Guide

How to make videos with Claude

AI models have crossed a threshold: you can now make professional-looking videos without hiring a video editor or a motion designer. This is the exact process I use to make motion graphics, music videos, product launch videos and edited shorts with Claude, the tools behind it, and the prompts.

Examples of Claude motion graphics videos

Everything in these videos was made entirely with Claude, all the animations and all the cuts. The ones from the walkthrough:

A music video, Not Available in Your Region. It's one of the examples on the meme music video playbook.
A product launch video that presents one SaaS product with nothing but animations and some music in the background, the kind of video the product launch playbook makes. HeyGen alternatives for launch videos compares this approach with avatar video.
A vertical short. All that went in was a raw recording of me talking; Claude did the rest of the edit.

Coming from After Effects? Here's how motion graphics in code compare.

The tools behind AI motion graphics

Models are smart enough now that you can just describe what you want, and they figure out which tools they need. But if the result doesn't match what you expected, it's often because of the tools the model had. Then you need to direct it and give it the tools it needs. So it helps to know what's going on.

A video has lots of layers. In one shot of my music video, the background could come from an image generator, the caption is code that writes the text on top, the singing person needs a video model that lip syncs, and the box around them is code again. Then there's sound, which a model can make or code can write.

Driving all of it is an orchestrator agent. The best one for this right now is Opus 5.5. Sonnet 5.5 is slightly worse in quality but much cheaper, so it's a good alternative. The agent takes the actions, writes the code and stitches everything together, usually with FFmpeg, the Swiss Army knife of video. To do this well it needs these components:

Research: the agent's web search and browser to get screenshots of a product, the browser or the API to search X, and yt-dlp (free) to download or transcribe other videos.
Script, or lyrics for a music video: the agent writes these itself. It just needs your feedback.
Text-to-speech for a voiceover: ElevenLabs, with a clone of your own voice or one of its preset voices. There are open-source alternatives too.
Music: Suno V6 is the best right now, and it's what I used for my music video. ElevenLabs Music isn't bad either. For simple music without vocals, and for sound effects, the agent can write the music in Python.
Transcription, so the agent knows exactly when each word is said and can time things to it. It usually defaults to Whisper running locally (whisper.cpp), or give it an OpenAI API key to transcribe in the cloud.
Image generation: Claude outputs text and code, not pixels. Use GPT Image 2 through the OpenAI API, Higgsfield for access to a bunch of image models, Cursor's built-in image model, or Midjourney through the browser (it has no API).
Video generation, for segments with real people that code can't draw: Kie.ai is cheap and gives you Seedance 2.5. Higgsfield and Midjourney work too.
The video framework, which puts it all together. That's the most important part.

Video frameworks: HyperFrames, Remotion and Manim

Your agent collects all the assets and then has to turn them into an MP4. Every frame is an image: a hidden browser opens the page, freezes time and takes a screenshot, moves forward 1/30 of a second and takes another, thousands of times. Then FFmpeg stitches the screenshots and the soundtrack into one MP4. Frameworks help with that:

HyperFrames lets the agent write HTML that turns into an MP4. GSAP is a good tool alongside it for making the animations in JavaScript.
Remotion is the same idea, with React instead of HTML. See Remotion vs Hyperframes for how the two compare.
Manim is for math animations like the 3Blue1Brown videos. The math animation playbook makes that kind of video, and Remotion alternatives compares Manim with the other code-video tools.
Custom: if you just tell Claude to make a video, it often writes its own little engine for the screenshots instead of downloading a framework. Using an existing framework is usually more efficient.

You don't need the details, but knowing roughly what happens lets you direct the agent when the output isn't what you expected. Which components a video needs depends on the video. The product launch video needs research for the product's assets, a script or plan, music and the framework, but no voice, image or video generation. The short needs transcription, but no text-to-speech, since my voice is already in it. The music video is the most complex and needs pretty much everything.

Level 1: Reference and describe

First, make sure your agent has access to the tools above, and knows it has them. Depending on the video, you might need to create some accounts or have your agent install a tool first.

Then reference a video and describe yours. Find a video in the style you want, copy the link and paste it into your agent with a description. I use Cursor with Opus 5.5 because I have a lot of credits there (more on making videos in Cursor). Claude Code works the same, and so does the ChatGPT app if you want the ChatGPT models.

[link to the reference video] Create a video in this exact style, but instead I want the video to be about this little cloud creature that finds love, and it should just be 10 seconds long.

In the walkthrough, the agent fetched the post through the FxTwitter API, took screenshots of the video to see what the style looks like, picked HyperFrames because it's installed on my machine, made the audio and rendered a 10-second video. The story was right, but it kept the timeline background from the reference, because that was in its screenshots. So be precise about what to copy from the reference and what not to.

