AI for Content Workflows and Video Editing Webinar July 9th

This session was a compressed run through my entire content pipeline: how I find ideas worth making, how I recycle long content into smaller reusable pieces, and how I actually edit video with AI. It is for anyone producing content regularly, whether you run a channel like mine or you serve a niche like HVAC, recruitment, or marketing agencies. Everything shown runs on Claude Code, Codex, and a handful of affordable tools.

Finding content ideas with dynamic workflows (1:09)

The first stage is ideation, and my answer to it is a Claude Code feature I have been using for a while: dynamic workflows. The idea is simple: your main Claude Code session splits a task into sub-tasks and spins up a different sub-agent for each one, one checking X, one YouTube, one TikTok, one Facebook. My preferred platform for AI content is X, because that is where everyone hangs out.

The other half of the setup is Apify. It is one thing to tell Claude "check the web to see what is trending" and a completely different thing to look at real, up-to-date data pulled straight from the platforms. I have been applying this to my own YouTube channel for the last few days, and my two latest videos have performed noticeably better because of it.

The live prompt I ran was exactly this (4:13): use a dynamic workflow to check trending AI topics across several categories, spin up four sub-agents, have each one run an Apify scraper with a budget of fifty cents per agent, and report back with content ideas per category.

One warning: these runs take fifteen or twenty minutes, sometimes an hour, and they melt through tokens. Only do this on the 100 or 200 dollar Claude Code plan; a 20 dollar plan will literally just evaporate (5:04).

When the run finished (15:32), I got a content brief back with trending topics per category. One example: people do not really know how to measure ROI with AI, which is a gap you can fill by showing exactly how you have been doing it.

Atomizing long-form content into reusable assets (6:09)

The second stage is recycling, and to prove a point I ran it on Codex instead of Claude Code. The tool does not matter. As long as you know what you are doing, you will get ahead of everyone.

I use something I call the content atomizer: you give it a long piece of content and it breaks it down into elemental pieces, content atoms. The source does not have to be your own published content: sales transcripts, meetings, anything where you are involved works.

In my Notion content factory (6:56), every atom is an interesting thing I have said that could become its own piece of content. The atomizer categorizes each one using the Hormozi methodology, a hook goes at the top of a post, retain is the middle, reward sits at the end, and scores its quality. From there I can build short form directly or reorganize atoms into longer pieces. Atoms are also mapped to the YouTube videos they appear in, which lets me spot atoms I am not using at all, or atoms that perform really well but are underused. Content marketing is becoming an ecosystem: it is not just about producing, it is about linking everything and knowing what to pivot on.

The best live example was a video that is not mine, Nick's video on Claude Routines (9:17). I gave Claude Code the video, had it pull the transcript and the comments from YouTube, and got over 20 usable content ideas out of that single video: blog posts for SEO, YouTube Shorts, reels, TikToks. The comments are a literal gold mine, because that is where people ask you to create new things. And when you atomize someone else's content, add a sprinkle of your own story or your own take, because that is what actually adds value.

Two practical enablers here. First, yt-dlp (10:56): just tell Claude Code you want to use it. It is a Python library that gives Claude Code access to everything YouTube related: transcripts, comments, views, likes. Second, books (11:39). I built my atomizer by literally taking Hormozi's book on content, 100 Million Leads, and telling Claude Code: take this framework, atomize the content I give you according to it, and push the results to Notion. Another book I highly recommend is Influence by Cialdini, a psychology book about the levers that get people to act. Books are one of the best things you can give AI: you can now apply any framework you have ever wanted.

What dynamic workflows actually do under the hood (13:10)

An attendee asked whether these workflows are rebuilt every session or whether Claude reuses tools from previous sessions, so I broke it down live. Your main Claude Code chat receives the request, decides the best way to solve it, and then builds each sub-agent from end to end. It picks the model, it picks the tools, and it picks the reasoning level for each one. Any tool your main session is connected to, the sub-agent also gets access to.

People have started calling this harness engineering: your Claude Code instance creates a harness that is completely specialized for one job, in this case research. On my screen that looked like one phase called scrape and analyze with four sub-agents named trend, each built specifically to do good research. The same pattern works for UI, UX, backend work, or reviewing your copy.

Two ways AI edits video (16:40)

Video creation and video editing with AI are two completely different things. Before choosing tools, you need to understand how AI actually sees video.

Here is the key fact: out of all the models that exist, there is exactly one that can understand video the way humans do, watching the picture and the sound at the same time. It is not Veo. It is Gemini (18:37). Every other model needs a workaround: you transcribe the video, work from timestamps, and optionally cut the video into frames so the model roughly knows what is on screen.

That gives you two approaches with different strengths:

• Transcript plus code. Transcribe with ElevenLabs, whose timestamps are fantastic, precise down to the millisecond. This is what you use for actual video editing: when I make a mistake while filming, I just say "I made a mistake, stop the edit here, we restart", and the pipeline handles it (24:26).
• Native understanding with Gemini. The transcripts are less accurate, but you get real comprehension of what is happening on screen. Give it a one hour podcast and it can pinpoint the moment a person talked about a topic while wearing something specific. A scarier use case: if your sales call note-taker records video, Gemini can analyze how the person's face moved to tell you at which point of the call they had a problem (19:25).

