Big Idea

The argument in one line.

Claude can fully edit a real CapCut project, syncing images to a voiceover and adding music, sound effects, text and keyframed zooms, by writing CapCut's project file directly instead of clicking through the app.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You run a faceless YouTube channel or content pipeline and want to cut manual editing time out of the loop.
You already use CapCut and Claude's desktop app and want a concrete, repeatable workflow for connecting the two.
You're open to connecting an MCP tool like Higgsfield to Claude for voiceover and image generation.

SKIP IF…

You only edit on CapCut Web, since this workflow needs local file access that CapCut Web doesn't expose.
You want full hands-on creative control over every cut, since this method has Claude make the first full pass.

TL;DR

The full version, fast.

Claude can edit a finished CapCut timeline by writing the project's draft file directly instead of clicking through the UI, which means a single prompt (sync images to the voiceover, add sound effects, music and text, fit the frame) produces a fully keyframed, scored edit that's still a normal, hand-editable CapCut project afterward. The full repeatable pipeline stacks three roles: Higgsfield generates the script's voiceover, images and character sheets; Claude orchestrates the whole thing and writes the timeline; CapCut stays the real, tweakable editor. A 3-minute video built this way ran about $9.50 in generation credits, with editing itself costing nothing.

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 – 00:32

01 · Claude edited this entire CapCut video

Cold open showing the finished stick-figure "Emu War" video already assembled inside a real CapCut timeline, with the claim that Claude built it without the creator touching the timeline.

00:32 – 00:51

02 · Works on the free version of Claude

Loui clarifies the method works on Claude's free tier, though he personally uses a paid Opus 5.5 plan.

00:51 – 01:26

03 · Set up your CapCut project

Open CapCut, create a new project, drag in the voiceover and 63 reference images, and rename the project so Claude can find it by name.

01:26 – 01:59

04 · The Claude prompt that edits CapCut

With CapCut closed, he gives Claude a short prompt naming the project and asking it to sync images to the voiceover, add sound effects, music and text.

01:59 – 02:39

05 · What you need installed + file access

Two hard requirements: CapCut desktop (not CapCut Web) and Claude's desktop app, since Claude needs direct file access to the local CapCut project folder; Claude requests folder permissions and he grants them.

02:39 – 03:22

06 · Claude's finished CapCut edit

Reopening the project in CapCut shows a fully built timeline: keyframed zooms on each clip, placed sound effects, an auto-found music track and synced text, playable with no manual edits.

03:22 – 03:55

07 · The stack: Higgsfield, Claude and CapCut

He names the three-part system for repeatable automated videos: Higgsfield as media generator, Claude as the engine, CapCut as the editor.

03:55 – 04:18

08 · Write a faceless YouTube script with Claude

A fresh chat gets a simple prompt for a 3-minute script on any interesting topic, destined to become a stick-figure explainer video.

04:18 – 05:18

09 · Connect Higgsfield to Claude (MCP setup)

Walkthrough of pulling an MCP server URL from a Higgsfield account and adding it as a custom connector in Claude's settings, then authorizing access in the browser.

05:18 – 05:53

10 · Generate the AI voiceover

A single prompt tells Claude to generate the voiceover for the script via Higgsfield; Claude returns multiple voice options to preview and download.

05:53 – 06:06

11 · Timestamp the script to the voiceover

A follow-up prompt asks Claude to timestamp the script precisely to the voiceover so the later image generation can sync scene by scene.

06:06 – 07:08

12 · Generate the character sheet and images

Claude is told to generate all images from the timestamped script via Higgsfield, but first produce a character reference sheet for approval; he picks one of two options before Claude generates the full image set.

07:08 – 07:58

13 · Let Claude edit the CapCut timeline

With voiceover and images downloaded, he repeats the CapCut setup (new project, rename, import assets) and gives Claude the same editing prompt for the second video.

