36 Channels, One System: Faceless Video Production with AI Agents

Today on Sonraki Adım with Umut Ilhan, the guest is Ertan Dönmez. Ertan is a producer with over a million followers under the name ErtanLabs. The topic is striking: running thirty-six YouTube channels with AI agents and earning a living from them. I listened to this claim as the summary of a working system, not as a fairy tale.

What thirty-six channels means

First, the frame. Nobody steps in front of a camera on these channels. Narration, visuals, and editing come from AI models, while a human sits in the director chair. Put simply: there is a kitchen no customer enters, yet dishes leave it on schedule. And the head chef is the same person.

Ertan shares a telling detail during the talk. While he speaks, finished pieces land on a remote server behind him and enter automatic scheduling. Production never pauses, not even mid-conversation. I compare it to a bakery: while you chat with customers up front, bread keeps baking in the back.

Money: costs, hit rate, language choice

Now the money side. Ertan says he pays roughly $700 to $800 a month for tools. He keeps paying in months with no earnings at all. I read this as workshop rent: a fixed monthly price for keeping the operation alive. Token and model usage fees sit inside that rent.

The hit rate is stated openly: twenty to thirty percent of channels take off. Meanwhile a different story circulates in the market: out of five hundred attempts two channels survive, those two go in the showcase, and the four hundred ninety-eight that flopped go unmentioned. Ertan sides with transparency. I agree: I distrust any narrator who hides failed attempts.

Language choice follows the money too. Turkish content pays little per view, so production switched to English. Same effort, different cash register. Turkish trials for passive income remain possible, but Ertan calls them a waste of time. I find nothing guilty in this math: with pricey tools, the register must be large.

One field stays deliberately untouched in Turkish: topics carrying claims about the body. The reason is ethical. Older viewers may believe whatever they see, AI systems can err, and the moral weight is heavy. I find this decision significant. Drawing a line at topics needing a physician is rare maturity in an age of fast production.

The rule changed: inauthentic output gets cut

Now the rules of the game. Google updated its monetization terms on July 15, 2025. The old repetitious-content heading was rewritten as inauthentic content. Template-duplicated videos that copy each other fell outside earning eligibility. The rule says: production may be fast, but the work must be original.

The Verge summarized the story in those days. YouTube executive Rene Ritchie described the change as a minor update. Reaction and commentary pieces were not the target; mass-produced runs that look like junk to viewers were. I stress this detail because it matters: using AI was not banned, careless duplication was.

A common trap sits here: screenshotting a channel, handing it to a Claude model, and asking for the same channel. Ertan says this does not work. When everyone runs the same method, most channels earn nothing. My metaphor is the photocopy: the original carries value, the pile of copies does not. Knowing what to chase is part of the craft, as much as writing prompts.

The suggested path is narrow and clear: focus on one or two ideas, never jump at every topic. As model power grows, attention scatters and each new idea sparks a fresh craving. But scattered power delivers nothing. I compare it to gardening: whoever waters three trees eats fruit; whoever gives one glass to thirty seedlings stays thirsty.

Voice: cloning your own voice

Voice is the heart of the operation. With ElevenLabs, a voice clone comes from a short recording. Instant cloning needs just one or two minutes of clean audio. I compare it to a photo of a voice: one frame identifies a face, one short recording identifies a voice. Fast results first, quality step by step after.

A higher-quality path exists too: professional cloning. It asks for thirty to one hundred eighty minutes of recordings and takes a few hours to prepare. The result sounds close to broadcast. Cost and effort run high, but producers working in their own voice hear the difference. Token use climbs at this stage and lands in the expense column.

One step further: carrying your own voice into other languages. A voice recorded in English may keep a light accent when speaking Spanish. The talk cites Thrace and Aegean dialects, plus the Irish accent in English. My summary line: voice is a passport, accent the hometown stamp. The model carries the stamp along.

Visuals: avatars, consistency, thumbnails

On the visual side HeyGen leads. It offers over a thousand ready presenters, support for more than one hundred seventy languages, and fast script-to-video turnaround. A ninety-second script becomes video in about two minutes. I see it as a standing cast: roles assigned fast, stage set with ease.

When ready presenters fall short, a custom twin steps in. One short recording builds a digital double, face and expressions preserved. The critical issue here is consistency . The look must stay fixed; clothing and stage language must not drift. Think of a series actor: viewers want the same face every episode.

The script side starts with rival analysis. First measure what viewers love, then write the narration. Agent support speeds the draft; the final call stays human. Titles and flow pass human approval too. My keeper line: the system proposes, the human decides.

Thumbnail choice stays human as well. A person picks the frame that earns the click; the production line arranges the rest. My metaphor is the shop window: whoever knows the customer dresses it, while the warehouse runs on rails. With a wrong window, even the best product stays on the shelf.

The editing line: Claude, Remotion, HyperFrames

At the center of the editing line sits Claude. Think of Claude Code as an agent living in the terminal. It reads files, runs commands, and pulls jobs through GitHub. Its builder Anthropic designed the tool for long jobs. My metaphor is the apprentice: the master describes, the apprentice carries, cuts, and arranges.

Remotion is the path of building video from software parts. Each scene is a component; when the script changes, the scene changes. The project lives in a public GitHub repository, open for review and contribution. I compare it to a recipe book: with the recipe written down, every dish lands at the same texture.

HyperFrames steps in for jobs like map animations. Coordinate and style rules produce smooth transitions. Creators shipped hits with January-era technology, and newer models pushed power further. I read this trend plainly: tools refresh every month, and whoever builds the system early moves first.

Time, computing power, and the last word

Production time and computing power complete the picture. Rented computing strength handles editing and rendering, with remote servers joining heavy scenes. Temporary storage needs and the current card shortage get a mention too. Much of the $700 to $800 monthly bill goes here. I read this line as an electricity bill: as the workshop grows, the meter spins faster.

My last word. This is not a shortcut but a trade setup. One idea, your own voice, human approval, steady measurement: with four legs firm, the table stands. That a producer past forty chases young developers to learn only adds value. Curiosity keeps working while tools keep changing.