Big Idea

The argument in one line.

Claude Code can perform all six jobs of video editing (research, cutting, planning, graphics, quality control, and learning from notes) when each job is scoped narrowly and backed by a skill that rewrites its own rules after every edit.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already edit your own YouTube videos and want to hand off captions, cuts, and motion graphics without losing a polished look.
You're building, or want to build, a Claude Code skill for a repeatable creative task rather than a one-off prompt.
You want a real sense of what AI video production costs in tokens and dollars, not just a demo reel.

SKIP IF…

You want a no-code, point-and-click tool. This workflow is built and run through Claude Code itself.
You're hoping for a single magic prompt. The system shown here took five months of iteration to look this polished.

TL;DR

The full version, fast.

A 15-year motion designer and video editor now lets Claude Code handle the entire edit of his YouTube videos: researching hooks, writing and fact-checking the script, cutting out mistakes and syncing audio to video, planning a beat-by-beat graphics map, building the graphics through parallel sub-agents, checking the final cut frame by frame against a six-point failure list, and rewriting its own rule books from his notes after every video. The video used about 1.4 billion tokens, worth roughly $445 at API pricing but covered by a flat Claude Max subscription, and took about 12 hours of Claude's working time versus three days to a week for a traditional editor. The core lesson: give Claude one narrow job at a time, not one prompt for a finished video.

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 →

Chapters

Where the time goes.

00:00 – 01:25

01 · Claude edited this video

Duncan opens by claiming every caption, motion graphic, and animated card in the video was built by Claude Code with no human editor, then previews the six jobs AI now handles and explains why his earlier AI-edited videos looked cheap until he started giving Claude one job at a time inside a reusable skill.

01:25 – 02:21

02 · Job 1: Research and script

Before writing anything, Claude studies how top creators in the niche open their videos and saves the best hook patterns to a library; when it writes the script, every claimed fact gets checked against a real source, and any fact it can't verify gets cut.

02:21 – 03:33

03 · Job 2: The cut

Claude transcribes every spoken word with exact timing, keeps only the last clean take of each line, and cuts the rest; it also measures and corrects a camera-to-microphone sync drift that was making his lips look slightly out of time with his voice.

03:33 – 04:39

04 · Job 3: The plan

A beat map lists every sentence and what should be on screen while it's said (140 beats in one past video); Claude studies reference edits from sites like promptmotion.com before building, and a full graphics plan gets approved before any single graphic is built, since a plan is cheap to revise and a finished graphic isn't.

04:39 – 06:16

05 · Job 4: Motion graphics

Claude splits the video into sections and hands each to a sub-agent that codes the graphic using a free tool called HyperFrames, built by Heygen; Opus handles planning and creative direction while cheaper Sonnet sub-agents do the building, renders now queue one at a time after parallel renders once maxed out his machine, and turning off motion blur plus rendering draft cuts at half-size made renders about three times faster.

06:16 – 06:58

06 · Job 5: Checking the work

Claude now reviews the cut at 10 frames per second instead of its original one-frame-per-second pass, checking for six specific failures: animations that play twice, graphics that flash for a split second, text shown too briefly to read, graphics covering his face, repeated words, and lips drifting out of sync.

06:58 – 08:04

07 · Job 6: Learning from notes

Every round of notes Duncan gives on a cut (36 on one video) gets turned into a rule-book update; when a new note contradicts an existing rule, Claude rewrites that rule instead of stacking a new one on top, which is meant to keep the system's 12 rule books from drifting into bloat or contradictions.

08:04 – 09:44

08 · What this edit cost

This video used about 1.4 billion tokens, roughly 500 million from Opus planning and directing and 900 million from Sonnet sub-agents building; at API rates that's about $445, but it's covered by his flat Claude Max 20x subscription, and the whole process took about 12 hours of Claude working plus his own review time, versus three days to a week for a traditional editor.

Atomic Insights

Lines worth screenshotting.

AI-edited videos look generic when you ask for a finished video in one prompt; giving Claude one job at a time is what makes the output usable.
A skill is a saved instruction set for one editing job, and it gets better every time it's used because notes turn into permanent rule updates.
Claude studies 'outlier' videos, ones that got 3x a creator's normal views, to build a reusable library of proven hook openings before writing a script.
The single highest-leverage script rule: if a claimed fact can't be checked against a real source, cut the line rather than risk it sounding made up.
A beat map pairs every sentence in a script with exactly what should be on screen while it's spoken; one video needed 140 of them.
Plans are cheap to revise and finished graphics are not, so a full graphics plan gets approved before any single graphic gets built.
Splitting a video into sections and assigning each to a parallel sub-agent turns one AI into something closer to a small design studio.
Turning off motion blur cut render times by roughly three times, because without it the computer draws every frame four times instead of once.
Checking a cut at one frame per second misses mistakes a viewer would catch instantly; checking at 10 frames per second catches far more of them.
When a new editing note contradicts an existing rule, the fix is to rewrite that rule, not stack a new one on top, to avoid rule drift and bloat.
This video cost about 1.4 billion tokens, roughly 500 million for Opus planning and 900 million for Sonnet sub-agents doing the building.
At raw API pricing a single AI-edited video like this runs about $445, though a flat monthly subscription plan can absorb that as a fixed cost instead of a per-video bill.

