Claude Just Became a Video Editor: Higgsfield Katana + Claude Motion
Read the Portuguese translation →
Author: Prompt Engineer 48 · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Claude Just Became a Video Editor: Higgsfield Katana + Claude Motion · Channel: Prompt Engineer 48 · 09:56 · Watch on YouTube
Topics: AI agents, Claude, AI content generation
Summary
Section titled “Summary”The video explains how Claude and Higgsfield Katana can recreate an existing video edit shot-for-shot using a reference and the user’s own photos, products, or footage. It walks through connecting Katana, choosing a community preset or uploading a permitted reference, and explains how the tool routes each cut between existing pixels, code, user footage, and generated shots. The second half demonstrates Claude Motion, which creates editable titles, cards, and other graphics through code rather than generated video footage. The presenter also notes that Motion is in beta for team and enterprise plans, generated shots use credits, and Katana references must be owned or used with permission.
Key points
Section titled “Key points”- Katana works inside Claude through the Higgsfield MCP connector or via the Katana page.
- Users can start from a community preset or provide their own 3–30 second reference edit.
- A community preset can use up to five photos; a custom reference workflow can use up to 40 images.
- Katana analyzes the reference and rebuilds its cuts, motion, color, titles, and audio, changing only what the user requests.
- Katana routes each cut through the least costly suitable method: retaining source pixels, using code, using the user’s footage, or generating a missing shot.
- Code-based work and cuts using the user’s footage cost zero credits; generated shots use credits.
- Katana checks the result against the source, including text spelling, frame count, and audio alignment.
- Claude Motion creates editable design graphics and animations with code, not a video generation model.
- Claude Motion is described as a beta for team and enterprise plans, so users should check whether they have access.
- Katana copies audio, so the reference must be the user’s own edit or one they have explicit permission to use.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:01 | Katana and the shot-for-shot editing idea | The presenter introduces Katana, explains how it uses a reference edit to rebuild a video, and outlines its two ways in: community presets and custom references. |
| 02:07 | Installing Katana and using presets | The video shows the one-time MCP connection, the community preset workflow, and examples of completed edits and their underlying scripts. |
| 04:12 | How Katana builds and checks an edit | The presenter examines edit scripts and cost-saving routing, describes result checks, and explains how to create and share edits from a permitted reference. |
| 06:16 | Claude Motion examples | The video introduces Motion as a code-based tool for editable design animations and showcases examples including product graphics, typography, and animated data. |
| 08:53 | Limitations and final advice | The presenter notes Motion’s beta availability, clarifies Katana’s separate operation and credit costs, and reiterates the reference-permission requirement. |
Quotes
Section titled “Quotes”It matches the cuts, the motion, the color, the titles, and the audio of the original — 01:03
nothing heavy runs on your laptop. — 02:38
Each cut takes the cheapest route that still matches the reference. — 04:43
Claude writes code that animates your content, so you can change any word, number, or timing. — 06:47
The last one is the reference. It has to be yours or you need permission because Katana copies the audio. — 09:24
Concepts and entities
Section titled “Concepts and entities”Claude, Higgsfield Katana, Claude Motion, Higgsfield MCP, Higgsfield, Anthropic
Full transcript (original language)
Section titled “Full transcript (original language)”00:01 All right, in this video we’ll see how Claude and Hicksfield Katana rebuild and edit shot-for-shot and you can see that on the right of the screen, the sneaker ad, the skater in the green puffer jacket, the red car in the desert. Now, mind you, these are some of the demos from Hicksfield and we’ll go through them one by one as we go. And on the left of every one of them you can see code, that is Claude writing the edit line by line while the video plays. I’ll show you the install the preset flow and the own reference flow. And then the
00:31 second half of the video is about Claude motion, which is the thing that draws the titles and the cards. Let’s go ahead and break it down. So, this is the launch video Hicksfield put out for Katana and it’s 78 seconds long. The one line that matters is on the post. It’s their AI video editing tool and it lives inside Claude. Here is the idea, you take an edit that already works like an aura farming reel, an anime edit, an outfit check, or a travel montage. You give that to Katana as the
01:03 reference, you give it your own face, your product, or your clips, and it rebuilds the edit shot-for-shot. It matches the cuts, the motion, the color, the titles, and the audio of the original, that is the part I want you to hold on to because it only changes what you tell it to change. Look at the screen here. You can see the analyzing card on the photo of the two people. Before anything is generated, the reference gets read and the work is split between two sides. Claude is the operator, it reads the reference, plans
01:33 every cut, and runs the workflow. Hicksfield generates only the missing shots and puts the final video together in its own sandbox. You started by typing {slash} Katana in Claude through the Hicksfield MCP or you can start from the Katana page on hicksfield.ai. There are two ways in. One is a community preset, you pick a ready-made edit from a grid, add up to five photos, and recreate it. The other one is your own reference. Any edit between 3 and 30 seconds with up to 40 images of your characters, products, or clothes.
