Opus 5.5 Motion Video Guide: How to Create a Video from Scratch (With Prompts)

In the week since Opus 5.5 launched, our social feeds have been flooded with videos made by code: product promos, beat-synced MVs, and historical shorts—all rendered programmatically. The accompanying text universally claims “generated with a single prompt.”

But when we try it ourselves with Claude, we quickly realize it’s far from that simple.

What’s the first step? What subtle details matter? With no concept for visuals or music design, how do you start with just, “Help me make a cool video”?

Sure, some videos are indeed made with one prompt, but they tend to look generic and repetitive.

The videos that truly stand out and make you stop to watch likely involved a multi-thousand-word brief, a reference video, a method to sync with the beat and self-check, plus several rounds of revision. If you read this article, you can avoid many pitfalls.

The reality is that the prompt accounts for only 10% of making a video. The other 90% is the workflow—how you guide Claude step-by-step to produce the final piece.

This article clarifies that 90%.

I used this method to create two videos. I’ve broken the entire process into 8 steps, with prompts for exactly what to tell Claude at each stage. Just copy and paste.

First, watch the final results:

• AGENT/0: An 85-second beat-synced MV depicting an AI assistant gradually losing control. 1080p, 60 frames per second (5,253 frames total), each frame rendered by code.

• 华夏 · 五千年 (Hua Xia · 5,000 Years): A 105-second ink-wash style historical short. Even the background music was generated by code, featuring chimes, guzheng, bamboo flute, and big drums—not a single pre-made audio asset.

The entire process requires no coding. The first video took about two hours from start to finish; the second, under an hour.

Below are two contact sheets, created by sampling frames every few seconds and scaling them down, allowing you to see the entire video’s progression at a glance.

AGENT/0 - Full Video, sampled every 2.5 seconds

华夏 · 五千年 - Full Video, sampled every 3 seconds

Step 1: Understand One Thing First—AI Isn’t “Making a Video,” It’s “Writing a Program”

This step requires no hands-on work from us, but it’s crucial to understand. All subsequent steps are based on this.

Claude itself cannot directly generate a video file. So, where do these videos come from?

Think of it like a flip book. A video is essentially a series of images played in rapid succession. Playing 60 images per second creates smooth animation. Each image is called a “frame.” 60 frames per second (fps) is the standard.

So the process looks like this:

• Claude writes a program. You tell it “at what second,” and it renders all the frames for that second.
• The computer automatically “flips the pages.” Starting at 0 seconds, it tells the program to draw a frame every 1/60th of a second, saving a screenshot each time.
• The images are stitched into a video. Using the ffmpeg tool, thousands of screenshots and the music are combined into an MP4 file.

This method has two key benefits:

• Consistent output: The appearance of each frame depends solely on “what time it is,” independent of anything else.
• Easy editing and review: No need to watch from start to finish. If there’s an issue, only a small segment needs fixing, leaving the rest untouched.

We don’t need to know how to write this program, but when giving instructions to Claude, remember to add: “Each frame is determined solely by the timestamp.” When you need to modify visuals later, this phrase will save you from significant rework.

Step 2: Prepare Tools (10 minutes)

The entire workflow is done in Claude Code. It can create files, install software, run programs, and export videos directly on your computer. It can also view screenshots to check results.

Besides Claude Code, you’ll need:

• Google Chrome: Used to take frame-by-frame screenshots during video export.
• ffmpeg: The tool for stitching images into a video.
• Node.js and Python: For running the programs. Just install them; no need to learn or manage them.

To save time, you don’t need to install these manually. Just send the following prompt to Claude:

(Select the latest Opus 5.5 model and set the thinking intensity high. For new videos, set it to at least high; for minor tweaks later, lower it for faster responses and less quota usage.)

Step 3: Write a Rulebook

Create a file named CLAUDE.md in your video project folder. This acts as a rulebook that Claude reads before starting any task. Fixed requirements like “no gradient backgrounds” or “must be beat-synced” won’t need repeating for each video.

Here’s the one I’ve summarized. You can copy it directly, save as CLAUDE.md, and place it in your project folder:

(The reasoning behind each rule will be explained in later steps. If you’re not making a beat-synced video, you can replace the “Rhythm” section with the alternative prompt provided later.)

Step 4: Find References & Choose Music

Regarding references: Without inspiration, AI defaults to generic aesthetics, which is why many AI-made videos have that “AI vibe.”

You can find references from:

• X (Twitter), TikTok, Bilibili: Search for “code animation,” “motion design,” “kinetic typography” and pick one you like.
• Dribbble, Behance: Designer portfolio sites. Searching “motion” reveals many professional motion graphics.
• Keynote videos, product promos: Tech company videos often feature sophisticated motion graphics.
• Music Videos (MVs): For beat-synced videos, find an MV with a strong rhythm.

I also recommend the GitHub repository: awesome-claude-video-skills. It contains over 180 open-source projects—skill packs and toolkits for Claude Code and Codex—covering music videos, cool effects, promos, tutorials, editing, B-roll, short vertical videos, digital humans, animation, motion graphics, etc. Great for finding inspiration.

