The argument in one line.
A YouTube creator replaced his editor with a multi-skill Claude Code pipeline that transcribes, plans cuts, generates motion graphics, and re-checks its own work, while the human only approves at four fixed gates.
Read if. Skip if.
READ IF YOU ARE…
- You publish long-form video on a repeat cadence and post-production is the bottleneck between recording and shipping.
- You already direct Claude Code with custom skills or subagents and want a concrete blueprint for a longer multi-stage pipeline with real checkpoints.
- You're building (or want to build) an internal tool — like a b-roll library or a publishing calendar — that an agent can query and write to automatically.
- You want to see what 'human-in-the-loop' actually looks like in a working system, not just as a talking point.
SKIP IF…
- You're looking for a consumer app or plugin to install — this is a custom-engineered pipeline built over months, not a product.
- You don't use a coding agent at all — the entire system assumes a terminal-based agent orchestrating scripts, not a GUI editor with AI features bolted on.
The full version, fast.
Brian Casel now edits every video on his channel with a custom Claude Code pipeline instead of a human editor. A wrapper skill orchestrates sub-skills for FFmpeg cutting, ElevenLabs transcription, and branded motion-graphics generation, plus two custom apps: Tubery, a searchable b-roll library, and a publishing calendar. The process runs in three phases — prep, produce, export — and re-transcribes the edited footage after cutting specifically so Claude can catch its own mistakes (duplicated lines, cut-off b-roll, off-center picture-in-picture framing). Casel only steps in at four gates: approving the cut plan, reviewing the first-pass edit, approving fixes, and never touching the final stitch. A run that used to take over a week now finishes same-day, with an hour or two of unattended agent work in the middle he can spend on other tasks.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:50
01 · Intro
States the problem: editing used to take over a week even with hired editors; a year of failed AI tools; as of a month ago, Claude Code edits every video on the channel, including this one.
01:50 – 04:50
02 · My AI video editing system
Maps the full system: raw footage in, Claude Code orchestrates a 'video editing' wrapper skill over 'video use' (FFmpeg) and 'create visuals' (Hyperframes), ElevenLabs for transcription, Tubery for b-roll. Shows the three-phase Prep/Produce/Export diagram with approval gates.
04:50 – 05:51
03 · What raw footage looks like
Shows raw scripted teleprompter takes with multiple retries, and unscripted screen-recording segments where he riffs and rambles — the messy input the agent has to cut down.
05:51 – 11:36
04 · Kicking off Claude Code
Walks his content repo and skill folder structure, explains why he's using Claude Fable 5 for this run (previously Opus 4.8), and kicks off Phase 1 (prep: transcribe, analyze cuts, propose b-roll and new visuals) with a one-line hint about the video's topic.
11:36 – 21:15
05 · Approving the edit plan
Reviews REVIEW.md and each segment's transcript.md with cut/b-roll/visual tags; approves the default color grade; drops in new b-roll clips and requests a teaser clip; gives the go-ahead into Phase 2 (first-pass edit + motion graphics), which will run unattended for 1-2 hours.
21:15 – 32:00
06 · Reviewing the edited segments
Returns after ~2 hours; reviews finished segments and generated motion graphics; requests fixes via screenshot + text (graphic overlap) and via '[FIX]' tags written into transcript.md (duplicate phrase, cut-off b-roll, off-center picture-in-picture framing that also gets written into the skill as a permanent check); approves final stitch.
32:00 – 37:17
07 · Final export and publishing
Fixes complete after another ~1 hour; final 4K master exported; automated cleanup renames and backs up the file to Dropbox, wipes working files, pushes new b-roll into Tubery, registers the video in the Sparkdrop publishing calendar, and auto-writes the YouTube description plus timestamped chapter list.
37:17 – 37:51
08 · The AI-native approach
Closing framing: the point isn't the specific tool but the approach — rebuilding a core business process around AI while deliberately designing the moments where human approval still matters.
Lines worth screenshotting.
- Claude Code now edits every video on this channel end to end, cutting turnaround from over a week to same-day.
- The pipeline re-transcribes the footage a second time after editing specifically so the agent can catch its own mistakes, like a repeated phrase or a b-roll clip cut off too early.
- Motion graphics are generated one at a time in a queue, not in parallel, because running the graphics engine on multiple clips simultaneously was more error-prone in testing.
- A full first-pass edit — cuts, b-roll placement, and motion-graphics generation — runs unattended for one to two hours before the creator reviews anything.
- Human approval happens at exactly four gates: the cut plan, the first-pass edit, the fix pass, and never again once the final stitch starts.
- Feedback is given by writing plain notes directly into a timestamped transcript file, including a shorthand '[FIX]' tag the agent expands into a full diagnosis on its own.
- A custom-built b-roll library app stores hundreds of past clips with AI-written descriptions, so new scripts can automatically reuse existing footage instead of resourcing it from scratch.
- A recurring defect — the presenter's head not centering in the picture-in-picture frame — got fixed once, then written directly into the skill's own instructions as a permanent verification step so it never has to be flagged again.