07:58 – 08:29

14 · Tweak, reassess and create thumbnails

He opens the finished timeline to make manual tweaks where wanted, tells Claude to "reassess" the project so it learns from the edits, then asks for three thumbnail options.

08:29 – 09:08

15 · Free Claude skill to automate everything

He packages the entire workflow, script to thumbnails, into a downloadable Claude skill so viewers can run the whole pipeline by approving a script idea.

09:08 – 10:01

16 · How much this video cost to make

Breaking down Higgsfield credit usage (46 images, 3 thumbnails, 2 character sheets = 241 credits at 3.9 cents each) for a total around $9.50, with a typical range of $5 to $9.50 for similarly produced videos.

Atomic Insights

Lines worth screenshotting.

A CapCut project is just a draft file on disk, so an agent with file access can write a finished timeline without ever opening the CapCut UI.
The instruction 'do not use computer use or open CapCut, build the timeline by writing the project files directly' is what keeps an agent off GUI automation and routes it to direct file edits.
Giving an agent a workable CapCut-editing prompt only takes four parts: which project, what to sync, what to add, and a framing constraint.
CapCut Web has no local project file an agent can write to, so this workflow only works with CapCut's desktop app.
Generating one character reference sheet first, and approving it before any scene images, is what keeps a character's look consistent across dozens of separately generated shots.
Timestamping a script against the actual recorded voiceover, not the script's estimated pacing, is what makes later image-to-audio sync accurate.
Because the AI-edited output is a real CapCut timeline and not a locked render, every clip, keyframe or text layer can still be adjusted by hand afterward.
Asking an agent to 'reassess' a project after you've made manual tweaks is how it picks up your specific preferences for the next edit.
The underlying pattern is three stacked roles, not one tool: a media generator for voice and images, an LLM as the orchestrating engine, and a real editing app as the assembly surface.
A 3-minute AI-generated video cost about $9.50 in Higgsfield credits (241 credits: 46 images, 3 thumbnails, 2 character sheets at 3.9 cents each), with CapCut's editing step costing nothing.
Generation cost scales with image count, not editing time, so a longer video gets expensive by needing more generated shots, not by needing more AI editing.

Takeaway

How to let an AI agent actually edit a real CapCut timeline

WHAT TO LEARN

Giving an agent direct file access to a CapCut project, instead of driving the UI, turns a short four-part prompt into a fully keyframed, synced timeline you can still hand-edit afterward.

01Claude edited this entire CapCut video

A CapCut project is just a draft file on disk, so an agent with file access can write a finished timeline, including clip order, keyframes and text, without ever opening the CapCut UI.
The result stays a normal, fully editable CapCut timeline, not a locked AI render, so anything the agent gets wrong is still a two-click manual fix.

03Set up your CapCut project

Before handing a project to an AI editor, do the manual prep yourself: drag in every asset and give the project a short, specific name so the agent can find it by name alone.
Renaming a messy default project to something like a short working title removes the single biggest source of an agent targeting the wrong file.

04The Claude prompt that edits CapCut

A workable CapCut-editing instruction only needs four parts: which project, what to sync, what to add, and a framing constraint.
Stating the sync relationship explicitly, images synced to the voiceover, gives the agent a single source of truth for timing instead of making it guess pacing.

05What you need installed + file access

This workflow only runs with the desktop app on both ends: CapCut Web has no local project file for an agent to write to, and the chat client needs local file access, not just a browser tab.
The instruction to skip computer-use and write the project files directly is what keeps the agent off GUI automation and routes it to faster, less fragile direct file edits.
When an agent first touches a folder on your computer, it asks permission per folder, so route it only to the specific project and asset folders it needs.

06Claude's finished CapCut edit

A finished agent edit is worth checking against concrete signals: are the keyframed zooms actually applied, is the placed music relevant, does the text match what you asked for.
Letting the agent find its own music from a pre-stocked folder only works as well as the folder is organized, since its picks are bounded by what's already there.
Because the output is a real timeline, an AI-edited result isn't an unchangeable one. Any clip, keyframe or text layer can still be nudged by hand afterward.