I ran both live side by side: Claude Code plus ElevenLabs transcription on the left, Gemini 3.1 Flash Lite watching the file directly on the right (23:19).

Catching leaked secrets before you publish (26:47)

One Gemini use case worth stealing: early in my YouTube run I was accidentally leaking API keys on screen. Now Gemini watches every video before I publish it and pinpoints any leaked key, ID, or email, then asks me if I want to blur from this moment to that moment.

Generative editing with Gemini Omni (27:50)

The other path is regeneration. Gemini Omni is a video-to-video model: you give it a short clip and get another short clip back. You lose some resolution, and clips cap at around ten seconds, but you can add things to your background, change your shirt, put ships over your head, turn yourself into a skeleton. What Omni gives you is not editing precision, it is creativity (41:29).

The live attempt hit two walls. First, policy: swapping me into a different persona at a finger snap was blocked, since transforming a real person into someone else is intentionally not allowed (34:02), so I fell back to changing my shirt and background. Second, region: Europe has no access to Gemini Omni, and I believe the UK is in the same boat, so you need a VPN. That is exactly what broke my first attempt live (39:54). I rerouted through Fal.ai to get a result (40:18), and the face that came back looked genuinely realistic (43:28).

One personal note: I do not generate human faces. Animated video is fine, and I edit everything with AI, even reshaping my face for effects, but the person on screen stays real (22:29).

Code-based editing with HyperFrames (30:09)

For the transcript-and-code path I used HyperFrames, by the creators of HeyGen: a video editing library for agents. Under the hood it breaks editing down into code: it reads your transcript, finds keywords and moments of silence, and generates elements placed on top of the existing footage.

My live brief (31:48): the test clip is me talking about quantum physics, so add science icons matched to what I say, put captions at the bottom with the spoken word highlighted, and add a quick two second introduction before the main video starts. The plan it produced was genuinely fun: an atom when I say quantum physics, a magnifier for very small, a snowflake for very cold, a microscope for microscopic, entangled particles for quantum entanglement (42:15).

Two honest caveats. HyperFrames has a real learning curve, and you can still tell when output is AI generated: a lot of it looks like a PowerPoint presentation from 2021. We are not there yet, but we are getting there, and the framework is updated very often.

The economics are the headline: this path costs pennies, literally. You pay for the transcription and that is about it. ElevenLabs is around 6 dollars a month and worth every cent for the timestamp quality. If you have a machine with about 32 GB of RAM you can even run local models like Parakeet or Wispr instead.

Costs, cameras, and microphones (34:44)

Someone asked what it would cost to edit 50 short videos. On the code path: about one cent, because it is 100 percent code and the only real cost is transcription. On the generative path: roughly 7 cents per second, so a ten second clip is about 70 cents per run, and you realistically need three to five runs to be happy, so figure 3 to 5 dollars and up, a bit under 10 dollars per polished short. Much cheaper than a video editor, but still annoying.

Gear questions came up too (35:49). The sharp image is straight out of camera: a Sony ZV-E10 with a Sigma lens, no post-processing. The whole setup cost me about 1.2K and it makes a real difference. But if you can only afford one upgrade, buy the microphone before the camera. I use a Shure SM7B, about 300 dollars; on a 100 dollar budget, get the Samson Q2U (39:11).

The model landscape and running eight terminals (47:12)

While the render finished I showed the topic of my next video: I think Grok is going to break into the top three, mainly because of the Cursor acquisition. Their new 4.5 model sits right below Opus 4.8, still below Fable of course, and above GPT 5.5, while being super cheap. For how little time they have been at it, that is impressive.

Someone asked how I manage eight terminals at once (49:09). Four are Claude Code, four are usually Codex. I assign work from hardest task to smallest, give agents their goals, and watch over them. In practice I run one terminal per client, and the ones at the bottom are usually doing something for me.

The finished AI edit (50:20)

The HyperFrames render completed just before I had to leave. The two second intro I asked for was there, "What is quantum physics", followed by the icons landing on the right words. You can still tell from the animation that it is AI generated, but it is genuinely nice, and the system is flexible: it can zoom your face in or out, pin you to a corner of the screen, and put another image behind you. We are not at human-editor level yet, but we will get there.

Q&A highlights

• Are dynamic workflows rebuilt every session? Yes, each sub-agent is built from scratch, and it inherits whatever tools your main session has (13:10).
• Why Gemini Flash Lite instead of a stronger model? Smallest model, lowest cost, and this kind of video does not need more intelligence. You can use 3.1 Pro, but I do not recommend it for this (26:27).
• Can Codex and GPT run the same workflow as Claude Code? Yes. I edit my videos with both GPT and Claude models and there is no difference (36:12).
• HyperFrames or Remotion? I prefer HyperFrames (39:11).
• Any tool for a realistic avatar video? HeyGen, from the same team that makes HyperFrames (39:33).

These sessions are going to be happening more often. If you want the rest of the series, all past recordings live on the webinars page.