Takeaway

Six separate jobs, not one AI prompt

AI EDITING SYSTEM

Treat video editing as six separate jobs (research, cut, plan, graphics, check, learn), each run through its own checklist, and the system gets cheaper and more reliable with every video.

01Claude edited this video

AI-edited videos look generic when you ask for a finished video in one prompt; asking for one job at a time is what made the output usable.
A 'skill' is a saved instruction set for one job that gets refined every time you use it, so the system compounds instead of resetting each video.

02Job 1: Research and script

Claude studies hook patterns from outlier videos, ones that got 3x a creator's normal views, and saves the best openings to a reusable library before writing a single line.
The single most important script rule: if Claude can't verify a claimed fact against a real source, cut the line rather than risk it sounding made up.

03Job 2: The cut

Claude transcribes every word with exact timestamps, keeps only your last clean take of each line, and cuts the retakes and dead air automatically.
It also catches camera-to-mic sync drift, a lag most creators wouldn't notice themselves, and corrects it across the whole recording.

04Job 3: The plan

A beat map pairs every sentence with what should be on screen while it's said; one video needed 140 of them.
Build a full graphics plan before building any graphic. Plans are cheap to revise; finished graphics are not.
Feed Claude reference edits you admire and ask what specifically makes them work instead of letting it guess at style.

05Job 4: Motion graphics

Splitting a video into sections and handing each to a parallel sub-agent turns graphics work from one AI into something closer to a small design studio.
Use expensive, more capable models for planning and direction, and cheaper, faster models for the repetitive building work.
Rendering is heavy enough to max out a machine when run in parallel; queue renders one at a time, and turn off motion blur for roughly 3x faster renders with no visible quality loss.

06Job 5: Checking the work

Checking a video at one frame per second misses mistakes a viewer would catch instantly; checking at 10 frames per second catches far more.
Give Claude a specific checklist of failure types to look for instead of a vague 'check it looks right.'

07Job 6: Learning from notes

Every round of notes should become a permanent rule, not a one-off fix, the same way you'd train a human editor.
When a new note contradicts an existing rule, rewrite that rule instead of adding a new one on top. Stacking rules causes drift, bloat, and inconsistent output over time.

08What this edit cost

A video like this used about 1.4 billion tokens, split roughly 500 million for planning and direction and 900 million for the actual building work.
At raw API pricing that's about $445 per video, but a flat subscription plan can absorb it as a monthly cost instead of a per-video bill.
Factor in your own review time, watching cuts and giving notes, on top of Claude's working time. It's lower-touch, not zero-touch.
Every video should get faster and cheaper than the last one as the system learns, so expect the first few videos to be the most expensive and slowest.

Glossary

Terms worth knowing.

Skill: A saved set of instructions that teaches Claude Code how to do one job well, refined over time by folding in notes from real edits.
Beat map: A list of every sentence in a script paired with what should appear on screen while it's spoken.
Sub-agent: A separate instance of Claude assigned one narrow slice of a task, often running in parallel with other sub-agents on the same project.
Loop engineering: Giving an AI a concrete task plus a way to check its own output, so it keeps iterating until the result is actually correct.
HyperFrames: A free, code-based tool (built by Heygen) that turns webpage-style code into a rendered video, used here to build captions and motion graphics.
Render: The step where code is converted into an actual video file; it's heavy, resource-intensive work that can max out a computer when run in parallel.

Resources

Things they pointed at.

04:01toolpromptmotion.com ↗

04:40toolHyperFrames (by Heygen)

09:06productClaude Code Club ↗

Quotables

Lines you could clip.

00:00

“Claude edited this entire video.”

cold-open claim with zero setup, forces curiosity about how→ TikTok hook↗ Tweet quote

01:59

“AI will happily make up a fact that sounds right, like a view count or a date.”

concrete, specific warning that applies to any AI writing workflow→ newsletter pull-quote↗ Tweet quote

08:35

“This is the worst it's ever going to be.”

short, confident, reframes the current (already impressive) system as a floor not a ceiling→ IG reel cold open↗ Tweet quote

The Script

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

Claude edited this entire video. Look at everything on screen right now. The captions, the motion graphics, and the cards flying out from behind me.

Not a single frame of any of that was cut, timed, or even touched by a human editor. I sat down and read a script and Claude Co. did everything else.