- Camera color grading and audio-sync correction between a separate microphone and camera feed are handled automatically with zero per-video input from the creator.
- The final cleanup step auto-generates the YouTube chapter list with real timestamps derived from the finished cut, replacing manual scrubbing through the timeline.
- The entire system runs on Claude Code, a general-purpose coding agent, not a video-specific AI product — every editing rule lives in custom-written skill instructions layered on top of it.
- The creator reports a same-video comparison: total review time across all human checkpoints is roughly 20-25 minutes, versus multiple hours of unattended agent work in between.
Four checkpoints turn a long agent run into something you can actually trust.
WHAT TO LEARN
The value isn't that an AI edits video — it's a pipeline broken into fixed phases with named approval gates, so hours of unattended agent work never run further than the point you last signed off on.
01Intro
- Editing used to take over a week even with hired editors; a wrapper skill over an FFmpeg-driving engine now takes it end-to-end same-day.
- For well over a year, none of the off-the-shelf AI editing tools moved the one metric that mattered — total time from raw footage to published video.
02My AI video editing system
- The system runs on Claude Code, a general-purpose coding agent, not a dedicated video product — the editing logic lives entirely in custom skill instructions layered on top.
- Re-transcribing the footage a second time after cutting is a deliberate, paid-for step that lets the agent catch its own mistakes instead of the human catching them.
03What raw footage looks like
- Every unscripted screen-recording segment gets far more aggressive cutting latitude than scripted camera segments, because rambling needs a heavier edit than a scripted take.
04Kicking off Claude Code
- Running multiple motion-graphics generations in parallel was more error-prone in testing, so the skill was redesigned to queue them one at a time.
- A one-line topic hint is enough to kick off the entire prep phase — the heavy lifting is in the skill instructions, not the prompt.
05Approving the edit plan
- Feedback is written as plain notes directly into a timestamped transcript file, including a lightweight '[FIX]' shorthand tag the agent expands into a full diagnosis on its own.
- A searchable b-roll library with AI-written descriptions lets new scripts automatically reuse hundreds of past clips instead of resourcing new footage for every mention.
06Reviewing the edited segments
- A recurring visual defect gets fixed once, then the fix is written back into the skill's own instructions as a permanent check — the system accumulates institutional memory instead of repeating the same correction on every video.
- Total human review time across every checkpoint is roughly 20-25 minutes on a ~19-minute finished video, versus multiple hours of unattended agent work in between.
07Final export and publishing
- Automating the tedious final steps — renaming files, backing up to Dropbox, wiping working folders, writing YouTube chapters with real timestamps — removes dozens of small manual tasks with one command.
Terms worth knowing.
- Skill (Claude Code)
- A packaged, reusable set of instructions and scripts that gives a coding agent a repeatable specialized workflow — here, video editing, b-roll placement, and motion-graphics generation are each their own skill.
- Video-use skill
- The underlying skill that operates FFmpeg directly to perform cuts, stitching, and export — a lower-level engine that a higher-level 'video editing' skill wraps and directs.
- Hyperframes
- The motion-graphics generation engine used by this pipeline; a separate 'create visuals' skill wraps it to apply consistent branding and layout templates.
- Tubery
- A custom-built web app that stores and indexes a creator's b-roll clip library with searchable, AI-written descriptions, so an agent can find and reuse relevant past footage.
- Approval gate
- A fixed checkpoint in an agent pipeline where the automated process stops and waits for a human to review and approve before continuing to the next phase.
Things they pointed at.
02:39productTubery (custom b-roll library app)
03:11toolHyperframes (motion-graphics engine)
34:57productSparkdrop (custom publishing-calendar app)
Lines you could clip.
00:00
“Back when I started making videos on my YouTube channel, it used to take me well over a week to get one of my videos edited and ready to publish.”
02:30
“That's how the Claude Code Agent is able to self verify its own editing work.”
21:46
“When I have it creating multiple motion graphics at the same time... it's more error prone when it's doing all that processing all at the same time. So I designed the skill to work in a sequence and put them into a queue.”
31:20
“The issues here were so minor and I'm using Claude Fable on this so I have a pretty high level of trust that it's gonna be good to go.”
37:54
“I took one of the most important and most time-consuming processes in my business and I retooled it from the ground up using AI. And then I designed the specific moments where I, as the human in the loop, still need to give my input.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Back when I started making videos on my YouTube channel, it used to take me well over a week to get one of my videos edited and ready to publish. Not even including the time it took me to plan and script and record the actual footage. And even when I hired video editors, it was still a major bottleneck in my process because I take such a detail oriented approach to my edits and the specific b roll clips and the sequencing and the pacing.
00:22And there's a lot that goes into it that you may not notice on the surface. And so for well over a year, I've been on the hunt for ways that I can incorporate AI into how I edit my videos. And I've tried all the tools out there but none of them really made an impact on the one metric that matters which is the actual amount of time it takes to go from recorded footage to publish ready until now.
