Claude can make a short film with every number right, even if you can't edit

What You Get

With Claude Code and an open-source style library, you can turn one idea into an mp4 with voiceover, music and subtitles. You don't write any code, and you don't need to learn editing software. This article has 4 prompts. Copy them and change a few words to use them.

Let's set expectations first. This method cannot do real people on camera or live-action stories. It is good at building content into video, such as charts, explainers, news flashes and title cards. You bring the ideas and check the results. The Agent does the rest.

First, Two Kinds of AI Video

Most people use "AI video" to mean video generation models. You give the model a sentence or a picture, and it draws a short clip directly. For real people and real scenery, this works very well.

This article covers a different method. Claude writes a web page. What the page looks like at each moment becomes one frame of the video. The computer captures the frames one by one, adds synthetic voice and sound, and compresses everything into an mp4. The whole process runs on your own computer.

For you, the difference comes down to three points:

Text and numbers stay sharp: If the code says 38.6%, the screen shows 38.6%. You can trust it for charts and data updates.
You can change one part: If you don't like second 10, redo only second 10 and leave the rest alone.
No per-second fees: You pay only for your chat usage with Claude, with no extra generation fee. But one video keeps the Agent busy for 30 minutes to an hour, so the usage adds up.

Step 1: Have the Agent Install the Tools

Making the video needs four pieces of software. FFmpeg is the tool that compresses video. The other three are Node, Python and Git. The names look intimidating, but you don't have to install them yourself.

You also need a style library: Lemo-Opuscar. It is an open-source project with MIT-licensed code. It includes 66 ready-made video styles. Each one explains how that kind of video should be shot, and it comes with its own rendering tool. Think of it as a director's handbook written for AI. The author says it was tuned for Claude, and results with other models are not guaranteed.

Open Claude Code (it requires a Claude Pro or higher subscription) and send it this:

I want to make videos on this computer with Lemo-Opuscar (github.com/lemomo-ai/lemo-opuscar). Please help me set up the environment.

Requirements:
1. First check what is already on the computer and what is missing. Tell me the result in one plain sentence.
2. Install what the repo README asks for: Node 20 or newer, FFmpeg 7 or newer, Python 3.11 or newer, and Git. Then install Lemo-Opuscar the way the README recommends.
3. Before installing each item, tell me what it is, why it is needed, and roughly how much disk space it takes. Wait for my OK before you install it. If something is already installed at a new enough version, leave it alone.
4. If you need my computer password, tell me in advance.
5. After setup, render a test video a few seconds long, and tell me where the file is. I should be able to play it, and that counts as success.
6. Whenever you need me to make a choice, ask me a multiple-choice question.

Why it is written this way: Rule 3 makes the Agent ask before it installs, so you can stop it at any time. Rule 5 uses a test video as the pass test. That is much more reliable than "it should be installed by now."

Step 2: Pick a Style

The online gallery shows sample clips for each style. Spend a few minutes browsing it first. Below are a few styles picked by use. Click a name to watch its sample.

What you want to makeStyles to trySampleExplain one concept, a course clipWhiteboard ExplainerWatch sampleMonthly report, comparisons, growth curvesData StorytellingWatch sampleHoliday greetings, a traditional-style title cardChinese Ink WashWatch sampleA story for kids, a gentle brand videoCrayon Picture BookWatch sampleA fun process, a growth story16-bit Pixel RPGWatch sample

The screenshots above come from the Lemo-Opuscar repository by LemoLab, used under CC BY 4.0.

Not sure which one to pick? Ask the Agent directly: "I want to make a video about [your topic] for [platform]. Recommend 3 styles from Lemo-Opuscar and explain why for each."

Step 3: Place Your Order With a Fixed Template

Use the same template every time and change only the words in brackets. The "Source" line matters most. Give it one official link, and tell it not to write any number it cannot find in the original. That way, every number in the video can be traced back to a source.

New accounts get 10 free credits, plus 1 more for each daily check-in. Spend 1 credit to unlock and it stays open for you.

Why it is written this way: "Audience" helps the Agent decide pacing and focus. The "fact-check table" lets you check the numbers at a glance. The last sentence asks it to report problems on its own, so you don't have to hunt through the video frame by frame.

Here is an example. To make a 30-second video about your company's quarterly user growth, fill it in like this: style "Data Storytelling", about 30 seconds, content "our user growth this quarter", and use your company's public earnings page as the source.

A few practical tips:

If you ask for 30 seconds, the result often runs a few seconds longer. To lock the length, write "no more than 30 seconds."
For your first video, pick a short one of 20 to 30 seconds. Make longer ones after the process works.
If you want to see a rough draft before it commits, change the last line to "First show me a few key frames, and make the full video only after I confirm."

Step 4: Watch the Video and Ask for Changes in Plain Language

When the video is ready, watch it from start to finish. If something is not right, just say so. You don't need technical terms. The Agent understands requests like these:

What you seeWhat you can sayThe overall look is too rough"The visuals are a bit rough. Redraw them with a finer brush style."The pacing is too fast"It's too fast. Hold each frame one second longer, and the total can go up to 40 seconds."One part is wrong"Change the number at 12 seconds to 42,000 and leave everything else alone."The style doesn't fit"Try this topic in the Whiteboard Explainer style."The sound is too loud"Lower the music a bit and slow down the voiceover."

Ask for one or two changes at a time, and review each round before asking for more. If you send ten changes at once, it may fix some and break others.

Going Further: Give It Real Material

If you want more realistic visuals, don't have it draw everything from nothing. Give it real material:

3D objects: Have it pick free real-world scanned models from Poly Haven. It only needs to handle the lighting and camera. That looks far more real than having it model things from scratch.
Your own material: Product screenshots, screen recordings and data tables all work. Tell it which folder the files are in.
Public footage: Ask it to use only public domain or licensed material. Record the author and license for each file, and show them in a corner of the frame.

Notes

A video Agent usually runs for 30 to 60 minutes. It uses a lot of chat usage, so don't queue many videos at once the first time.
Videos with real people on camera or live-action stories are not possible this way. For those, use a video generation model.
Some voice tracks in the sample films are not licensed for reuse. The README lists which ones. For your own videos, use the default voices in the style library or voices you have the right to use.
Don't put other people's footage, logos or brands in your video without permission.

References

If you want to make a faceless explainer video without showing your face, see Faceless Explainer Video Tutorial, Step by Step. To learn what a Skill is, see the Skill Handbook.