Most people think AI editing means auto captions and a few zooms, but Claude can now do every job a real editor does. In this video, I'll walk you through the six jobs Claude does to make a video like this one. Along the way, I'll show you the one habit that makes it better with every new video.

And at the end, I'll break down what one of these edits really costs and whether I think a system like this is worth it. For over 15 years, I was a video editor and motion graphics designer for brands like Apple, PlayStation, and Nissan. And up until very recently, I edited all of my own YouTube videos by hand.

So when I first asked AI to edit one for me, I could tell right away that it looked cheap. If you've ever tried it, you know exactly what I mean.

Most AI edits have that same generic AI look. I kept at it for about five months and two things finally changed. The quad models got a lot better and I built a better system for making polished videos.

But one thing made the biggest difference. I stopped asking for a finished video in one go and I gave quad one job at a time. And that part you can copy today.

Then I turned my whole editing process into a skill. A skill is a safe set of instructions that teaches quad how to do one job and mine gets better every time. I use it.

The first job AI has to do, and it's the part everyone skips, is the research and the script, because a great edit can't save a confusing video. If you've ever watched a video that tries to cover 10 ideas at once, you know you're lost by idea three. My early scripts did exactly that.

So now before anything gets written, Claude studies how the top creators in my space use hooks to open their videos and it saves the best openings in a library. The first line of this video came from that library. It uses the shape of an opening from an outlier, which is a video that got more than three times its creator's normal views.

Then Claude writes the script, and here's the trap. AI will happily make up a fact that sounds right, like a view count or a date. So have Claude check every fact against a real source, like a file or a web page.

If it can't prove a fact, cut the line. That one rule keeps your videos clear and honest. After that, all you do is read the finished script to camera, which is exactly what I'm doing right now.

The second job is the cut, and it's the one most people picture when they hear the word editing. When you film, you're going to mess up. You'll trip over a line, start a sentence twice, or pause way too long.

One of the first things Claude got good at was cutting out those mistakes and awkward pauses. Back in May, it already caught about 8 out of 10 of my retakes on its own. Here's how it works.

You give Claude your raw recording and it writes down every word you said with the exact time of each word. Then it keeps your last clean take of every line and it cuts everything else. It even fixed a problem I didn't know I had.

My camera's picture runs a few frames behind my microphone, so my lips didn't quite match my voice. If you record with a separate camera and mic, yours might too. Claude measures that gap all through the recording and it lines the picture back up.

Giving Claude a real job like this, plus a way to check its own work so it keeps going until it's right, is called loop engineering. And it's exactly the type of thing we teach inside the Claude Code Club. I even put a starter version of my editing skill in there so you can try this on your own videos.

The link's in my description. But a clean cut is still just me talking. So the third job decides what you see while I talk.

the third job is the plan and it starts with a beat map a beat map is a list of every sentence you say and what should be on screen while you say it for my last video that list had 140 beats without a plan ai edits look like slides my early edits looked exactly like that and one had a big green color bar down the side of the screen that was just plain ugly if you want beautiful videos you need to give claude references otherwise it just guesses what looks good So show it videos from editors you love and ask it what makes them look so good.

I did that and it pulled out things like how the text moves and how one scene turns into the next. There's a great site for finding references like these. It's called promptmotion .com and it's full of motion graphics that people made with Claude and every video comes with the exact prompt behind it.

So you get the look you want and the words that made it. Then, before you build a single graphic, have Claude make a plan of every graphic mapped to your beatmap. A plan is cheap to change and a finished graphic isn't, so this saves you a lot of time and tokens on fixes later.

The fourth job is building the graphics. And this is where it stopped feeling like one AI and started feeling like a whole design studio. Claude splits the video into sections and hands each one to a sub -agent.

A sub -agent is a helper. It's just another copy of Claude that does only that one job. And in one of my videos, six sub -agents were building at the same time.

They don't use editing software. They build every graphic with code using a free tool called HyperFrames, which is made by Haygen. It works like a webpage that turns into a real video, and that's how you get captions that light up word by word.

I use the more powerful, more expensive models like Claude, Opus, and Fable for the planning and the creative thinking, and I use cheaper, faster models like Claude Sonnet for the sub -agents that do the building. One thing to watch out for though, a render is when that code gets turned into actual video and it's heavy work for your computer.

I once had nine sub -agents rendering at the same time and I totally maxed out my machine. They even started stopping each other's renders. So now my renders wait in line and go one at a time.

Two more things made my renders a lot faster. First, I turned off motion blur. To fake the blur, the computer draws every frame four times.

Without it, each frame gets drawn once, so my renders got about three times faster, and I can't even see the difference. Second, while I'm still reviewing a cut, Claude renders it at half size, so my drafts come out even faster. Only once I approve the cut does Claude render the full quality version.

