Big Idea

The argument in one line.

AI video edits stop looking cheap when you split the work into six narrow jobs, save each as a skill, and feed your notes back so the instructions improve with every video.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A creator who films talking-head videos and wants Claude Code to handle captions, cuts, and motion graphics without opening an editing timeline.
Someone who tried asking AI for a finished edit, got generic-looking results, and wants a repeatable process instead of a better one-shot prompt.
A Claude Code user curious how subagents, skills, and model choice (Opus vs Sonnet) apply to a real production workflow with real token costs.

SKIP IF…

You want a click-by-click tutorial for HyperFrames or Whisper setup. The video names the tools but does not walk through installation.
You edit professionally and need broadcast-level polish today. He admits a great human editor still wins on some shots.

TL;DR

The full version, fast.

One-shot AI video edits look cheap because the task is too big, so the author splits editing into six jobs: scripting, cutting raw footage, design, motion graphics, review, and feedback. Whisper gives word-level timestamps and FFmpeg cuts the retakes. A beat map plans every graphic before anything is built, and subagents build graphics in code with HyperFrames. Claude then checks its own frames and rewrites its instructions from your notes. One edit took about 600 million tokens, roughly $160 at API prices, and six hours of Claude time.

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 – 01:51

01 · Claude edited this video

Claim, 15 years of editing credibility, and why splitting the work into six jobs and saving them as a skill fixed cheap-looking AI edits.

01:51 – 02:20

02 · Step 1: Hooks and script

Have Claude study three outperforming videos in your niche, save the hooks and formats in a library, and write scripts from it.

02:20 – 03:03

03 · Step 2: Cut the raw footage

Whisper transcribes every word with a timestamp, then FFmpeg keeps the last clean take of each line and cuts the rest.

03:03 – 04:17

04 · Step 3: Design

Give Claude screenshots of editing you love, use prompt-motion.com for examples, and build a beat map before any graphics.

04:17 – 05:45

05 · Step 4: Build with subagents

Subagents build each section's graphics in code with HyperFrames, Opus plans and Sonnet builds, renders are capped at four and drafts run at half size.

05:45 – 06:18

06 · Step 5: Check the work

Claude inspects 10 frames per second for six known mistakes using a checklist.

06:18 – 07:04

07 · Step 6: Learn from notes

Notes from each cut are turned into rewritten instructions so the skill improves over time.

07:04 – 09:05

08 · What this edit cost, and the starter kit

About 600 million tokens, roughly $160 at API prices, six hours of Claude work, and a pitch for the starter skill inside the Claude Code Club.

Atomic Insights

Lines worth screenshotting.

Asking an AI for a finished video in one shot fails because the job is too big, so it makes bad choices. Six narrow jobs fix it.
A skill is a saved set of instructions for one job, and it gets better each time you feed it your notes.
Whisper is free and gives a timestamp for every word, which lets Claude keep your last clean take of each line and cut the rest.
Graphics made without a plan look like PowerPoint slides because nobody told the model what good looks like. Show it screenshots of editors you love.
A beat map lists every idea in the script and what is on screen for it. Changing a plan costs almost nothing, changing a finished graphic costs a lot.
One graphic per idea that stays up long enough to read beat a graphic for nearly every sentence, which viewers called too much.
Split a long video into sections and give each to its own subagent so many graphics get built at once.
Use the stronger, pricier models for planning and creative direction and the cheaper, faster model for the building.
Nine subagents rendering at once maxed out the machine and made them cancel each other's work. Capping at four and rendering overnight fixed it.
Render drafts at half size while giving notes. Only render full quality after the whole cut is approved.
Make Claude check 10 frames per second for the specific mistakes you always catch yourself, like a graphic sitting on your face or a doubled word.
After each round of notes, Claude finds the instruction that should have prevented the mistake and rewrites it so the next video does not repeat it.
No single prompt made the video. About five months of accumulated notes did.
The second AI video used under half the tokens and half the hours of the first (1.4 billion tokens and 12 hours) because graphics and tools were reused.
Building a library of reusable assets as you go cuts cost and tokens on every later video.

Takeaway

Split the job, save the lessons

WHAT TO LEARN

AI output stops looking generic when you give it narrow jobs, plan before building, make it check its own work, and turn every note into a better instruction.

Break a big creative task into separate jobs instead of asking for the finished result in one shot, which pushes the model into bad choices.
Show the model examples of what good looks like, such as screenshots of work you admire, before asking it to design anything.
Plan first with a beat map, because changing a plan costs almost nothing and changing a finished graphic costs a lot.
Put the expensive models on planning and creative direction and the cheaper, faster models on building, and cap parallel renders so your machine survives.
Draft at half size while giving notes, and render at full quality only after the whole cut is approved.
Give the model a checklist of the specific mistakes you always catch, and have it check like a person would.
Ask which instruction would have prevented each note, so the fix goes into the skill and the mistake does not return.
Reuse graphics and tools across videos: the second edit used under half the tokens and half the hours of the first.

