To automate a YouTube clips channel with AI, learn the manual process first, then write each step down as rules an agent can follow: verify the channel, collect source videos, transcribe them, label the content type, pick clips from the transcript, check the boundaries, cut and pace, write the SEO, schedule. My system published three clips in 35 minutes.
This post expands a 49-minute masterclass I published on October 11, 2026, where I run the whole process by hand in Premiere and Photoshop, then show the automated version. It's embedded below.
- Medium-form clips nurture buyers better than shorts. In a clip from his course, Alex Hormozi says his highlights channel generated twice the revenue of his main channel with roughly half the views.
- The automated run took 35 minutes. I gave Codex three source videos at 11:15 and it published three 9 to 10 minute clips to a test channel, scheduled for 11am Eastern.
- Volume is capped by source material, not time. The slow part is transcribing and finding clips, so six clips would have taken the same 35 minutes as three.
- You can't automate what you can't articulate. It took me about 100 hours to write the system, and it only worked because I had done every step by hand for years.
- Two jobs still want a human: verifying the channel once, and A/B testing titles and thumbnails, which YouTube's API doesn't support.
Watch it on YouTube
What is a medium-form clips channel, and why not just post shorts?
A medium-form clips channel (also called a highlights channel) posts 3 to 25 minute clips cut from your long-form content: podcasts, webinars, interviews, coaching calls. It sits between the full episode and the 30 to 60 second short.
Shorts are good for discoverability. They are bad at making a viewer stick around long enough to trust you. A 10-minute clip that solves one specific problem does both, which is why the biggest podcasts run one. Joe Rogan and Lex Fridman have clips channels pulling millions of views a month.
For business content the case is even stronger, because the viewer is qualified. Hormozi's highlights videos are problem-specific and tactical, so the person watching is closer to buying. That's how half the views can produce double the revenue.
Can you fully automate a clips channel with AI?
Yes. Most people talking about automating a clips channel on Twitter have only done it in theory, so I built it. My system takes source videos and goes all the way to a scheduled upload with no human in the loop. You're only limited by how much source material you have and how many tokens you can spend in a day.
In the test run I gave Codex a lesson from Warren Buffett, Steve Jobs's lost interview, and a My First Million episode with Rory Sutherland. 35 minutes later all three clips were live on a dummy channel. For comparison, Hormozi's team says one person, heavily assisted by AI, gets 15+ videos out per day. With the right workflows that number could be one or two videos an hour.
Two caveats. Channel setup is a one-time human job (more below). And YouTube's API can't run A/B title or thumbnail tests, so if you want those (you do), someone clicks through YouTube Studio for a couple of minutes per video.
What are the steps to automate a YouTube clips channel?
There are seven core steps, then publishing. Here is each one done by hand and done by the system.
| Step | By hand | With the AI system |
|---|---|---|
| 1. Channel onboarding | Create and verify the channel (phone + ID) | An agent walks you through it and connects everything. Still needs you once. |
| 2. Source intake | Find your long-form videos, download them to a Drive or Frame.io folder | Pulls new recordings automatically (yt-dlp for public videos) |
| 3. Transcription | Premiere's auto-transcribe (or DaVinci, CapCut) | MLX Whisper Turbo running locally on a Mac |
| 4. Content type | Intuitive for a human editor | Declared per folder: Q&A, podcast, panel, interview |
| 5. Clip identification | Watch the source at 2x, mark clips in the timeline | Agent reads the full transcript with a criteria file in one pass |
| 6. Quality check | Re-watch the start and end of each clip | Reads the first and last 30 seconds of each candidate's transcript |
| 7. Extraction and pacing | Copy to a new sequence, remove pauses and filler words in Premiere's text panel | FFmpeg cuts, removes blank space, adds the CTA |
After that: export the clip's transcript, generate the title, description and tags (I use my Optimize YouTube SEO skill), create the Notion card, make the thumbnail, and schedule. I mapped every step with timings at the masterclass process map.