07The stack: Higgsfield, Claude and CapCut

The repeatable version of this isn't one tool, it's three roles stacked together: a media generator for voice and images, an LLM as the orchestrating engine, and a real editing app as the final assembly surface.
Swapping which tool plays generator or editor doesn't change the pattern. Any image or voice generator plus a file-capable agent plus a native-project-file editor reproduces the same pipeline.
Keeping the editor a real app with a draft file, not a proprietary render, is what preserves manual override at every stage, which is the actual advantage over a pure text-to-video generator.

08Write a faceless YouTube script with Claude

A single open-ended prompt, topic and target length, is enough to seed the entire downstream pipeline, since every later step derives from this one script.
Specifying the eventual visual treatment up front lets the model write pacing and scene breaks that already anticipate the image generation step.

09Connect Higgsfield to Claude (MCP setup)

Connecting a third-party tool as a custom MCP connector is a one-time setup that then exposes that tool's actions as things you can simply ask for in plain language.
Once a tool is connected this way, the chat interface itself becomes the control surface, no separate app to learn, just new verbs available in the same conversation.

10Generate the AI voiceover

A plain instruction to generate the voiceover is enough once the tool is connected; the service returns multiple voice options to preview before you commit to one.
Previewing and picking from several generated voice options before downloading avoids re-running the entire downstream pipeline over a voice you don't like.

11Timestamp the script to the voiceover

Timestamping a script against the actual voiceover, not the script's estimated pacing, is the step that makes later image-to-audio sync accurate, since spoken pacing never matches a silent read.
This step turns a flat script into a structured scene list with real start times, which is the input format the image generation step actually needs.

12Generate the character sheet and images

Generating one character reference sheet first, and approving it before any scene images, is what keeps a character's appearance consistent across dozens of separately generated shots.
Choosing between a couple of character sheet options up front is cheaper than discovering a consistency problem dozens of images in.

13Let Claude edit the CapCut timeline

The same CapCut setup steps and the same editing prompt reproduce the whole process on a second video, which is the actual proof the workflow is repeatable and not a cherry-picked demo.
Reusing an identical prompt template across projects is what separates a real pipeline from a one-off trick tied to a single file.

14Tweak, reassess and create thumbnails

Manually tweaking a few things, then explicitly asking the agent to reassess the project, is what lets it pick up your specific preferences for the next run.
Thumbnail generation can ride on the same session once the edit is approved, since the agent already has full context on the subject matter.

15Free Claude skill to automate everything

Packaging every step into one reusable skill collapses a multi-step manual process into a single approve-the-idea decision point.
The only manual step left in a fully packaged version of this pipeline is importing the final footage into CapCut; everything upstream of that is automated.

16How much this video cost to make

241 Higgsfield credits, 46 images, 3 thumbnails and 2 character sheets, at 3.9 cents each landed this specific 3-minute video around $9.50, with a typical range of roughly $5 to $9.50.
Cost scales with image count, not the editing step itself, so a longer video gets expensive by needing more generated shots while CapCut assembly stays free.
For anyone budgeting a channel on this pipeline, per-video generation cost, not the AI editing, is the real recurring expense to plan around.

Glossary

Terms worth knowing.

MCP connector: A Model Context Protocol server that a chat client like Claude connects to as a custom connector, exposing a third-party service's actions (like generating a voiceover) as things you can ask for directly in chat.
CapCut draft file: The project file CapCut saves locally on disk, holding the full timeline, clip positions and keyframes. Writing to this file directly is what lets an agent build an edit without opening the CapCut app.
Character reference sheet: A single image showing every character in a planned video side by side in a neutral pose, generated and approved first so every later scene image stays visually consistent.
Higgsfield credit: The unit Higgsfield bills generation by. Each image, thumbnail or character sheet uses one credit, and a paid plan bundles a monthly allotment at a fixed per-credit rate.
Keyframed zoom: A subtle in-and-out zoom applied to a still image over time by setting different zoom levels at different points on the clip, giving a static image a sense of motion.