Glossary

Terms worth knowing.

Skill: A saved set of instructions that teaches Claude how to do one specific job. It can be updated after each use so the job improves over time.
Subagent: Another copy of Claude given one narrow task. Several subagents can work in parallel, such as building different sections of a video.
Beat map: A list of every idea in the script and what appears on screen while it is spoken. It is built before any graphics so changes stay cheap.
HyperFrames: A free tool from HeyGen that turns a web page into real video, so graphics and word-by-word captions can be built with code.
Whisper: A free speech-to-text tool that produces a timestamp for every word, which lets software cut footage by what was said.
FFmpeg: A free command-line program for cutting, joining, and converting audio and video files.
Render: The step where code or an edit is converted into an actual video file. It is heavy on the computer and slows other work.

Resources

Things they pointed at.

03:27linkPrompt Motion ↗

04:36toolHeyGen HyperFrames

02:40toolWhisper

02:40toolFFmpeg

08:48productClaude Code Club ↗

Quotables

Lines you could clip.

01:19

“I stopped trying to one-shot everything. It was too much for Claude to handle, so it just made bad choices.”

Clear diagnosis of why AI edits look cheap, no setup needed→ TikTok hook↗ Tweet quote

06:36

“No single prompt made this video. About five months of notes did.”

Short, contrarian, reframes where the quality comes from→ newsletter pull-quote↗ Tweet quote

03:38

“Changing a plan costs almost nothing and changing a finished graphic costs a lot.”

Transferable rule for any build process→ 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.

Claude Code edited this entire video and people are racking up tons of engagement across social media with systems like these. The captions, the motion graphics, the cards flying out from behind me, I didn't touch an editing timeline once. I was a video editor for 15 years for brands like Apple, PlayStation, and Nissan, so I know what a cheap AI edit looks like and this isn't one.

So here's the part that you can copy today. I stopped asking Claude for a finished video and I gave it six jobs, one at a time. In this video, I'll walk you through all six, step by step, what to tell Claude, what to check, and what to do in your own project.

And at the end, I'll show you exactly what this edit costs. Now, is it perfect? No.

A great human editor would still beat it on some shots and you'll need to give notes and try again like the same way you would with a person. But this is the worst it will ever be. Every video you make teaches the system something, so the next one comes out better, faster, and cheaper.

So here's why it takes six steps. When I first asked AI to edit one of my videos, I could tell right away that it just looked cheap. If you've ever tried it, you know exactly what I mean.

Most AI edits have that same generic AI look, and 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 trying to one -shot everything. It was too much for Cloud to handle, so it just made bad choices. Instead, I split the work into six jobs.

Scripting, editing, design, motion graphics, review, and feedback. Then I saved all six into a skill. A skill is a safe set of instructions that teaches Cloud how to do one job, and mine gets better every time I use it.

If you want to get access to the Cloud Editor Starter Kit, along with all my other skills and resources, already done for you, the link is in the description. A cool video won't perform well if the script sucks.

So have Claude study the best videos in your topic and pull out the hooks and formats that they use. Then save those in a library that you can use over and over again. So try this.

Ask Claude for the hooks and format from three videos in your niche that beat their channel's usual views. Save them in a library and then have it write your scripts the same way. When the script is done, you read it to the camera, which is exactly what I'm doing right now.

Next comes the footage and raw footage is a mess. You trip over a word restart a sentence or sit in silence. So give Claude your raw footage and ask it to transcribe it using whisper.

Claude will download whisper from the internet on its own and it's free to use. It's going to create a timestamp for every single word that you said. And now Claude understands how to do the edit is going to keep your last clean take of every line and cut the rest using another free program called FFmpeg.

So try this film one short take with a few mistakes on purpose, give quad the file and ask it to transcribe it with whisper, cut the retakes and list everything it removed. But a clean take is still just me talking.

So next is the fun part. Step three is design. If you ask for graphics with no plan, you get boring PowerPoint slides.

Claude does that because nobody told it what good looks like. So pick videos from editors you love and tell Claude to take screenshots of them. Then ask it what makes them work.

It will notice things like how text moves in and how one scene transitions to the next. A really good place to find examples is a site called promptmotion .com. It's full of motion graphics people made with Claude, and each one actually shows the exact prompt behind it.