Two details matter more than they look. First, verify the channel. YouTube only unlocks clickable links in descriptions, chapters and A/B testing once you complete advanced features verification, and custom thumbnails need phone verification. Second, declare the content type. A coaching exchange runs 3 to 10 minutes. A podcast through-line can run 8 to 25. The agent shouldn't guess which one it's looking at, because the expected clip length depends on it.
How do you find a good clip in a long video?
Look for a through-line: one subject that starts somewhere, builds, and pays off. In the masterclass I scrub through the Steve Jobs interview at 2x speed (press L twice in Premiere) until he starts the story of cold-calling Bill Hewlett at 12 to ask for spare parts. That's a clip: a clear opening, a story, and a payoff (he got the parts and a summer job at HP).
By hand, I mark it with a keybind that colors the clip in the timeline and a named marker, so I know where it is when I come back. The system does the same thing by reading the SRT file against a written criteria file, in a single pass over the whole transcript. It doesn't chunk the video into pieces. Chunking saves tokens, but it cuts through-lines in half.
That criteria file is the meat of the whole system. The more specific your description of a good clip, the better the clips.
Where should the CTA go in a clips video?
Put one at the end, and on clips longer than 7 or 8 minutes, add a second one about 30% in. More people are still watching at 30% than at 100%, so more of them see it.
Place the mid-roll right before a cliffhanger. If the speaker says "so here's the whole secret," cut to your 10 or 15 second CTA before he says it, then come back. It's the same trick every commercial break on television uses: people sit through it because they want what comes next.
How long does a clip take by hand versus with AI?
Walking one clip through every step by hand, with a thumbnail and a scheduled upload, took me a little over an hour in the masterclass, and that was working fast. Every extra source video adds more time, because someone has to sit through it.
The automated run took 35 minutes for three clips, and nobody had to watch it. Since the slowest part (transcription and clip identification) runs once per source video, adding clips from the same source barely adds time.
Who this is for (and who it isn't for)
It's for businesses and creators who already record long-form content (a podcast, coaching calls, webinars, interviews) and want it working harder. It's also for media teams who want to see each step written down so they can streamline it, and for editors who sell clipping as a service.
It isn't for anyone without source material. The system multiplies content you already have. If you don't record anything long-form yet, start there.
Common mistakes
- Automating before you can articulate. If you can't describe a step clearly enough to hand it to a person, an AI agent will do it badly too.
- Letting files live everywhere. Export to the same central folder every time. An agent that has to keep searching for new folders wastes tokens and makes mistakes.
- No naming convention. Optional if you post by hand, essential if an agent is involved.
- No used-clips log. Without a record of what's been cut from each source (for example 3:48 to 5:35, the HP story), a second editor or a second run will post the same clip twice.
- Skipping human review early on. Clip selection is still the hardest part for agents to fully get. Review the clips until you trust the criteria file.
Frequently asked questions
How many videos per day can an AI clips channel post?
As many as your source material supports. Hormozi's team posts 15+ per day with one person and AI; a fully automated system could do one or two an hour if the footage exists.
What tools do you need to automate a clips channel?
For the manual version, Premiere Pro and Photoshop. For the automated version, yt-dlp, MLX Whisper, FFmpeg, Claude or Codex, and Notion for the content calendar.
Do highlights channels need thumbnails?
No. Hormozi ran his highlights channel without thumbnails for a long time. Simple thumbnails help, and ChatGPT can now make one from a reference thumbnail and a frame in a single prompt.
Can the YouTube API run A/B title and thumbnail tests?
No. The API can upload and schedule, but title and thumbnail tests still have to be set up in YouTube Studio by hand.
Related reading
- You can't automate what you can't articulate, the principle this whole system is built on
- Can AI edit videos like a human editor?
- How to schedule social media posts with Claude
- Why API access matters for AI agents
- The full masterclass process map
- Tools: yt-dlp and MLX Whisper
Want the system itself? I sell it three ways: DIY (every workflow, skill, prompt and agent), done with you (I implement it and train your team), or done for you (my agency runs the channel). Book a call and we'll figure out which fits.