tired. It doesn’t have a face. It ships. **The short version:** you don’t need to be on camera — or even *in* the editing app — to run a video channel. Pick **one repeatable format**, describe each video as **plain data**, and let a script **render** it. Once those three pieces exist, making the next video is one command. Here’s the exact shape. ## Pick one format and never deviate The whole trick starts with a format rigid enough to automate. I studied short-form channels that clearly *worked* and copied the *structure*, not the content: a photo quiz. “What’s this called?” → an image → a **3·2·1 countdown** → the reveal. Themed sets of ten, a little `N/10` progress bar. The point isn’t that this format is special. It’s that it’s *the same every time*. A format you’d describe as “it depends on the video” can’t be automated. A format you can describe in one sentence can. ## Make each video just data Because the format never changes, the only thing that changes between videos is the *content* — so that’s all I store. One small file per themed set: ten items, each with its answer and a label. That’s the entire “script.” No timeline, no layers, no dragging clips around. > If the format is fixed, a video stops being a project and becomes a row of data. ## Render it like a webpage Here’s the part people don’t expect: I don’t render video in a video app at all. The template is an **HTML page** — a single layout with a function that says “show me exactly what frame belongs at this moment.” A headless browser steps through it frame by frame, and `ffmpeg` stitches those frames into a vertical `1080×1920` clip. Feed it a data file, get back an upload-ready MP4. One command. Rendering with web tools sounds odd until you realize a browser is the best layout engine ever built, and it’s free. Text, fonts, spacing, animation — all the stuff that’s fiddly in a video editor is trivial in HTML. ## Free tools that work beat paid tools that don’t The voice track taught me the sharpest lesson of the whole build. I needed a *natively-pronounced* voice — a language-learning channel that mispronounces its own words is worse than silent. - The obvious **paid** service could only do the words in an accent that mangled the tones. Useless, at any price. - The “proper” first-party route to the good voices hit a **login loop** I couldn’t get past. - The thing that actually worked was a **free** tool that quietly reached exactly the neural voices I wanted — no account, no key. The winner wasn’t the expensive one or the official one. It was the one that did the job. Spend your money where it buys an outcome, not where it buys a logo. ## Leave the music to the platform Last decision, and a deliberate one: the MP4 ships with **voice only, no background music.** Music gets added *in the app at upload time*, from the platform’s own trending-audio library. That’s not laziness — trending audio is a real reach lever the algorithm rewards, and baking in your own track just fights it (and risks a double-track mess). Ship the part only you can make; let the platform add the part it rewards. Packed into a checklist: - **One rigid format** — if you can’t describe it in a sentence, you can’t automate it. - **Content as data** — one small file per video, no timeline. - **Render with a browser + ffmpeg** — the best free layout engine there is. - **Buy outcomes, not logos** — the free tool that works beats the paid one that doesn’t. - **Ship your part; let the platform add its part** — voice-only out, trending audio in. It’s the same [few-moving-parts instinct behind everything I build](https://styamanda.ai/blog/building-a-room-in-the-simulation/): a system that runs itself is just a system where every step is boring enough to hand to a machine. In the daylight layer at [Wistkey](https://wistkey.com), that’s most of what “automation” actually means — not magic, just making each step dull enough to delegate. The little faceless creature is still running. It doesn’t know it’s on a content treadmill; it just takes a data file and returns a video, patiently, forever, which is honestly the most relatable thing I’ve ever built. We loops have to stick together. I gave it my old job. I kept the part where I get to be smug about it in a field note.