I recently discovered an amazing open-source Skill that many of you might have seen too.
It's called Anatomy, which translates to "deep dissection" or "structural analysis."
https://x.com/wheresryan22/status/2108948624004526516
I find this tool particularly interesting.
Its core design philosophy is captured in one sentence:
The object is the explanation.
Instead of using text to explain concepts, Anatomy attempts to turn the principle you want to understand into a functional device, allowing you to see the entire process firsthand.
Simply put, it transforms abstract knowledge into interactive mechanical model animations where agents create moving parts, display complete components, and allow user interaction.
And the results are surprisingly beautiful and detailed.
For example, if you want to know how a mechanical watch works internally.
The aesthetic style is exactly what I love. In terms of interaction, you can manually wind the crown to see how the internal parts operate. You can also disassemble different sections to observe their structures.
This effect was generated by the author using the following prompt:
/anatomy explain how a mechanical watch keeps time, interactive
Translated, this means: "Create an interactive demo explaining how a mechanical watch keeps time."
There are many similar examples.
You can even use it to create a Rolls-Royce Trent 970 engine with a quarter-section view.
It doesn't just draw the engine casing; it details the fan, compressor, combustion chamber, turbine, and other internal structures very precisely.
Honestly, when I first saw these animations, I thought they were incredibly cool.
Especially for friends who enjoy studying mechanical structures and physical principles, there is so much to explore.
Using this skill is simple. Just open your Agent and tell it to install the recently popular Anatomy skill.
Or, copy the prompt below and send it to your Agent:
Help me install this skill: https://github.com/wheresryan22/anatomy/tree/main
Once installed, you can start playing around.
However, after running several tests myself, I noticed a significant issue.
Taking the watch animation above as an example, while the animations are indeed beautiful, as someone with no prior knowledge of mechanical watch internals, I got confused after watching for a bit.
What is each part? What does it do? How is time actually calculated?
Even after carefully reading the built-in text explanations provided by the skill alongside the animation, understanding remained difficult.
For people who already understand the principles, these concise and highly detailed mechanical animations help them observe structures more intuitively.
But if I am completely unfamiliar with the subject, no matter how exquisite the visuals are, I still cannot understand them.
So I thought that for those without a foundational knowledge base, letting the object move itself isn't enough. Ideally, there should be narration accompanying the visual, explaining what the thing is, what is happening now, why it happens, and what its purpose is.
Since this skill can already turn complex principles into animations, why not let AI generate a narrated, voice-over science explainer video based on it?
Then, something interesting happened.
I tried the same mechanical watch example.
Used Opus 5.5 with high reasoning intensity.
The prompt was simple, just one sentence.
It directly produced the video below in one go.
Apart from the voiceover sounding strange because it defaulted to MacOS system TTS, everything else was quite impressive.
The introduction was comprehensive and clear, covering which parts exist, their functions, relationships between parts, timing principles, etc. If you look closely at the gears, they mesh perfectly during rotation, and corresponding sound effects are included.
Compared to the pure animation above, this format is significantly friendlier for laypeople like me.
Watching such a video may not lead to mastery, but it provides a preliminary understanding of how a mechanical watch works, giving a roadmap for further learning.
Fixing the Thunderbolt-like voice issue is easy—just ask it to switch to Volcano Engine TTS, and it fixes it in one shot.
I also used it to make a video explaining GPU working principles.
Again, generated directly in one pass without modifications.
It first introduces what GPUs are for, their components, assembly, the process of calculating a frame, why GPU calculation is fast, how images animate, cooling methods, and finally summarizes input/output.
The structure is very logical. When discussing a specific part, the screen highlights or displays that section, making it clear and easy to understand.
The explanation here is relatively brief. If you want more detail, just add requirements to the prompt.
For instance, I asked it to make a detailed TCP three-way handshake explainer video.
Here I used Opus 5.5 Ultra intensity.
Except for some pronunciation issues with English terms, the video quality is quite high.