Resources

Things they pointed at.

00:00toolHiggsfield ↗

08:39toolFree Claude skill (full pipeline automation) ↗

01:42linkFull Claude prompts list ↗

00:51toolCapCut

Quotables

Lines you could clip.

00:00

“This entire CapCut video was edited by Claude. The voiceover synced to the images, the music, the sound effects, the text, and I didn't touch the timeline once.”

the entire premise in one cold-open line→ TikTok hook↗ Tweet quote

02:22

“Do not use computer use or open CapCut. Build the timeline by writing the project files directly through file access to my Mac.”

the literal reusable instruction viewers will copy-paste→ newsletter pull-quote↗ Tweet quote

09:31

“This project cost around nine and a half dollars to create.”

a concrete, repeatable cost number→ 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

This entire CapCut video was edited by Claude. The voiceover sung to the images, the music, the sound effects, the text, and I didn't touch the timeline once. What's cool is that this isn't just an AI video that you can't change.

It's built in a real CapCut timeline, meaning that we can make any tweak we want. I'll show you exactly how to automate your CapCut editing step by step. I've also made a free Claude skill for you that you can find in the description that'll run this entire process from the scripts to the thumbnails of your video.

So here is the finished CapCut timeline that Claude edited entirely. Let me show you exactly how I got here. What's great is that this works on the free version of Claude, so you don't need a subscription.

I do have a subscription, so I'm gonna use Opus 5 .5, which I absolutely love. But again, you can do this on the free version. So all we need to do is open CapCut and create a new project.

I'm gonna go ahead and drag my voiceover into my media bin. And then I've created 63 frames or images that I want in my video. So I'm gonna copy all of them and paste them into that media bin.

You can import whatever media you have. What we also need to do is change our project name to something that we can tell Claude.

So I'm going to call it Emu War 2 and go ahead and click enter. Then what we can do is go ahead and close this CapCut project. Make sure that in your projects it's saved as Emu War 2 and then let's close CapCut entirely.

Then inside of a brand new chat window I'm going to add my prompt telling Claude that I want you to edit my CapCut project with that project name Emu War 2. Sync the images to the voiceover, also add sound effects, add music and relevant text, making sure it fits the frame perfectly. These last two lines are important or you're actually going to see CapCut doing the manual editing.

Do not use computer use or open CapCut. Build the timeline by writing the project files directly through file access to my Mac. Now I'm going to have a full prompts list that you can copy and paste into your Claude.

There's a link in the description for this. Two things. You actually need CapCut installed on your computer.

You can't have or be using CapCut web. We also need Claude installed on our computer so that it can access the whole timeline and the actual project file. that is stored on our computer from CapCut.

If you don't have either installed on your computer, this process won't work. When you run this process in a new chat, Claude is gonna ask for permission to access our timeline. All we need to do is click Allow Once.

Right here, Claude is also asking for two folders on my computer. I actually have a Creatives folder with music and sound effects that Claude wants to access, so I've just given it permission to access that Creative folder too. I'm gonna give Claude a couple of minutes to edit this project, and I will pick back up once it is done.

We're going to go ahead and reopen CapCut. And let's find that project Emu War 2. And you can see that this is the fully edited timeline.

It's got all the sound effects. If we click on a clip, it's applied keyframes so that we get zoom on our clips. And it's added in text files.

It's also found a music file directly from our computer. And if we click play... In 1932, Australia sent soldiers with machine guns to fight an army of 20 ,000.

A month later, the soldiers went home. The enemy? Emus.