00:44So as of about a month ago, all of my videos on this channel have been completely edited using Claude code, including this video that you're watching right now. And so that enables me to record a video in the morning and have all of the post production done by the end of the same day. From editing out my bad takes to generating custom motion graphics to selecting b roll and stitching it all together, all of that is now handled with a complex set of agent skills orchestrated using Claude code.
01:12So today, I'll show you the entire process from end to end. But what I really want you to take away from this is a glimpse into what it looks like to take an AI native approach to our work. Now, this is one of the more complex AI systems that I've assembled in my business.
01:27It took a lot of iteration and trial and error, and I'm still improving it on every run. But this is what it looks like to rethink at a fundamental level the most important processes that drive our business and see how we can retool those using AI. And most importantly, I'll show you the points where the human in the loop, me giving my input and my approval, really matters and how I design that into this process.
01:50So let me start by mapping out all the components at play in my video editing system. We start with my raw recorded footage of me on camera and my screen recordings. And then I use Claude code to initiate and orchestrate the entire process.
02:04I invoke a custom skill that I created, which I call video editing, which interacts with a different skill called video use. And that is sort of an engine that operates FFmpeg.
02:15My custom wrapper skill is what dials in the editing process for exactly how I need it for the way that my videos need to come together. Transcripts drive a lot of this, so I'm using Eleven Labs for that. And you'll see how I actually retranscribe the video multiple times throughout this process.
02:32So that costs a bit more, but it's how the Claude Code Agent is able to . Now, and placing b roll clips throughout my videos has historically been a really time consuming process.
02:44And so I built this custom app that I call Tubery and that's where I store and index my library of b roll clips that I ever use and reuse throughout my videos. And Claude is able to analyze the script of the current video and choose the most relevant b roll clip from my library and place it in. But for most of my videos, we also generate new visuals like motion graphics.
03:06And for that, I created another custom skill called create visuals. And that one makes use of a different skill called hyperframes. So that's the engine that helps me create motion graphics like the one that I'm showing you here.
03:20And so my custom wrapper skill dials in the branding consistency that you see across all of my motion graphics. Now while the vast majority of the work is done by Claude Code, there are some critical points where I, as the human in the loop, need to give my approval and my input.
03:37Here's how that process plays out. We start with my raw recorded footage and then Claude runs phase one where it does a bunch of prep work. It's running transcriptions, it's analyzing where to make the cuts, where to place b roll, where we need create new b roll, and it preps all that for my initial review.
03:55And so once I approve, that kicks off phase two. And that's where Cloud works for a while and does its first pass edits. It creates and generates some motion graphics and it presents each segment for my review.
04:07Now that phase can sometimes take over an hour, which is a good time for me to do other work like creating the thumbnail for this video. Now I sometimes suggest a few fixes and all I need to do there is write a few notes into the text transcript, and then Claude recuts those segments with my feedback incorporated.
04:24And once I approve, Claude proceeds to its phase three and that's where it stitches together all the segments and exports the final video. So then we have a few final cleanup tasks like deleting the working files and saving new b roll clips to my library in Tubery and then prepping everything for publishing on YouTube.
04:43And all of that is automated using my skill with Claude Code. So that's the process. Now let's run it on a real video.
04:50Now just to show you what we're working with here, here's a look at some of my raw footage. Some of my segments are scripted where I speak off of a teleprompter and I do multiple takes for most of the lines. Take a look.
05:03When I started making videos here on my YouTube channel, it used to take me when I started what?
05:12Back when I started making back when I started making videos on my YouTube channel, it used to take me and some of my segments have me showing something on screen, which is unscripted. I'm mostly just riffing, aka rambling a bit too much.
05:26Have a look. Our I went out to dinner, took a walk. It might have even been closer to two hours that it worked on that phase of might have might have even been closer to two hours that it worked on that phase.
05:40I saw that it spawned multiple background. Eventually, my good takes are in there somewhere, but it takes a talented editor to piece it all together or a well crafted AI agent.
05:51Alright. So my starting point for this whole process is to take the raw footage that I've already recorded, which, you know, lands in my Dropbox folder so that I have backups over there for all of the raw footage and that's right here.
06:06So I usually record the footage sort of like this where I number each segment of the video and I'm either recording just directly to camera where I'm, you know, reading off of a teleprompter or I'm recording my screen and my camera in two different feeds where I'm showing something on the screen and I'm sort of riffing without a script while I turn to the camera.
06:26If you've seen any of my videos, those are like the two types of segments that I do. And then I number them and organize them like in the sequence that they will end up in the video like that. So basically, I back these up into Dropbox first and then I move these into another video folder and that's here in this repo.
06:46So here I'm working in my content development repo. It's darkened here because I don't wanna push this folder and its contents into GitHub because there are a lot of like really large four k video files in there.
06:59Now, this repo has a whole bunch of custom skills. Most of these are custom, like two of them are third party skills, but the rest are my custom designed and engineered and crafted skills that follow my entire content process.
07:16And I've gone through multiple weeks of iterations on these things. The one we're gonna be focused on today is mainly this video editing skill.