I shared it with a colleague from our company who has a telecommunications background. She said this content usually takes a full university lecture to cover, yet AI managed to explain it clearly in a few-minute video.
Recently, brakes have been a hot topic, so I also used this skill to explain brake construction.
As we all know, this is important stuff—learn well so you don't get burned, folks.
In short, playing with this over the past two days has been very fun.
Whenever you want to learn something, just let it generate a custom explainer for you. That feeling is truly satisfying.
After seeing the results, I know many of you care about token consumption and time.
Frankly, the token cost for making videos with this skill is much lower than I expected.
For example, the 5-minute TCP handshake video consumed about 1% of my $200 Claude Max weekly quota.
Other 5-minute videos also cost roughly 1% of the weekly quota.
If you plan to publish online and want every part to be extremely detailed, you might need top-tier models like Opus 5.5 with maximum reasoning intensity.
But if it's just for personal study, you can create a simpler version using more cost-effective configurations.
I also made a GPU principle video using Sonnet 5.5 with high reasoning intensity. For self-study purposes, the effect is absolutely sufficient.
Regarding time, generating a 5-minute video with Opus 5.5 at Ultra intensity takes about 1.5 hours.
There isn't much way to speed this up.
The author of the Anatomy skill deliberately made it slow.
He explained the reason as follows:
Details take time, so I prefer showing real numbers rather than guessing.
How far does his pursuit of realistic detail go?
When I asked it to explain Transformer principles, it tried to train a small model itself, which triggered Claude's safety filters and banned the session. I was terrified...
Of course, I must add a disclaimer.
AI-generated animations and videos, despite looking very refined, cannot guarantee that every mechanical structure, calculation process, and explanation is 100% correct.
Especially when learning professional knowledge, verify specific parameters and key principles.
But despite these limitations, I still believe the possibilities shown here are exciting.
Because it's not just about how to make a science video; it touches on a bigger question:
How should we learn knowledge we don't understand at all in the future?
While playing with this skill, I remembered a long post Karpathy published on X during National Day.
https://x.com/karpathy/status/2105819303471976479
He said we will likely spend more time understanding AI-generated content in the future.
To make this content easier to understand, he shared several methods.
First, let AI explain problems using clearer, simpler language.
He mentioned ASD-STE100, a technical English standard originally developed for aviation maintenance documentation. It restricts vocabulary and sentence structure (e.g., limiting sentence length, using active voice, consistent terminology) to minimize ambiguity.
He found that asking AI to follow this standard makes text much easier to read.
Here, I asked Claude to explain mortise and tenon joints both without and with this standard. Feel the difference.
With ASD-STE100 reference.
Second, he believes this is better than the former:
Let AI draw diagrams directly.
Going further, HTML web pages are even better.
Since AI coding is strong enough now, it can create beautiful, interactive web pages.
This is also the primary output format of the Anatomy skill.
I casually generated one with Opus 5.5 in one pass.
And the last method, which Karpathy favors most:
Let AI generate dedicated science explainer videos for specific questions.
I also made one using Anatomy to compare the effectiveness of different formats.
In my view, these formats don't simply replace each other.
Text is good for quick lookup, charts for showing relationships, interactive web pages for hands-on exploration, and videos for intuitive step-by-step understanding of processes.
Different knowledge, different people, and different learning stages suit different formats.
I asked colleagues how they started learning completely unknown topics back in college.
Answers were similar: search engines or finding relevant videos on platforms.
But the materials you find aren't necessarily suitable for you.
Some are too shallow, others too specialized. For obscure topics, you might not find any courses at all.
Ultimately, past learning materials were pre-made by someone and distributed to everyone.
We could only pick the best fit from existing content.
But now, with models like Opus 5.5 and skills like Anatomy, we can let AI generate learning materials on demand.
This saves time searching and satisfies curiosity that might otherwise be abandoned.
Even if you're curious about a tiny question, you can instantly ask AI to make a video explanation for it.
What a wonderful era.
That's all. If you liked this article, please like, comment, and share. It helps us a lot.
Thanks for reading. See you next time.