All of this renders right on your own computer with free tools like FFmpeg, so how fast it goes depends on your machine. A full render can take a while, so I usually let it run overnight while I sleep. That way it doesn't lock up my computer during the day.

The fifth job is checking the work. And honestly, it's the one I underestimated. At first, Claude checked a video by looking at about one frame every second, but you don't watch a video one frame a second, so it kept missing mistakes you'd spot right away.

Now it looks at 10 frames every second and it checks for six different mistakes. One, an animation that plays twice. Two, a graphic that flashes on screen for a split second.

Three, text that shows up too late for you to read it. Four, a graphic covering my face. Five, a word I said twice in a row.

And six, my lips drifting out of sync with my voice. So if you use Claude to make anything visual, tell it to check the way you'd check, not the way that's fastest. And that brings me to the sixth job, which is the habit I promised you at the start.

Claude learns from every video. Every time you watch a cut, you'll have notes, the same notes you'd give to a human editor. On one video, I gave Claude 36 notes on a single cut.

But Claude doesn't just fix the video, it finds the rule in its prompt that would have stopped the mistake, and it rewrites that rule. My music was too loud, so the music level got a rule. two of me showed up on screen at once so now there's a check for that right now my editing skill has 12 rule books like one for the script one for the visuals and one for checking when a new note changes an old rule claude rewrites the old rule instead of stacking a second one on top this reduces drift bloat and hallucinations that's why this video looks the way it does it's not one magic prompt.

It's five months of my notes built into a skill. So those are the six jobs Claude does in my edits now, research and script, the cut, the plan, the graphics, the checking and learning from my notes. I still film it and I still have the final say, but I didn't touch an editing timeline for any of this video.

so what does one of these edits really cost for the version you're watching claude used about 1 .4 billion tokens about 500 million of those were opus planning and directing and about 900 million were the sonnet builders if you paid for those tokens through the api it's about 445 dollars but i'm on the quad max 20x plan so for me it's all part of my monthly subscription from the first research to the finished video quad worked for about 12 hours plus the time i spent watching cuts and giving notes but But I expect every video after this one to be faster and cheaper because every video teaches the system something new.

This is the worst it's ever going to be. Now, compare that to the normal way. Someone has to write the script, map the beats, storyboard it, design every graphic, animate it, edit the audio and add the sound effects.

For my years as an editor, that's about three days to a week of work. And that's for someone who's fast and really good. So is it worth it?

For me, yeah. I'm handing more and more of my work to AI. And honestly, I find it fun, it's interesting, and it's challenging in a whole new way.

If you want to try this for yourself, the starter version of my editing skill is inside the Cloud Code Club. It does these same jobs in a simpler way. It cleans up your recording, adds word -by -word captions, builds simple graphics, and it checks its own work.

It won't look like this on day one because mine took five months of notes, but it saves your notes the same way, so it gets better every time you use it. If you want to master Cloud Code in 45 days and get this starter kit plus all my skills and resources already done for you, come join over 8 ,000 builders in inside the Cod Code Club.

It's only $9 a month and the link's in my description. If this helped, hit like and subscribe. It really does help the channel a ton.

See you in the next one.

The Hook

The bait, then the rug-pull.

A professional editor opens by claiming he never touched an editing timeline. Not the captions, not the motion graphics, not the cards flying across the screen. Claude Code did it all, and he walks through exactly how.

Frameworks

Named ideas worth stealing.

00:24list

The six jobs of AI video editing

Research and script
The cut
The plan
The graphics
The check
Learn from notes

The six discrete jobs Claude is given one at a time instead of one combined 'edit this video' prompt, each with its own checklist and its own rule book.

Steal forany creator building a repeatable Claude Code editing or production pipeline

05:08concept

Loop engineering

Giving Claude a concrete task plus a way to check its own work, so it keeps iterating until the output is actually correct instead of stopping at a first pass.

Steal forany multi-step Claude Code automation, not just video editing

06:26list

Six things Claude checks for in every cut

Animation that plays twice
A graphic that flashes for a split second
Text that shows up too late to read
A graphic covering the face
A word said twice in a row
Lips drifting out of sync with the voice

The specific failure checklist Claude runs at 10 frames per second against every rendered cut.

Steal forany AI-assisted QA pass on generated video or motion graphics

CTA Breakdown

How they asked for the click.

VERBAL ASK

09:06product

“If you want to master Claude Code in 45 days and get this starter kit plus all my skills and resources already done for you, come join over 8,000 builders inside the Claude Code Club. It's only $9 a month.”

Soft, value-first pitch placed at the very end, only after fully delivering the six-job breakdown and the real cost numbers; positions the paid community as the shortcut version of his own five months of iteration.

MENTIONED ON CAMERA

04:01toolpromptmotion.com ↗

09:06productClaude Code Club ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.