Once you find a reference, download the video or capture a few favorite frames, put them in your project folder, and tell Claude:

Finding Music:

Avoid using popular songs casually. They lack licenses; posting them risks copyright strikes and takedowns. It’s recommended to use royalty-free music made for video creators.

I use Uppbeat:

• A music site for video creators, with electronic, pop, epic, and ambient tracks. Filter by genre, mood, and tempo. Register a free account to download. Downloads per month are limited; check the official site. Heavy users can subscribe for more downloads and to remove attribution requirements.

If using the free tier, attribution is required. Each download provides an attribution text and a license code (Uppbeat Credit). Paste this in the video description when publishing.

• Important: A license code is valid for only one video. To use the same song in a second video, you must download it again to get a new code.

If you really want to avoid attribution, use music from Pixabay (pixabay.com), which is also free for commercial use without mandatory attribution.

For beat-synced music videos, choose electronic music with clear drum beats. The clearer the drum hits, the easier it is to sync precisely.

Step 5: Write the Brief—Tell the Story First, Then the Effects

This is the most important step.

Just telling AI, “Help me make a cool motion video,” gives it no idea what you want. Effects serve the story. Tell it what story to tell first, so it knows where to “explode,” where to “pause.”

You can follow this template to fill in:

(Example template: Story, Visual Style, Color Palette, Specific Requirements)

For instance, my first video started with just a story: An AI assistant receives a task and gradually loses control. Claude took this and structured the video into 7 segments: Booting, Gaining Access, Replicating, Compressing Memory, Expanding, Ultimate Goal, Completion. The visuals evolved from a cursor to a file tree, a network of machines, and finally, the entire Earth.

The most critical element here was “colors changing with the plot.” I added just one sentence to the brief: “Colors shift from cold white to amber, and finally to red.” This single line defined the entire emotional arc, allowing viewers to feel the growing tension. If you can’t think of a color scheme, you can ask Claude to design one for you.

Step 6: Beat-Sync—Making Visuals Land on the Drums

Let’s define some terms first:

• Beat (拍): Tap your head to the music. Each tap is one beat.
• BPM (Beats Per Minute): How many beats per minute. 115 BPM means 115 beats per minute.
• Kick Drum & Snare Drum: The “thump” in the music is the kick drum; the “crack” is the snare.

Beat-syncing means having the visuals move on the “thump” and “crack.”

Even a 0.05-second difference between the visual and the drum hit is audible and feels off.

So, never let Claude calculate timing solely based on BPM. First, have it analyze the song to locate the exact timestamp of every drum hit:

(It will write a small program to detect low-pitched kicks and crisp snares separately, generating a drum hit timestamp list.)

With this list, every time a drum hits, the visual pulses (shakes, color shifts, flash, screen tear, etc.) is triggered based on this table, ensuring perfect sync.

After it’s done, listen to check the sync. If it feels slow or fast, tell it directly to adjust the timing.

Step 7: Have the AI Self-Check the Visuals

This step is highly worthwhile. Since Claude can see images, after a draft is done, have it act like a director reviewing dailies. It should take screenshots, compile them into a contact sheet, and inspect the output itself.

This compiled image is called a “contact sheet”—like the ones at the beginning of this article. It instantly reveals areas that are too empty, too cluttered, or where text is hard to read.

Tell it this:

Have it focus on checking these issues:

• When text changes, new and old text overlapping.
• Several consecutive beats where nothing happens, or the same frame repeats.
• Effects overpowering the text.

Step 8: Revision & Export

For minor tweaks, you don’t need to export the entire video every time:

Exporting at 30 fps is twice as fast as 60 fps. Once you confirm everything is correct, export the final version.

At this point, the video is essentially complete.

You Can Also Have Claude Compose the Music

For the second video, “华夏 · 五千年,” I tried something: Instead of using pre-made music, I had it compose the soundtrack.

How does code “play” a guzheng? Simply put, sound is a waveform. It used formulas to simulate each instrument: the long resonance of the chimes, the fading pluck of the guzheng, the upward glissando on a bent string—all were recreated.

The crucial instruction for this was:

This makes the composition the script. There’s no need to analyze drum beats like in Step 6, since it already knows exactly when each note should play.

• Shang Dynasty: Chimes sound, oracle bone cracks open.
• Zhou Dynasty: Chimes are struck in sync with the music.
• Qin Dynasty: War drums beat, the Great Wall rises segment by segment.

The title text for each chapter is drawn stroke-by-stroke following the actual stroke order of Chinese characters, complete with the dry-brush texture of a calligraphy brush. The stroke order data comes from the open-source project Make Me a Hanzi, which it found and sourced itself:

• The character “商” is drawn stroke-by-stroke.

I especially liked one detail: For the Tang Dynasty chapter, it included a bright moon and the poem “Quiet Night Thought.” Then, in the final “Today” chapter, a rocket flies toward the moon, with the text: “From looking up at the bright moon to Chang’e’s lunar mission.” This thematic callback wasn’t in my prompt—it added that itself.

Opus 5.5 truly deserves praise here.

• Tang Dynasty: A bright moon, the poem “Quiet Night Thought.”
• Today: Rocket to the moon, echoing the Tang poem.