Once you have the design, Claude now needs a game plan called a beat map. a beat map is just a list of every idea in your script and what's on screen while you say it so make claude build that map before it builds any graphics because changing a plan costs almost nothing and changing a finished graphic costs a lot i actually recently added a new rule that came from the comments on my last video before nearly every sentence got its own graphic and many of you said it was just too much.

So now instead each idea gets one graphic and it stays up long enough to read. So try this. Take screenshots of the things you love and either have Claude write design guidelines from them or have it mimic what it sees.

You can also give Claude a screen recording or a video file. Once the plan exists, the work piles up fast. A long video can have dozens of graphics, so I split the video into sections and I tell Claude to give each section to its own subagent.

A subagent is just a helper, which is another copy of Claude, but it has one narrow job. This lets Claude work on many things at once. They don't open any editing software.

Each graphic is built with code using a free tool from HeyGen called HyperFrames, which turns a web page into real video. That's also how you get captions that light up word by word. I put the stronger, more expensive models like Clod, Opus and Fable on the thinking.

So the planning and the creative direction, like all the design choices, the cheaper, faster models like Sonnet do all the building. There is one big thing to watch out for and it almost crashed my computer. So a render is the moment that code becomes actual video and it can really slow your computer down.

I once let nine sub agents render at the same time and they maxed out my machine and started canceling each other's work. Now the system only renders four at a time and I also render overnight so my machine isn't bogged down while I'm trying to work. I also like to keep my drafts small, so while I'm giving notes, Claude will render everything at half size so the drafts come out much faster.

Full quality renders only happen after I approve the whole cut. So try this, ask Claude to build one graphic as a webpage first, then turn the rest into a storyboard. Look at it, fix it, and only then turn it into video.

This will save you a ton on time and tokens in the long run. Step five is arguably the most important part, have Claude check its own work. So here's how I do it.

Claude checks 10 frames every second, and there are six mistakes that it looks for. An animation that plays twice, a graphic that flashes for a split second, text that lands too late to read, a graphic sitting on my face, a word I said two times in a row, and my lips drifting away from my voice. Make Claude check like a person, not like a computer.

So try this. Write down the three mistakes you always catch when you watch your own videos and hand that checklist to Claude to watch out for. The last thing that Claude does is save all of my feedback and improve.

the skill with each round. So every time you watch a cut, you're going to have notes, the same ones that you would give to a human editor. Quad fixes the video and then it goes one step further.

It finds the instruction that should have prevented the mistake and it rewrites it. So future videos won't have the same problem. No single prompt made this video.

About five months of notes did. So try this. After your next cut, give Claude your notes and ask which instruction would have prevented each one.

That's the whole system. Scripting, editing, design, motion graphics, review, and feedback. I still film every video and I make the final call because your choices and your taste is what's going to set you apart.

But I never have to open an editing timeline or After Effects again, which feels crazy to say after 15 years. so what did this actually cost well quad used about 600 million tokens roughly 100 million of those went to opus which did the planning and the design direction and about 500 million went to sonnet which did all the building at api prices that works out to about 160 dollars but i'm on the quad max 20x plan so for me it's just part of the monthly subscription so from the first bit of research to the final video quad worked for about six hours it doesn't count the time i spent watching cuts and giving notes my first ai video took about 1 .4 billion tokens and about 12 hours.

So this video used less than half the tokens and about half the hours because I reused a lot of the graphics, the tools, and the skill from the first one. So here's a tip, build a library of reusable assets as you go. It will cut down on your costs and tokens on every video after that.

So compare this with doing video editing the traditional way. Someone has to write a script, map every beat, create a storyboard, do designs, animate each graphic, mix the audio and add sound effects.

When I did that for a living, a video like this would take three days to a week and I was pretty fast. If you're a professional video editor, you'd still probably do some of these shots better and I would too, but this is the worst it will ever be. And it's already good enough that I keep handing more and more of my editing over to AI.

Honestly, it's fun and I like a new challenge. If you want to try it for yourself, the starter version of my editing skill is inside the Cloud Code Club. It does the same jobs, but in a simpler way.

It cleans up your recording, adds word -by -word captions, builds simple graphics, and checks its own work. It won't look like this on day one because mine took five months of notes, but it keeps your notes the same way so it improves each time you use it, and it's all tailored to your brand and your taste. if you want to master cloud code in 45 days and get the starter kit plus all my skills and resources already done for you come join over 8 000 builders inside the cloud code club it's only nine dollars a month and the links in my description if this helped hit like and subscribe it really does help the channel a ton i'll see you in the next

The Hook

The bait, then the rug-pull.

The whole video is the proof: the captions, motion graphics, and flying cards were all built by Claude Code, and the host says he never touched an editing timeline. He then hands over the process as six separate jobs you can copy.

Frameworks