07:26And this is gonna run the entire process from end to end. Now, within this, we're sort of like wrapping and referencing multiple other skills such as this video use skill.
07:37We're also calling on the hyper frame skill and actually we're gonna be using another one of my skills called create visual and that's what we'll be using to generate custom motion graphics that will be shown throughout this video as you've probably seen in some of my other videos.
07:53So all of that stuff is gonna be handled with VOD code following my process that I've built into these pretty complex set of skills. Now to kick off the whole process, my first move here is to just run video editing.
08:06So I'm just gonna give it a hint and I'll just say, today we're editing the video about using Claude as a business brain.
08:19So that is actually the topic of the video that we're gonna be editing here today. Probably by the time that this video drops, that video will have already published.
08:28So you can actually see the final result of the video that we are gonna be editing today. It should already be live on my YouTube channel. Now, when I kick this off, it's going to start what I call phase one.
08:40There are basically three big phases that we're gonna run through throughout this editing process and phase one is what I think of as like a prep phase. So once I initiate this, Claude is going to, you know, gather up all that raw footage. It's going to transcribe it by sending it to Eleven Labs and getting the initial transcripts for each of the segments.
09:03And then Claude is gonna do a deep analysis of the content within all of the segments that I've recorded. It's gonna actually like markup the transcript to decide where it's gonna make some cuts, where it's going to remove dead space.
09:17It's also going to find b roll that it could use. Perhaps we have existing b roll that we've already produced in previous videos that we can reuse here. And for that piece, it would actually go into my custom built app that I call Tubery.
09:33And that is my library of b roll content that we've developed over the last several months. And all of this is like filed away with, you know, really descriptive descriptions and keywords and things. And so Claude is able to access this using the API that I built on it.
09:49That's how it finds the right b roll to apply. But actually more times than not, it's going to find opportunities in the script to generate new b roll, mostly new custom motion graphics.
10:04So it's gonna find all those opportunities, it's gonna mark them up, find the right spots, plan out where it's gonna make all the cuts. It's also gonna do some color grading and it's gonna, you know, give me a few options on that. So, all that is just prep work even before it even starts to get to work.
10:18So, it's gonna start that phase one now and kick off this process. And so I'm gonna pause here and I will I'll return once that prep phase is all done and ready for me to review and green light into the next phase. By the way, I wanna mention that for today's video edit, I'm actually using Claude Fable five.
10:39Now, all of the video edits that I've done on the previous videos published in the last couple of weeks on my channel, those were done using Claude Opus 4.8. Now, I'm recording this on the day after Claude Fable five finally came back and became available again to use, so I'm gonna take advantage of it while I can, especially while we are here before this July 7 cutoff date when they're gonna actually start to charge extra for it.
11:04So I'm gonna make use of it, but I just want to note that I have been able to run this process totally successfully without a hitch using Opus 4.8. I expect Fable five will will be a little bit smoother, but really it's just following all of the specific instructions my very complex video editing skill here with all sorts of like instructions and scripts and things built in.
11:28So today, we are using Label five.
11:35Okay. So Claude's initial phase where it did all of the prep work, that is all done. So I think that initial phase took maybe about ten, twelve minutes.
11:45What do we have here? So inside of the editing folder in this in this little project folder, there's a whole lot of new files that Claude sort of cooked up here. Now most of these files that are within these different segment folders, most of these are like working files.
12:01There's a lot of like JSON stuff in here. There are like little fragments of things that it's doing. It's it's sort of like doing its own little analysis and storing all these working files.
12:11So most of that stuff, I don't really need to touch. That's just for Claude's technical use throughout this project.
12:20But I did design the process to prepare things in a way that's easy for me to quickly review and give my input back. The whole idea when I design a complex agent skill like this is to make it really optimal for me to just step in and do my input and then walk away and do something else.
12:41So the main starting point for me is at this first gate is for me to look at this review.md. So Claude created this.
12:50This is sort of like its main deliverable to me, and it's sort of like a checklist of things that are ready for me to read through and review. Now, actually, kinda skip past this one and I go directly into each of these segments. So these numbered segments are like the segments that will make up the video.
13:08There's gonna be like a camera introduction, and then there's gonna be like a screen segment, then another camera, then another screen segment, and then an ending segment. Right?
13:16So that's like one, two, three, four, five. Now, something to be clear about. Claude has not started actually editing any video yet, and it also has not yet created any motion graphics.
13:27This is all prep work. It's like kind of laying the groundwork so that I can then give it the go ahead to start that phase two.
13:35So we're gonna dive into the first segment here and what do we have in here? So first thing I like to look at is this grade folder. And basically, Claude is going to do a little bit of color grading on my camera to either warm it up or make it a little bit punchier.
13:50So this first one is like the default footage that comes out of my camera. It's like warm, cinematic, color grade. It's kinda hard to tell the difference between these, but, you know, you can see it closely.
14:00I think that this one works fine. So I'm actually not gonna give it any course correction on this. It's just gonna go ahead and default to this color correction for all of the camera shots throughout the video.