This is the great emu war. We haven't tweaked anything, but if we wanted to tweak anything, maybe there was a text file that was misplaced. Once again, we have that full timeline capability to go in and do some final tweaks to customize this.

So you've seen the capability that Claude has to edit a project in CapCut. But how do we set up this whole entire process for automated videos to create videos like the one you just saw? Well, the first thing we need is our media generator.

So we're going to be using Higgs field. The next thing is that we need Claude, which is our engine. And the last thing is we need our editing platform, which is CapCut.

I've showed you the first two steps. Let me show you how to run this entire process so that you can make a video like this using Claude. We'll be using Higgs field to generate our voice and images, and I'll show you exactly how much the project cost at the end of this video.

What we're going to do is start a new cloud chat and put in a very basic prompt for a YouTube script right here. I said, write me a three minute script for YouTube topic. Anything interesting.

I will later animate the script into a stick man video. Now you don't necessarily have to create a stick man video. You can create whatever type of video or have whatever type of subject you want in the video.

And I've just specified that we're going for faceless YouTube automation. While our script is being written, we wanna connect Higgs Field directly inside of Claude. This gives Claude the capability to generate images on our computer.

If you don't have a Higgs Field account, there is a link in the description where you can sign up for Higgs Field. Once you do have a Higgs Field account, log into higgsfield .ai and go to MCP. Then we can copy this link right here.

Let's go ahead and copy that and go back to Claude. In our sidebar, we're gonna go down to the bottom left, click Settings, go to Connectors, Add, Add Custom Connector.

We can change the name to Higgs Field. And in the MCP server URL, let's go ahead and paste that link that we copied. Click Continue.

Under Authentication, go ahead and click Sign In Now. You can leave this as normal and select Add. Then once you see this, go ahead and click Connect.

It's going to redirect you to your browser where we can finish connecting. We're going to give Claude consent to access Higgs Field by clicking Allow. And once we see this final prompt, we can go ahead and open Claude.

You'll see in the bottom right that we have now connected Claude to Higgs Field. Claude has generated our Stickman script, which we are now going to turn into a voiceover. In our prompt bar, I'm going to write, generate the voiceover for the script using Higgs Field.

And we hit send. If this is the first time that you're using Higgs Field and Claude, Claude will give you the full set of voiceovers that you can choose from. Once you find one that you like, just select that and it'll generate your voiceover using that option.

in a few minutes claude has generated our voiceovers we can go ahead and click one and preview in 1932 australia sent soldiers with machine guns to fight a war against once you're happy with your voiceover go ahead and click download the next thing that we need to do is actually generate a timestamp script according to this voiceover so i'm going to write another prompt please timestamp exactly according to the voiceover so we can generate visuals exactly according to the timestamp Once your timestamp script is done, we're going to put in a new prompt.

This is where we finally get to generate those images. I've put generate all the images for our videos using the timestamp so it can sync perfectly. But first generate a character reference sheet for approval.

Claude and Higgsfield generated two character sheets for us. The first one looking like that and the second option that we can choose from. If we scroll down a bit, you can see that sheet A is Claude's pick.

So I'm going to say sheet A is approved. Proceed to the other images. Once we send this prompt, Claude and Higgsfield are now going to generate all the final images for our entire video.

Once again, this doesn't necessarily have to be Stickmin. You can create whatever type of character you want. And if you don't want to use images, just ask Claude to generate videos for your final edit.

Once our images are done generating inside of Claude, it'll give you a zip file that you can download to your computer. Claude has also generated a preview for us. We can see the entire video, but again, this is where CapCut steps up.

We don't have any ability to customize this. So what we're going to do is go ahead and open our CapCut. This is where it gets crazy.

We're going to make that new project. We're going to go to our downloads folder, import our voiceover and import all the shots that it created. Let's change the name of this project, Emu War 3, and click enter.

We can go back to that homepage and make sure that Emu War 3 is saved. That is it for our involvement with CapCut. Go ahead and close it.