02:07 So, the install is one click on the Katana page at Hicksfield.ai/katana. There’s a button that says install Hicksfield in Cloud. It opens a connector dialog in Cloud and the address mcp.hicksfield.ai/mcp comes already filled in, so you don’t type anything. If you have already connected Hicksfield before, there is a second button that says continue I have Hicksfield MCP and you just use that one. This is a one-time step. After this, Hicksfield is a tool Cloud can call the same way it calls any other
02:38 connector. And what that gives you is this, the MCP is what lets Cloud do the generating and the compositing on Hicksfield’s side, so nothing heavy runs on your laptop. Now, the preset side, there’s a grid of community presets and the catalog has 19 aura farming presets like last Katana, power suit, outfit check, tiger eyes, dark aura, nocturne, car edit, and aura fast. You hover over a card and a recreate button shows up, you click that. Then it asks for your photos, up
03:10 to five, and there’s an optional prompt box where you can say what to change. And the last button says recreate in Cloud. That hands the whole job over to Cloud. That is the point where the job starts and credits get used. So, let me show you what a finished one looks like from the demos. So, here’s one, this is the skater edit, the line 9 clock at 2:47 a.m., the Ollie plus 130 pop-up, the tag plus 500 burst, and the rush neon wall at the end. And this one is the red car on the desert road with the title Noctis drawn over the sand at the
03:41 end. Both of these have the code of the edit running on the left, so you can read how it was cut. There are also clipping presets, travel added, many lies, and physical body, and those cut your own footage to the rhythm of a reference instead of generating new shots. This is the part I was most curious about, so let’s look at what Claude is actually doing. On the left of the sneaker edit, you can see a script. The top of it is a list of scenes, each one
04:12 with a start, a frame count, and a color like yellow, white, and red. Under that, the lines that call generate image with a prompt and little thumbnails of the images next to each line. Those thumbnails are the shots Katana had to make. And on the right, the finished shot plays. The laugh with the glasses on the yellow newsprint, the white horse jumping out of the red paper tear, the cardboard locker room with the pink sneaker character. Now, here’s the routing idea, and it saves money. Each cut takes the cheapest route that still
04:43 matches the reference. If the original pixels can stay, they stay. Flashes, type, grade, and camera moves are done in code. If you gave it your own footage, it uses that. It only generates a new shot where something is missing. And the routes that use code or your own footage cost zero credits. That’s from Hicksfield’s own description, and it’s why the status line with the cost matters when you run it for real. Look at this watch clip. It’s only 5 and 1/2 seconds. The code has a track function, so the dial stays locked while the wrist moves and turns, and a generate video
05:15 call at the bottom for the closing shot. On the car one, the code panel stacks a list of cuts with names like approach, chase, high wheel, and tail light, and the timeline strip under the video follows along with each one. Before it hands you anything, Katana compares the source and the results second by second. It checks text spelling, frame count, and audio alignment, and you get one finished MP4 a side-by-side with the reference, and a spend summary.
05:46 Now, the option that Hicksfield recommends, create from your own video. You click that, you upload a reference that is between 3 and 30 seconds, you add your images up to 40, and you can type a prompt. One rule, and it’s in their guidelines, the reference has to be your own edit or one you have explicit permission to use. Katana copies the audio and can keep the original footage, so choose the source carefully. So, keep that in mind before you pick a source and use your own edit. If you want to see how far one setup goes, you can run versions 1 to 4 from
06:16 the same reference with different people or products, and then make more of the one that works best. And if an edit turns out great, there is a share preset flow, upload the reference, add images, write a prompt using add image one style mentions, and hit publish preset. That makes it a community preset. Okay, now the second half. Katana makes the hero edit, Claude Motion makes the design graphics around it, the title cards, captions, specs, and calls to action. Anthropic’s own article
06:47 describes it, you type {slash} motion in the Claude message box, describe the animation, review it, adjust it in the editor, or ask Claude to change it, and download an MP4. And one line from that article explains everything you’re about to see. Claude writes code that animates your content, so you can change any word, number, or timing. It does not use a video generation model. So, no generated footage and no AI-generated people in that part, it’s text, shapes, charts, and images moved by code. Here is one from Higgs Field. On the right, you can
07:18 see a black and white reel, the ocean, then the horse, the wolf, the runner on the track, the lightning, the eagle, and the last scene says go deeper. Under the video, there are seven scene thumbnails, and the one that is playing gets a green box. On the right is the Python for the scenes with names like ignition, forward, and lift off. This one is a pre-order page for a pet robot called Blip. At the bottom, you can see the scene strip and the video goes from the two wide eyes opening to the Blip title to the big numbers one
07:50 and three to four little creatures to a clock dial and ends on pre-order Blip. This one is for sunglasses called Glare. The first frame says, “Don’t look.” Then the word glare repeats down the screen and then the blue lenses come in. And this is a keyboard heat map with the title some keys work hard. The backspace key has coffee cup, control CV are three keys holding paper and the last scene says press responsibly. This one is called Howl with a pink
08:21 waveform across a face and a decibel readout that goes up to 110. Then a cow, a big black and white cow title with the lettering warped over it. A reel for the number 16 with a red record and cut out photos stacking up. This is Strata with a product launch card that says Strata 2.0 is live.
08:53 And last, Vault, a pair of white gloves opening a metal tin with the line deserves a vault. Every one of these is made up branding for a demo, so there is no real company logo in any of it. Let me tell you the catches because I’m not going to oversell it. First one, cloud motion is a beta for team and enterprise plans, so check that you have access before you plan around it. Second, Katana does not need motion. Katana runs on its own and motion is the
09:24 extra layer if you have access. Third, the credits, the roads that use code or your own footage cost zero and generated shots are where the credits go. So, watch the cost status line when you run it. The last one is the reference. It has to be yours or you need permission because Katana copies the audio. So, in this video we have seen how Claude directs, Higgsfield Katana rebuilds the edit shot for shot, and Claude motion draws the graphics on top. If you understood find an edit you love
09:56 that you own, open Katana in Claude, and make it yours. The link is at the top of the description. Let me know in the comments which edit you would remake first. I’ll see you in another one. Check out the next video.