14:11So that'll be good. That's a nice touch. The other thing that it's doing like under the hood a little bit, and I had to go through some iteration on this is, you know, I had to make sure that my audio from the mic is directly aligned with my mouth and the camera footage.
14:28Sometimes it was off by a couple of seconds based on my camera setup here. So I actually built in a process for Claude to self fix that. It does that through some of these like WAV files.
14:39So it's like matching up WAV files to the video files and it gets a little bit technical in there. The other thing that I will immediately start to look at in each of these segments is transcript dot m d.
14:51And throughout this process, I'm gonna use this transcript dot m d for each of the segments as my source of truth to see what's happening in the edit and to give my feedback on the edit.
15:07So I designed this sort of like interface using markdown files between me and the editor throughout the process. So what do we have here initially? At the very top is sort of like an overview for me, but ultimately, we have like time stamped line by line transcriptions of everything from this particular segment.
15:28This is intro segment. This is probably like the first three or four minutes of the video. So it did send it out to Eleven Labs.
15:35It came back. We have this transcript now. And Claude went ahead and time stamped everything.
15:41Or Eleven Labs did that, but Claude put it into this format. Now you'll also notice that there are these little tags. One like this says cut and and then it's sort of like a closing tag.
15:52So I worked with Claude to create this little like syntax where we have like these wrapping tags sort of like HTML that defines exactly what Claude is going to remove. So all of this was probably like bad takes, like I I said something a few times and it's gonna keep the good take, which is typically the last time I say it.
16:11So there are these like cuts throughout and I don't really need to do anything but I do see them here. If I see something that's off, I could I could course correct it but I really don't touch it. I'm just gonna let it go with these proposed cuts that it's gonna do when it goes ahead and creates this.
16:26Now the other thing that it does is it checks my b roll library to see if I have any b roll clips that we wanna use anywhere. I'm gonna come back to that in just a second.
16:36But it's also proposing that, hey, in this part of the script, I think it would be good to generate a new motion graphic. We can create something new that helps to visualize the mental model that I was describing in the script.
16:50So Claude just identified that and said, hey, at this moment at six minutes and nineteen seconds, that's where we should put this new motion graphic and here's a quick concept what for what that should be. And actually, think that's perfect. I think that's I think that really fits.
17:04So I'm just gonna leave it in and not gonna course correct it. Good to go. Let's jump over to the second segment, which I think is a screen, which is like a screen and picture in picture segment.
17:16In these segments, which are not scripted, it's just me riffing, that means I'm rambling a lot. So you can see that it made significant cuts and it sort of like summarized all of the cuts that it made. And here I'm I instructed Claude to have a lot more creative freedom and empower it to go ahead and just cut out the fluff.
17:37You know, I tend to describe things a lot and maybe I say the same thing multiple times in different ways, but really we only need the best take of that. Claude is gonna do its best to cut that stuff out.
17:48It's also gonna, you know, remove any like dead air, dead space. It's going to eventually cut out all of the ums and the ahs and I do a lot of those when I'm screen casting and I'm just riffing.
18:01So it's gonna try to catch most of those while keeping it natural. One thing I forgot to do before I initiated this process and it's fine to do it now, actually that's the purpose of having this checkpoint in the process, is I did need to record some new b roll.
18:18This is not motion graphics. This is like screen recordings. So for that I usually just like whip out a screen studio for these like quick like five second recordings.
18:28Here's what I'm telling Claude. I'm saying I have some new b roll clips that I just dropped into the b roll folder. Use one titled brain files around where I talk about this, where I say that in the script and then use the AI native workshop when I mentioned that in the promo segment.
18:45Be sure to push those clips into Tubery. That is my custom app for this is Tubery here.
18:52Have it's like so I have this like clips library where I store all of b roll clips and it's kinda nice. You can like hover over it and it has these little previews of what each of these b roll clips are.
19:06And I have my agent or Claude code go in and it automatically summarizes all of these.
19:13So it uses these detailed descriptions to automatically pick out past b roll clips and reuse them in relevant parts of new scripts.
19:24Right? So it's really handy to have this custom built app that is just my library for all of my b roll. I have like hundreds of clips that we're generating all the time.
19:34And then one more thing is near the end of the first segment, I sort of introduced the video. Sometimes I like to have Claude create a clip of a future part of the video where it shows me in the meat of the video.
19:49I like to tease that near the beginning. And so that's exactly what I'm telling Claude to do here. And the video editing skill knows that that's a pattern that I like to do and it's expecting me to ask for that and it sort of just knows how to do that.
20:04So I'm gonna just give it those things and then finally I'm just gonna say everything else is good to proceed with the next phase.
20:15Okay. So it is just getting to work now and we're now entering that phase two where it's going to actually start to generate some custom motion graphic b roll clips. It's going to actually edit every one of these segments and it's gonna make all the cuts.
20:29So it's gonna go ahead and actually do all these edits. I'm gonna let that cook. It's gonna take a while.
