How we run two faceless YouTube channels with an AI video factory

The short version

We run two faceless YouTube channels, Rise & Fall Files and DAINER AI, from one pipeline. From 12 to 14 Oct 2026, our uploader script put 46 videos on YouTube through the official API, each with a verified publish schedule. Three test uploads are not counted.

On 14 Oct Rise & Fall Files had 11 of the 1,000 subscribers and 52.5 of the 4,000 watch hours that YouTube's Partner Program asks for, so it can't earn ad revenue yet.

This guide covers the stages, what we measured, what broke, and the rule to read before you build anything.

The six stages

A faceless channel means nobody appears on camera. The video is voice, visuals and text. A factory means the steps from idea to scheduled upload repeat the same way every time.

1. Script. Claude (Opus) writes the script from research notes. A person picks the topic and approves the angle. Our production rule needs fresh research for each film.

2. Voice. Cartesia text-to-speech reads the final script. A Rise & Fall film uses about 10,250 Cartesia credits. The Shorts reuse narration from the long video, but a newly voiced outro uses extra credits.

3. Scenes. Smaller Claude helpers (Sonnet) build the scenes in HyperFrames, an HTML-based video framework, in a fixed chalkboard style. Opus checks every render before it moves on.

4. Render. When several renders are queued, we send the video-rendering stage to a Linux machine running Docker. Audio finishing and review stay on the Mac. In one test, a Short rendered in 108 seconds on Linux, running 4 render jobs at once. On the Mac it had taken about 10 minutes with 1 job at background priority. We compared both outputs frame by frame (1,621 frames each). They were almost identical, with a small visible font difference.

5. Review and thumbnails. A read-only Codex run reviews each release and checks the script's facts. Thumbnail options are made in ChatGPT, and a person picks one. We aim for three thumbnail variants per long video, though review sometimes leaves two.

6. Upload and follow-up. A Python script uploads through the YouTube Data API: video, thumbnail, captions, playlist and a scheduled publish time. It then checks with YouTube what was actually scheduled. After that come the pinned comment and each Short's link to its long video.

What broke: uploading through a browser

Our first uploads used an AI agent driving a web browser, clicking through YouTube Studio like a person. One upload used about 16.3 million Codex tokens. On 2 Oct the browser session died halfway through a batch.

So we turned the steps the agent had worked out into a script. yt_upload.py reads a list of what to upload, uploads through the API, and checks the result. It passed 15 local tests and made private test uploads on both channels on 12 Oct. It uses no AI tokens.

The lesson we kept: let an agent work out a job once, then write the steps down as code.

What one episode used

For DAINER AI episode L03 (9 to 12 Oct), our ledger estimated US$66.45 of Claude usage at API prices. Some of the output was estimated, and Codex and Cartesia weren't priced, so this is neither the episode's total cost nor our subscription bill. The ledger also logged 3.37 hours of summed agent activity. That's machine time, not human labour or calendar time.

Codex logged 21.0 million tokens for the episode, and 17.04 million of them were in the upload thread. That's why the upload was the first step we automated properly.

The YouTube rules to read first

Inauthentic content. On 15 July 2025 YouTube renamed its "repetitious content" policy to "inauthentic content" and said this content had always been ineligible for monetisation (YouTube Help). It covers:

• videos that are interchangeable or follow one template with little variation
• slideshows or scrolling text with little narrative
• mass-produced content without the creator's own view

Reused content is judged across the whole channel, and the original creator's permission does not make it safe. A shared intro and outro is fine when the main content differs, and so is a series where each video has its own story.

Disclosure. YouTube requires a label on realistic synthetic or meaningfully altered content. Its examples exempt cloning your own voice for narration (YouTube Help). Check the audio as well as the visuals before deciding.

Monetisation thresholds. For ads revenue sharing today, you need 1,000 subscribers, plus either 4,000 qualified watch hours in the last 12 months or 10 million qualified public Shorts views in the last 90 days. Meeting that lets you apply for review (YouTube Help). From 1 Feb 2027, YouTube says new applicants will need 8,000 watch hours in 365 days or 20 million Shorts views in 90 days (YouTube blog, 10 Aug 2026).

API privacy. The current upload reference says unverified API projects are not restricted to private viewing (YouTube Data API). For your first test, explicitly request private and check the status that comes back.

Our numbers

For the week of 4 to 10 Oct:

| Channel | Views | Watch hours | Click-through | Net subscribers | |---|---|---|---|---| | DAINER AI | 439 | 0.3 | 1.5% | 0 | | Rise & Fall Files | 418 | 7.0 | 1.0% | +1 |

In the lifetime retention readings from 12 Oct, Sega had 54% of viewers still watching at about 30 seconds. Average percentage viewed was 22.9% for Sega, 28.6% for Pan Am, 28.1% for Walkman and 23.9% for Kodak. Blockbuster didn't have enough data yet.

Sega also had higher click-through than the other films. That doesn't tell us whether the thumbnail caused it. We still need completed thumbnail comparisons to judge that.

How to start your own

1. Pick one format you can repeat with a different story each time, not one template with the words swapped.
2. Write one script with Claude, with sources. Ask a second model to flag unsupported claims, then open the sources yourself before approving.
3. Make one video by hand with tools you know, and time each step.
4. Upload it manually as private and watch it on your phone.
5. Choose a repetitive step worth scripting. We chose uploads because browser automation used so many Codex tokens.
6. Read the retention graph of your first videos before you decide to make more of the same.

I would test whether people watch the first video before building the rest of the pipeline.

Related guides

• HyperFrames: make videos from an HTML file, the framework our scenes are built in.
• How to run a social media account with AI, for where the videos get shared.
• A 30-day plan to start an online business with AI agents, for the wider build.