Inside of that same clawed chat, I've said edit Emu War 3 perfectly syncing the visuals to the voiceover. Add subtle zooms in and out using keyframes to each of the images. Super subtle, it'll add a little bit of motion.

Add sound effects and relevant music, and add any text you deem fitting, but make sure it fits in the frame. Do not use computer use or open CapCut. Build the timeline by writing project files directly through the access to my Mac.

You can go ahead and click send. Let's wait and we can go ahead and open that project once it's done. We have the complete timeline if we wanted to make any changes.

Maybe a text file wasn't sized correctly or we didn't like the music. We can make any change. after making those changes close that cap cut project and ask claude to reassess that exact project it'll learn from all the tweaks that you made so that your next project will be even better than the one it edited once i was happy with the edit i asked claude for three thumbnails based off the subject matter of our video and it went ahead and made those thumbnails for us that we can just use directly in the youtube upload now i've created a skill for you that will automate this entire process all you need to do is click approve on the script idea you can go ahead and download that skill in the description once you have it inside your cloud you just go to your profile at the bottom left select settings go skills and then you can add that skill by clicking upload skill once again that's going to enable your claude by itself to write the script generate the voiceover do the timestamp script generate that character sheet generate all the images it'll understand the cap cut briefing on how you like to edit your project so all you need to do is import that footage into cap cut and claude will do the rest so how much did this three minute video actually cost to create

While we use 241 credits for the 46 images, the three thumbnails, and those two character reference sheets. On Higgs field for $39 a month, you get a thousand credits, which works out to 3 .9 cents per credit. If we take 3 .9 cents and times it by the 241 credits, this project cost around nine and a half dollars to create.

Now, this is more on the expensive side of what I've run as I've done this exercise a couple of times. And that range generally goes from about five dollars to this case, which was the highest 9 .5. So you can expect the range of the project to be around there if you're creating a three minute video.

Obviously, the longer the video is, the more expensive it's going to be as it has to generate a lot more images. And as you've seen, you can now create a faceless YouTube channel with no editing using Higgs field, Claude and CapCut. I'll see you in the next video.

The Hook

The bait, then the rug-pull.

Matt Loui opens by playing a finished stick-figure video, The Great Emu War, that Claude built inside a real CapCut project: every clip synced to the voiceover, zoomed with keyframes, scored with music and sound effects, and still fully editable afterward because nothing about the timeline is locked.

Frameworks

Named ideas worth stealing.

03:22model

The Higgsfield → Claude → CapCut stack

Higgsfield (media generator)
Claude (engine)
CapCut (editor)

The repeatable three-role pipeline for producing a whole automated video with no manual cutting: one tool generates voice and images, an LLM orchestrates and writes the edit, a real editing app stays the final, tweakable assembly surface.

Steal forany automated content pipeline that needs a generator, an orchestrating agent, and a real editable timeline instead of a locked render

01:26list

The CapCut editing prompt template

Name the exact project
Sync the images to the voiceover
Add sound effects
Add music and relevant text, make sure it fits the frame
Do not use computer use or open CapCut
Build the timeline by writing the project files directly through file access

The exact instruction shape that gets Claude to write a CapCut draft file directly instead of clicking through the UI, repeated verbatim across two different videos in the tutorial.

Steal forany prompt asking an agent to edit a native app's project file instead of driving the UI through computer-use

CTA Breakdown

How they asked for the click.

VERBAL ASK

08:29product

“I've created a skill for you that will automate this entire process. All you need to do is click approve on the script idea.”

A soft version lands at 00:19 ("I've also made a free Claude skill for you"), then the full explanation and download pointer comes back at 8:29 once the whole manual pipeline has been demonstrated, so the CTA lands after the value, not before it.

MENTIONED ON CAMERA

00:00toolHiggsfield ↗

08:39toolFree Claude skill (full pipeline automation) ↗

01:42linkFull Claude prompts list ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.