20:34It might even take like up to an hour or so. I'm gonna let it work on that and I'm gonna take a break and go to dinner. By the way, if we haven't met yet, I'm Brian Castle and this channel is about working AI natively.
20:45Not just reaching for AI here and there, but taking the core processes that run your business and rebuilding them around AI. Exactly like what you're watching me do here with my video editing right now. And if you wanna make that shift yourself, I put together a free workshop on going from AI dabbler to AI native.
21:03I walk you through the method behind systems like this one so that you can start to retool your own work. It's completely free over at buildermethods.com/workshop. Alright.
21:14Let's keep going. Alright. So Claude cooked on that for well over an hour.
21:18I went out to dinner, took a walk. It might have even been closer to two hours that it worked. Saw that it spawned multiple background processes to get this done.
21:26Also, part of the reason why it takes a while is because I have engineered the skill to work in a certain sequence, especially when it comes to generating custom motion graphic visuals.
21:40In my earlier testing, I found that when I have it creating multiple motion graphics at the same time using the Hyperframes engine and the other scripts that are involved, it it's more error prone when it's doing all that processing all at the same time.
21:53So I designed the skill work in a sequence and put them into a queue. So it creates all those, which takes a while, and then it goes through and it actually does edits for all the segments. So, you know, there's been a lot of like iteration and technical tweaking to get this really finely tuned to the point where I could step away for hours and it just works.
22:14Then I come back for this review session. So this review session typically only takes me about twenty, twenty five minutes where I do a pretty deep review of its work and I'm gonna do that now. So what we have here is a finished edit for each of the five segments that are ultimately gonna be stitched together to make the final video.
22:35And here's the final message that Claude gave me. You know, it says like all five segments are ready. Looks like the total run time for this video is gonna come in somewhere around nineteen minutes and that's pretty typical for my videos.
22:47It gave me a few notes in here but really all I need to do is start to dive in to each of these segments and, you know, I watch them back. So this first segment is only like a minute thirty two. That's like the intro section.
23:00Now, off camera I did already review these closely and I have some fixes that I'm gonna ask it to make and I'm gonna show you those here. But the other thing that I like to do before I even start to watch these back is I like to check out the new custom motion graphics that were created and generated and those are inside of this viz visuals folder.
23:21And so these were created using hyperframes but then I have my own create visual skills with which gives it my branding and my specific structure and content templates and things like that. So I did review these and these are mostly pretty good and I can sort of like look at these and so these are gonna be graphics that show up in the video but I'm also gonna save them into my library and I can use them in things like social media posts a lot.
23:45Now this is actually the one issue that I noticed across all of the custom graphics is that I don't like this. The little text thing sort of overlaps. So I'm gonna ask it to fix that just like I would ask it ask Claude to fix any other like UI bug in website or something.
24:02I'm gonna take a screenshot, copy that to clipboard, and I'm going to paste it in here and I'm going to say, in this motion graphic, the can't see in and out text and the red circle come too close to the right edge of the box.
24:20Let's move that red circle and its text to the center of the line between two boxes.
24:31I think Fable will know what I'm going for there. Okay. The next fix that I need to ask it to do is something that I've noticed multiple times in some of my previous edits.
24:42So I'm gonna ask it to fix this, but I'm also going to ask it to bake this verification into the skill itself so that I don't need to ask it to fix this again on the next one.
24:53If you notice, you know, my head sort of locks back and forth a little bit, but most of the time I find that my head is a little bit too far to the left within this framing. So, you know, it doesn't perfectly like center my head and my body in this picture in picture frame.
25:10And I've noticed that's the case in most of these segments. I noticed in both of the screen picture in picture segments, my head is not horizontally centered in the PIP frame.
25:26I'll need you to fix that in segments two and four.
25:31And this is something that I've noticed in previous video edits. So I also want you to find the correct location in the skill instructions and add a verification step to look for this specific issue.
25:48I want you to take multiple screenshots of picture in picture segments and check to see if my head is not horizontally centered.
26:02And if so, fix that before finalizing this segment. Alright.
26:07So we've got that one in. Okay. So then I go through and I actually watch back every second of every segment.
26:13And as I do that, I usually kinda make the video small. And I'll put it sort of like up here. And then watch it back.
26:22I will read through this transcript because this is where I can mark in any fixes that I wanna make. So this was actually one of the only glitches that I think it actually did a really good job this time.
26:35But I did notice this one. So somewhere right around here, it I sort of said the same thing twice back to back. So basically at the end of this phrase, I'm actually running paid traffic because I'm focusing on organic traffic.
26:49But then I say, because first I need to focus on organic funnels. So they're slightly different wording but I'm saying the same thing twice and it would be awkward to hear me say that like that back to back in the final video. So I have actually set up the skill so that if I wanna be really lazy about it, all I really need to do is just type in fix like that, brackets.
27:13And Claude will just find that and analyze the whole area around here and just know that like something is wrong here. Brian pointed it out. I need to fix it.
27:21And it typically will fix it. If I wanna be a little bit more descriptive, I can add it to the prompt and that's actually what I'm gonna do here. I'm just gonna say, segment two, I sort of said the same thing twice here.
