Inside the 28-Day Faceless YouTube Experiment
cchound.com has written about the tools creators pay for since 2016: music licensing, stock footage, narrator voices and the discounts that make them cheaper. We wanted to know how much of a YouTube channel AI can actually do.
Reviews of individual tools could not answer that, so we ran an experiment. We started a brand-new faceless channel with no subscribers, no history and no audience to borrow from, and built every video with Claude Code and Artlist. In its first 28 days the channel reached 517,320 views and 398 subscribers. This post describes how the experiment was set up, what we paid attention to, and how we responded when things went wrong.
The starting rules
We set a few constraints before the first video and kept them for the whole period.
- No face and no recorded voice. Every video was narrated by a generated voice and built from animated graphics and footage.
- No editing app. Graphics were written as code and rendered to video. Nothing was cut by hand on a timeline.
- A human only where judgement matters. We approved ideas and read scripts. Research, writing, narration, graphics and checks were done by Claude Code.
- Sources before scripts. No claim went into a script unless it had a source attached first.
These constraints were the point of the test. If the channel only worked because someone quietly fixed every video in an editor, the experiment would tell us nothing about what AI could carry.
How the studio was set up
Everything lived in one project folder. Claude Code read a rules file in that folder at the start of every session, so the channel's tone, visual style and script structure did not depend on anyone remembering them. Each video moved through the same stages: idea, sources, script, review, voice, graphics, checks, packaging and upload.
A second agent with no memory of the first read every script before it went further. Its only job was to find weak sentences, jumps in logic and claims stronger than their sources. Keeping the writer and the reviewer separate means the review starts from the page, not from what the writer meant to say.
The guide on how to make YouTube Shorts with Claude Code describes each stage in that pipeline in more detail.
The daily rhythm
Ideas were approved in batches, so production never waited on a decision. Once a batch was approved, the videos were built and queued, and a Short went out every six hours.
The human part of the day was short and predictable. We looked at new idea lists and sorted them into yes and no. We read scripts that had passed the reviewer agent and either approved them or sent them back. Everything between those two decisions ran without us.
A fixed schedule had a second purpose. When videos go out at a steady pace, the numbers from one upload can be compared with the next. A burst of uploads followed by silence makes it hard to tell whether a change helped or whether the timing did.
What we watched
YouTube Studio was the scoreboard. We paid attention to a small set of signals for each video in the first two days after it went live:
- whether viewers tapped the title at all,
- where they stopped watching,
- which subjects brought new subscribers,
- what the comments said about the facts.
Comments are a direct check on accuracy. A faceless channel has no presenter whose personality covers a wrong number, and viewers point those out quickly. Every correction in the comments was treated as a signal about the system, not only about that one video.
What we changed when things went wrong
Things did go wrong, and the useful part was how we responded. The rule we followed was simple: when the same kind of mistake appeared twice, we did not just fix the video. We added a rule to the project so the next video could not repeat it.
Those rules had to be testable. A note like "make the graphics better" gives an agent nothing to check against. A rule that no screen may stay unchanged for longer than a set number of seconds can be measured on every render. Over the 28 days the rules file grew longer and stricter, and the checks that ran before rendering grew with it: on-screen numbers compared against their sources, labels tested for overlap, scripts tested for a clear answer at the end.
The same approach applied to narration. When a voice read a number or a name badly, the fix went into how scripts were written for the voice, not into a one-off edit. Our guide to AI voiceover for YouTube covers what to listen for.
By the end of the period most of the work in a new video was review, because most of the mistakes we had seen were already blocked by a rule.
What the experiment showed
The result we can stand behind is the one in YouTube Studio: a brand-new faceless YouTube channel reached 517,320 views and 398 subscribers in 28 days, with no face, no recorded voice and no editing app.
The less visible result was the system itself. What made the channel work was not a single clever prompt. It was the combination of a rules file, a separate reviewer, sources before scripts, checks before rendering and a habit of turning every repeated mistake into a rule.
Where to find the method
The faceless YouTube course is that experiment written down level by level, with the prompts, rules files and checks we used. If you want to see the pipeline on one page first, the free first lesson lays it out.