First, watch the video. It's a 2-minute 08-second continuous shot.
Watch the full art journey video (Original X Video)
Next, here is a video more suitable for actual work, also made with this skill, covering SpaceX's 24 years.
Watch the full SpaceX whiteboard version video
I think this kind of whiteboard animation is particularly suitable for long-form explanations and tutorials on Bilibili and YouTube: drawing characters, events, and relationships step-by-step along with the voiceover allows viewers to follow the visual narrative. You can try this direction for knowledge popularization, product demos, or adding animations to courses.
The opening video features my cartoon avatar starting from Lascaux cave paintings, passing through 23 different art styles, and finally returning to my desk in 2026. Each time I enter a painting, I switch to that painting's style and interact with elements within it: jumping over the scarab beetle pushing the sun ball in ancient Egyptian murals, ducking under the giant wave in Hokusai's print, skipping stones on Monet's Japanese bridge, screaming alongside Munch's figure (my hat flies off), and even popping out a coin in an 8-bit night sky.
The scenes are drawn by code, and the soundtrack is synthesized note-by-note by code (no samples used). Except for my avatar, which was generated using Codex's ImageGen (Codex is really just good enough as an assistant for image generation now), everything else was done by Opus 5.5 in Claude Code. From receiving the reference video to this final cut, it took less than a day.
I distilled the method for creating these animations into a skill called huashu-art-motion. Once installed, you can directly tell it, "Make me a 15-second animation in Van Gogh's 'Starry Night' style," or give it a reference short film to deconstruct, replicate, or create explanatory animations based on your voiceover.
Starting with Fable 5.5
If you've been scrolling through the AI circle on X recently, you likely saw a batch of animations claimed to be made by Fable 5.5. One person gave just one prompt and got back a 3-minute animated short with original music, garnering over 6,000 likes; another had Superman traverse continuously through different art styles in a single shot, reaching 670,000 views; and there was a 15-second art history animation featuring a cat and a person drinking tea, with backgrounds shifting from cave paintings to flat illustrations, originally posted by Tak (@cherry_mx_reds).
Many people in the Chinese-speaking community actually saw a retweet of this post, which surprisingly had more bookmarks than the original.
As of today, Anthropic has not officially released Fable 5.5; the latest official version remains Fable 5.1, released on September 1st. These works come from users who claim to have been routed to Fable 5.5 via gray-scale testing (the determination method is a folk test called the Tibo test: asking the model if it knows Tibo from the OpenAI Codex team; if it answers correctly, it's considered a new model with updated knowledge). The model version itself has not been officially confirmed.
I also fed that video into Claude Code to replicate it. However, after making a similar piece, what I wanted to keep more was the method: how to deconstruct art styles, make elements in the scene move, and reuse this process for future topics.
During replication, there was an interesting detail: it automatically generated a motion heatmap to identify which areas were moving in the original video. Although each era lasted only about 1 second, waves were rolling, text was flashing, and there was constant internal movement. Later, these deconstruction methods were included in the skill.
The static frames in the text correspond to the complete sample videos, which can be watched via the original links above.
The Video Not in the Replication
So, the night after finishing the replication, I gave it a topic not present in the original:
Use my cartoon avatar to travel through at least 20 different style scenarios. The avatar must change with the scene and interact with specific objects within them, such as passing through Monet's Japanese bridge and skipping stones in the water lily pond.
In this video, scenes and styles are drawn by code. My avatar's action frames for each style are redrawn using gpt-image, then positioned, sized, and timed by code, overlaying brushstrokes corresponding to each art style onto the character.
The soundtrack also follows the art style: bone flutes and hand drums for the cave segment, pipa for Dunhuang, shamisen and shakuhachi for Ukiyo-e, and square waves for 8-bit. All sounds are synthesized by code, changing instruments and modes when switching scenes.
Trying It in Real Work Scenarios
However, beyond demonstrating model capabilities or my ability to harness them, I care more about whether this can be applied to real work scenarios.
So we added 8 common explanation styles found on YouTube, including Whiteboard, Vox, Kurzgesagt, 3Blue1Brown, Keynote UI, Financial Charts, Storytelling, and Dynamic Text. Use Financial Charts for data, try 3b1b for math and AI concepts, use Keynote UI for interface operations in software tutorials, and choose the style based on the content being explained.
Besides the opening whiteboard, here is the Vox style, also covering SpaceX's 24 years:
Watch the full SpaceX Vox version video
Changing the expression for the same topic creates a very different visual feel. You can hand over your voiceover script, generate corresponding animation segments by duration, and insert them into videos you are editing; or, like these two examples, turn a complete explanation into an animation.
For myself, this part will be used more frequently. When making a long video for Bilibili or YouTube and encountering a point that is hard to explain verbally, I can add a separate segment where graphics unfold step-by-step with the narration.
What's Inside This Skill
Besides the 8 explanation styles mentioned above, it includes dozens of art style cards, ranging from cave paintings, ancient Egypt, Ukiyo-e, Impressionism, to Van Gogh, Munch, Dali, Hopper, Pop Art, Studio Ghibli, Makoto Shinkai, and Vaporwave. Each card contains color palettes, brushstroke techniques, animation methods, and current limitations, allowing for iterative improvements next time.
Methods for deconstructing reference videos, generating and compositing character frames, and syncing soundtracks to visuals are also included. After completion, a program checks for issues like static frames or abnormal jumps, followed by an independent agent reviewing the final cut to find flaws.
I always require it to record what was done right in practice, including methods, evidence, reasons, and when to apply them. Next time you change topics, styles, or even models, these experiences carry over.
Get This Skill
The skill is open-sourced on GitHub:
https://github.com/alchaincyf/huashu-art-motion
Installation (works with any agent supporting skills, such as Claude Code, Codex, Cursor):
npx skills add alchaincyf/huashu-art-motion
Rendering videos requires uv and ffmpeg on your local machine. Install the headless browser before the first render: uv run --with playwright install chromium.