27:32I added a fix marker. I want you to delete the first duplicate phrase which was the This is actually a good moment to tell you how the self correction system is built in.
27:46So what happens is this transcript is not the original transcript that we had at the very beginning of the process. This was a re transcription of the video after Claude made all of its edits.
28:00So I do have Claude, you know, show me the initial transcript where it says what its editing plan is going to be, like it marks what it's going to cut out and where it's gonna put b roll and whatnot. Then it makes all those edits.
28:15It cuts a lot of the content out of the original footage. And then it sends it to Eleven Labs again, and then it gives me this transcript.
28:23So why do I do that same cycle and actually pay for that extra transcription process? Well, that's how it actually can catch its own mistakes.
28:32Right? So it reanalyzes the transcript after it does all the programmatic editing work to spot things like that, like where I I say a phrase twice or maybe some dead air that was inadvertently left in the script or, you know, something just doesn't make sense and so it needs to piece together some of my sentences.
28:53It's gonna look for those things and if it spots its own errors, it's gonna just read and run the whole edit again. And that's why it takes like two hours for it to do that work. But again, I could be doing something else while it works And I know that when I come in here, there's only like one or two small fixes that I need to ask it to make.
29:09So it's incredible. Alright. One more issue that I noticed also in the second segment near the end of it around thirty five, we show some b roll.
29:17It was one of the motion graphics that was created. We show that here. And so what I noticed was that it sort of cut off that b roll a little bit too soon.
29:26It didn't give it enough time to play out, which is a little weird for the viewer. And I think that there was actually an extra word thrown in there. So the b roll clip is not played for long enough.
29:38Also there seems to be an extra word stuck in there. I added a fix marker in that spot.
29:45Alright. So it's gonna find that and clean it up. Alright.
29:47So now here is one more thing that I noticed in segment four which, you know, the editor really wouldn't have picked up on its own and that is there are like two spots in here where I'm describing something and I didn't realize this at the time I was recording it, but I was sort of showing something on the screen that was here at the bottom right corner of the screen, which means it's gonna get covered up by the picture in picture.
30:10So there was about a minute in the beginning and a minute near the middle where I do that. And it's gonna be really confusing for the viewer if this thing is hiding what I'm showing. So I sort of prepared these two.
30:23In segment four, we need to fix this picture in picture covering up some content on the screen. Here are the details.
30:31So I gave it like the span of time where I want the picture in picture to sort of fade out and then fade back in. And then the only final issue which was in the final segment like the ending, just didn't use the b roll for when I talk about the AI native workshop. So I'm just gonna add that in.
30:48Other than these issues, all other segments are locked and good to go.
30:54Please go ahead and make these fixes and then when you're finished you can immediately proceed to the final phase which is to merge all of the segments together and export the final four k m p four file.
31:09In my previous edits I probably would have had it just make those fixes and then I would review them again and then I would give it the go ahead to do the final stitching. But the issues here were so minor and I'm using Claude Fable on this so I have a pretty high level of trust that it's gonna be good to go.
31:26So I sort of just wanna get the whole thing done and if there are any remaining issues after that, I could always go back and ask it to fix that segment and rerun it if I have to. But I'm pretty confident it's gonna be good to go.
31:39So it's gonna make those edits and that's probably gonna take another, you know, at least thirty minutes or so. It's gonna work on that and meanwhile, I can go, you know, work on something else like I started working on the thumbnail for this, which I still do in Adobe Photoshop.
31:53So, yeah. I'm gonna work on that a bit while that's cooking and and then this thing should be pretty close to ready to go. Alright.
32:00So all of those fixes are now complete. That took a little while longer than I expected. It took again over an hour to make those edits because it went through multiple rounds of checking its work and, you know, tuning some things and I didn't have to step in at all.
32:14I just waited for it while it went through and I just worked on the thumbnail and some other stuff while it worked. So here we are. It is all done.
32:22The final video came in about nineteen minutes. So that looks good. So but, you know, I did just go through here and I found those actually, watched the whole thing through but I did notice that all those specific fixes that I asked for have been completely made.
32:37So all of that is good to go. And so now I have the final m p four file here, but still all of the original working files which included, you know, a lot of like raw footage and like little sub cuts and lots of little working files and things.
32:54Lots of stuff is sort of just accumulated in this folder. So we needed to do some cleanup but actually there's a number of like final cleanup items that I need to do at the end of every video editing project. And so, you know, I figured out like what all those little cleanup tasks are and I just built them all into the very end of the process in this skill so that with literally one click of a button, all of it can just be done automatically.
33:21So I'm not gonna go too deep into the weeds here but basically what we're gonna do is we need to rename that final file, you know, which currently like named with the date and it's sort of like temporary. Now, you know, obviously, it's really easy for me to just click and rename it right now, but that's just like one of several little tedious steps that has to be done right every time.
33:42So I like to rename it and make it sort of a keyword friendly file name which is then ready to actually upload to YouTube. We're gonna do that. It's also gonna move this file out of this folder into another location on my system which is backed by Dropbox.