My music video process and prompts

For a more complex video like my music video, I went through it step by step:

Reference and research. My first prompt said to go through the same process as an earlier video of mine, gave it the topic (Europe falling behind on AI), and told it where to search for memes, including an account on X whose tweets were the direction I wanted.
Lyrics. For a music video you want the music first and build everything on top. Claude suggested topics, then wrote lyrics, and I told it what was good and what wasn't. It took about seven rounds.
Song. Once the lyrics were locked in, I told it to use Suno and the lyrics to make the song.

The whole process with every prompt is in a PDF in the free resources.

Turn your workflow into reusable skills

When you've been through a process and want to repeat it, turn it into a skill. Two of mine:

A short-form creator skill. It turns a raw talking-head recording into a fast-paced short by removing silences and earlier takes, matching supplied or sourced visuals to the narration, and adding lower-third captions. It made the short from the start. It even decided which screen recording it needed, recorded my screen typing into ChatGPT, and cut that footage in.
A WebGL music video skill, adapted from one someone shared on X. The elements move with the beat of the song, and that's defined in the skill.

Both are in the free resources, with other skills I've uploaded over time. In FrameJam, skills like these are playbooks. For shorts, start with the vertical shorts playbook, or read how playbooks work to make your own. Ready-made skills from others are in Claude Code video skills compared.

Make a video with FrameJam

Another way to start is FrameJam, a tool I built to make my own editing process easier. It's free and open source, and it lives inside Claude Code, ChatGPT or Cursor: you open it in the browser in your sidebar, like a plugin. Install it first.

On the styles page you pick a look for your video. In the walkthrough I chose Cyanotype Mac, clicked to use it, then typed the FrameJam skill that comes with the plugin and my prompt:

/framejam:start Create a 10-second video about a rocket launch.

The agent is connected to FrameJam over MCP (here are its tools), so it fetched the style I picked and made a rocket launch video in that blue, colored-pencil look.

Level 2: Plan your video with a storyboard

A storyboard maps out the video as a sequence of images first. Between your starting point and your perfect video, it gives you steps where you decide the path. Otherwise the AI can go in all kinds of directions, land somewhere you didn't want, and you have to go back and correct it.

Use this style and make a storyboard for a 15-second video that shows off the features of FrameJam.

The storyboard opens in FrameJam. Mine had nine frames, and I commented on them as I went through:

On the first frame: "Use a different mascot that looks more like the FrameJam logo."
On a later frame: "Make sure these are added dynamically and bounce."
On the whole storyboard: "After frame eight, add another one to show how easy it is to install."

Then finish the review. The agent listens for your feedback for a while; if it has stopped, copy the command FrameJam shows and paste it into the chat. Version two had the new mascot and the extra frame, and you can flip back to version one to compare. More on reviewing storyboards.

Level 3: Iterate with video feedback

Once you're happy with the storyboard, tell the agent: "Now turn it into a video." Classic video editing is a timeline with lots of buttons and things to drag around. Now editing is describing what you want, getting an output, and telling the agent what to improve and where exactly. You need that round of iteration for good videos, and it's what FrameJam makes easier.

The first version had no sound yet. My comments on it:

For the whole video: "Add some music to this."
On one part: use the images from FrameJam's actual styles page here, because the styles it showed weren't real ones.
On a range I highlighted: "Think of another way to bring in this banner, or just don't use Sent at all."

When you're done, press ⌘↵ to finish the review, and the agent makes version two. My music video went through eight versions this way. More on reviewing videos.

Free resources

FrameJam is free and open source. The music video process, the WebGL music video skill and my other skills are in the first email when you sign up for my newsletter. To edit a raw recording instead of making a video from scratch, read how to edit videos with Claude. For a short motion piece cut to the beat, try the motion design showreel playbook.

Questions

Which model is best for making videos with Claude?

Opus 5.5 right now. It was a big jump for video, both in what's possible and in quality. Sonnet 5.5 is slightly worse but much cheaper, so it's a good alternative.

Should I use HyperFrames, Remotion or Manim?

HyperFrames has the agent write HTML (GSAP works well alongside it for the animations), Remotion does the same with React, and Manim is made for math animations like 3Blue1Brown's. If you don't pick one, Claude often writes its own little engine instead, but an existing framework is usually more efficient.

Do I need FrameJam to make videos with Claude?

No. Level 1 works in Cursor, Claude Code or the ChatGPT app with a link and a prompt. FrameJam makes it easier to pick a style, review a storyboard and give feedback on exact parts of the video.

What does FrameJam cost?

Nothing. It's free and open source.

Works with

Install FrameJam in Claude, ChatGPT or Cursor, or join the Discord to see what others are making.