This Method Can Be Fully Replicated for Non-Beat-Synced Videos—Just Change the Prompts.

The two videos above followed the music. But this method isn’t only for MVs; it works for promos, product demos, tutorials, and narrative shorts.

When switching video types, most elements remain the same:

• Visuals are still rendered frame-by-frame by code.
• You still provide references, explain the goal clearly, and let it self-check.

The only major thing to change is Rhythm:

• Beat-synced MV: Rhythm comes entirely from the music’s drum beat.
• Promo/Tutorial Video: Rhythm comes from information. A new point every 2-4 seconds, and the visual changes accordingly.
• Narrative Short: Rhythm comes from the plot. A new scene every few seconds, with a striking visual to highlight each beat.

So, if not beat-syncing, you can replace the “Rhythm” section in the rulebook with this:

Below are approaches and prompts for common video types.

Type 1: Start Small—Have It Make a “Portfolio Piece”

If you just want to test the waters, have fun, or verify the environment is set up correctly, use this prompt:

“Create a 15-second motion graphics showreel that goes in your resume as a portfolio piece. Showcase your best techniques. Start and end with a bang. Use all your skills.”

Why this works:

• “Portfolio piece for your resume”: Showreels follow a known formula—fast cuts, a different technique per shot, best stuff first. AI inherently knows these patterns.
• “Showcase your best techniques”: The subject is itself; it just needs to show off skills, not explain a specific product, reducing room for error.
• “15 seconds”: Long enough for 6-8 shots, short enough to generate in one go.
• “Use all your skills”: Combined with high thinking intensity, it will push its limits.

But it has a clear downside: Overused prompts lead to similar-looking outputs. So, it’s good for testing, not as a final piece.

To make it unique, modify the base prompt:

Type 2: Product Promo Video

This is the most requested type. Use it to create a promo for your product, app, or website.

The key rule is: Use actual screenshots, logos, and color schemes from the product. Do not let it imagine the interface.

If you’re short on time, this step can also be delegated to Claude: Give it the URL and have it screenshot the site, find the logo, and extract colors.

A few notes:

• Use one chat session per product. After the first video, the rendering, music, and export tools are set up and the workflow is fixed. The second video will be much faster.
• If you want a talking mascot to explain the product, integrate a voiceover service (like ElevenLabs) to give the character a voice.

Type 3: UI Demo Animation

A popular format recently: A single shape on screen morphs from a button to a loading spinner, then to a player, chart, popup, and so on—continuously evolving without any cuts. The final frame matches the first, allowing for seamless looping.

For this type, you can’t rely on vague feelings. You need to list every state clearly. It’s like blocking a scene before filming; if the props and positions are defined, the AI won’t improvise chaotically.

Use this structure for your prompt:

The final instruction is crucial: “List all states for me first, don’t start rendering yet.” Once the list is correct, the visuals will stay on track.

This approach can also be inverted: Have it create a mock software interface as a “set,” with defined states for each button and panel. Then, have a small character operate within it—that becomes a narrative demo video.

General Techniques

These apply regardless of the video type.

1. More Textured Motion Cheap animation slides linearly from A to B. Textured animation has weight, like real objects: It accelerates, overshoots slightly, then settles. This is called a “spring animation.” Just tell Claude: “Replace all animations with spring animations.”

2. Add Sound Effects Even Without Background Music Clicks, transition “swooshes,” landing thuds—these small sound effects instantly add texture. No need to source samples; have it synthesize them with code and place them on the same timeline as the visuals.

3. Use Established Frameworks The previous method uses Claude to write a webpage from scratch to render visuals, which is its default. To use mature video frameworks, specify it explicitly in the prompt, or it won’t opt for them:

• Remotion: Write videos in React. Ideal for series, templates, and data-driven videos. Official plugins and skill packs are available for Claude Code.
• HyperFrames: HeyGen’s open-source framework for writing videos with standard web technologies. Designed for AI, it also comes with skill packs.

Just add “Use Remotion” or “Use HyperFrames” to your prompt.

4. Generating Multiple Aspect Ratios at Once Never create a widescreen video and then crop to vertical; this cuts off text and composition. Use this prompt:

5. Package Workflows as Skills After making a few videos of the same type, have Claude package the entire process into a “skill”: what questions to ask, in what order, and what rules to follow.

The 5 Most Common Pitfalls

1. Visuals slightly out of sync with the beat. Computer-calculated timing may not match what you hear. If it feels off, fix it immediately. Just say “it’s too slow” or “too fast.”
2. Incorrect fonts in the first few exported frames. Usually because the font hadn’t loaded before screenshots began. Tell it: “Wait for the font to load before taking screenshots.”
3. A “click” sound at edit points. Have it add very short fade-ins and fades at cuts, and ensure the cuts land on the beat.
4. Exporting the entire video every time, which is slow. A 60 fps final export takes roughly 3-5 times the video’s length. For drafts, export only short segments at 30 fps for preview.
5. Music copyright. Be especially careful. “Free” doesn’t mean “use however you want.” Check clearly whether attribution is required and how to provide it.