33:57So I have like the final version and I have the original unedited raw footage all backed up in Dropbox. And so once it verifies that it has moved it there and it is there, it's gonna go ahead and just wipe this working folder so that we get that stuff off of my hard drive.
34:13It's taking up a lot of space because this is a lot of four k video. So it's gonna delete that for me after it kinda does the file management stuff. A couple other things, we're going to we're gonna take all the new b roll that we created during the course of this project and it's gonna push that stuff up to Tubery.
34:31Tubery is my custom built app that I use to, you know, manage all of my b roll so that like next time if I cover a topic that covers the same, you know, the same topic, I can easily just pull one of those pre existing motion graphics that we generated here. I also have a custom app, another custom app called Sparkdrop and that's where I manage my pipeline and my publishing calendar.
34:55And going to publish this video manually, I usually do that for the long form YouTube videos, but I still like to get it registered in my publishing calendar so that I see I have like LinkedIn posts and tweets and newsletters and stuff like where all of my videos are popping in my in my content plan.
35:14So it's gonna handle getting it up in there. Oh, and one more thing and this is a really tedious thing that it's gonna do automatically for me. It's gonna create the chapters list for YouTube.
35:25So I do already have the packaging all planned out for this video. In this case, I'm going to test out two different titles and I've already created the thumbnails for these, so that's good to go.
35:38This is like my template YouTube description. What Claude is gonna do is write the intro of the description which describes this video and what it's about.
35:48And then it's also going to fill in down below this list of chapters. That's not just a list, that has to be time stamped with the exact time that each chapter begins. So it's gonna analyze the script, analyze the final recording and it's gonna actually make that text list of timestamps rather than me having to scrub through the video and find those timestamps and do it manually.
36:12All of it is done automatically from Claude. Okay. So that is the cleanup process.
36:17Let's kick it off. Here we go. All the cleanup tasks have been done automatically.
36:22The final video has landed in the correct place. It's backed up in Dropbox along with the original footage and so that is here.
36:30You know, my next step would be to actually upload that into YouTube. I also created the thumbnail for this video. So actually I I created a and a b.
36:38I like to do some a b testing on those. I created those in Photoshop while the video editor was working.
36:44So again, I'm able to actually multitask and be so much more efficient than I used to be before I was doing it this way. Alright.
36:51So that is all set and we have the description and timestamps all set up. So here is the description.
36:58It wrote that up. I like to keep it kind of short. And then down at the bottom, we've got the list of chapters all with the timestamps good to go.
37:06And I really like the way that it broke it up, way that I like it, titled these the way that I like to. So this is basically all ready to go for me. My next step is just to throw it up into YouTube and schedule it to publish.
37:17So that's the whole system end to end. It's a lot. But the point is the approach.
37:22I took one of the most important and most time consuming processes in my business and I retooled it from the ground up using AI. And then I designed the specific moments where I, as the human in the loop, still need to give my input. And that's what taking an AI native approach actually looks like.
37:38Again, you can get instant access to my free workshop on shifting to AI native by going to buildermethods.com/workshop. And hit subscribe so you don't miss my next video on going deeper with AI. Let's keep building.
The Hook
The bait, then the rug-pull.
Brian Casel spent over a year testing AI editing tools that never moved the one metric he cared about: time from recorded footage to published video. This is the system that finally did — a set of custom Claude Code skills that now edit every video on his channel, while he steps in at exactly four checkpoints to approve, redirect, or fix.
Named ideas worth stealing.
01:56model
Three-Phase Edit Pipeline (Prep / Produce / Export)
- Prep — transcribe, plan cuts, propose b-roll/visuals (renders nothing)
- Produce — cut, composite, place b-roll/visuals, self-verify every cut
- Export — stitch all segments, export one 4K master
The full pipeline is three sequential phases, each ending at a human approval gate, running from raw footage to a finished, published video.
Steal forAny long-running agent workflow (video, docs, code) that needs a clean human-checkpoint structure instead of one giant unsupervised run
02:09model
AI Video-Editing System Architecture
- Raw footage (camera + screen recordings)
- Claude Code (orchestrator)
- video-editing skill (wrapper) -> video-use skill -> FFmpeg
- create-visual skill (wrapper) -> hyperframes engine (motion graphics)
- ElevenLabs (transcription)
- Tubery (b-roll library app)
A layered skill architecture: thin creator-specific wrapper skills sit on top of general-purpose engine skills (FFmpeg for cuts, Hyperframes for graphics), with a custom app as the searchable memory layer for reusable assets.
Steal forStructuring any custom Claude Code skill set — separate the 'how I want it done' wrapper from the reusable 'engine' skill underneath
How they asked for the click.
VERBAL ASK
21:10link
“It's completely free over at buildermethods.com/workshop.”
Mid-roll self-ad read while Phase 2 is running unattended, framed as directly related to the video's topic (going AI-native); repeated as the closing CTA alongside a subscribe ask.
MENTIONED ON CAMERA
Visual structure at a glance.
