跪拜 Guibai
← Back to the summary

vibe-knowledge-film: An Agent Skill That Codes Knowledge Videos Frame by Frame

Vibe knowledge videos are blowing up! Without editing or burning credits, I used WorkBuddy to produce two finished videos directly.

A no-filler Agent Skill test: vibe-knowledge-film is not a video generation prompt; it's a skill that breaks down "how to make pictures explain an idea" into steps: read the instructions first, then set the topic, turn the plan into a storyboard with one subtitle per line, draw each frame, do sample quality checks, and finally verify the finished piece with numbers. Two engineering solutions worth noting: first, it doesn't generate images but has AI write code to draw frame by frame, so the same frame drawn twice is identical, and you can change any frame; second, quality control is done in two layers: a 16-frame collage to see the overall picture, and an 8.4-second single frame enlarged to original size for details, catching problems that thumbnails can't show, like subtitles covering the picture, labels overlapping, or light-colored text on a light background. This article reproduces the process in five steps, with 7 complete reusable prompt texts, three real rework instances, and measured numbers from the finished videos.

I'm Xiao Hu, computer science bachelor's and master's from Zhejiang University, an action-oriented breaker, author of "AI Super Individual: Complete Guide to One-Person Company", focusing on AI training, super individual building, and company efficiency consulting.

Recently, something has quietly become hot on Bilibili: searching for "Vibe Knowledge Awards," the top video has over 6 million plays!

Hello everyone, I'm Xiao Hu.

It's not a short drama, not a talking-head video. No narration, no real person. Whatever the subtitle says, that's exactly what appears on screen.

Even though it's popular, making one yourself still has a high threshold. Many people who want to create knowledge content get stuck at the same place: they can explain the idea clearly, but can't produce the visuals.

Hiring someone to make an animation costs thousands; learning to edit videos is too heavy; and generating it on an AI video platform burns hundreds of credits for a two-minute clip, and changing one word means regenerating the whole thing.

In the past couple of days, I dissected a batch of these videos and tried a different approach: no on-camera appearance, no editing, no AI video platform generation. Using 7 prompts, I had WorkBuddy take an open-source skill from reading instructions all the way to a finished video, and ended up with two ready-to-publish clips, plus a set of files I can edit later. The star is HeyChengC/vibe-knowledge-film, an open-source skill specifically for making this kind of short video.

It doesn't generate images; it makes AI write code and the computer draws frame by frame. If you're not satisfied with one frame, you just change it and rerun.

First, let me show you the results. Both are horizontal, high-definition, with sound, and double-click to play:

The first one explains "Why does a piece of paper folded 42 times become thicker than the distance between Earth and the Moon," with a cold color palette, star points, and calibrated instruments.

The second one explains "Why losing 100 yuan hurts so much that you need to pick up 200 yuan to feel better," with a warm color palette, flat illustration, wallet and balance scale.

Same skill, two different art styles. Here's the step-by-step breakdown.

Step 1: Bring in the skill, but don't start right away

Start a new task, and the first instruction is just to read the manual thoroughly, not to do anything.

Please look at the skills already installed on my computer, find the one for making Vibe knowledge videos, and read the usage instructions. Then tell me in plain language: what visual styles it can produce, what kind of content each style suits, what steps are involved from a single line of copy to a playable video, and what I need to prepare in advance. Don't start working yet; just clearly explain the full process, in short bullet points.

It returned a screenful of information. I didn't read all the details, only noted three things: it can do two art styles; there are four steps from a line of copy to a finished video; and it will pause for my confirmation at three points: topic selection, storyboard, and sample video.

Step 2: Set the topic, find visual metaphors, and write the storyboard

I chose the topic "A piece of paper folded 42 times can be thicker than the Earth-Moon distance," and first asked it for a plan:

Following steps 1 and 2 of the skill, give me a complete plan for a 30-second knowledge video with the topic "Why can a piece of paper folded 42 times be thicker than the Earth-Moon distance?" Include: a one-sentence thesis, how to explain each concept in plain language, a familiar everyday metaphor for each concept that even a layperson would recognize, the moment in the video that makes people say 'oh,' and the source of every number and which parts are just analogies. Use the cold-color style. Keep it in short bullet points; don't start working yet.

The plan that came back included a table: one piece of paper, one ruler, one Earth-Moon line, three props used throughout. The thesis was one sentence: it's not that the paper gets thicker, it's that it doubles every time. The starting thickness was marked as "assumption," the Earth-Moon distance was verified with its source, and the analogies were separately flagged. I checked it and couldn't find any issues, so I let it continue.

Then I turned the plan into a storyboard where each subtitle is one line:

Following step 3, turn the above plan into a storyboard with one subtitle per line. Each row must include: how to draw this frame, what's on screen, what it does, the small text on screen, and what the viewer understands after watching it. Requirements: do not force images for questions or conclusions; each scene should support more than two sentences; the ending returns to the opening image. Also tell me at which second the title appears. Keep it short; don't start working yet.

The table came back with six frames. I covered the subtitle column and only looked at "what's on screen, what it does" and could tell it was about folding. That was enough. It left the question and conclusion rows blank in terms of illustration, which I accepted. The title is placed at 8.8 seconds.

Step 3: Draw the frames according to the storyboard

I created an empty folder, pasted the storyboard table in, and asked it to draw all six frames, but not assemble the full video yet.

The six frames are: measure the paper, fold, title, mileage ruler, Earth-Moon line, and return to the paper. On the folding frame, the paper really folds step by step, with numbers climbing alongside. On the Earth-Moon frame, the thickness column really grows from Earth and passes the Moon.

I asked it to extract 16 frames from the whole video and combine them:

Please check the folder for this video on my computer, read how the frames were made, then extract 16 frames from the whole video and combine them into one image. Tell me frame by frame: what each frame draws, whether any subtitle blocks the picture, and which second the title appears. Do not modify any files; only inspect.

The combined image came back; all six frames are there, Chinese text, title, and subtitles are clean.

The combined image only gives a rough view. So I also asked it to open the frame at 8.4 seconds at its original size and go through it line by line:

Please open the frame at 8.4 seconds at its original size and show it to me, then answer line by line with the three questions: what is this (an everyday word), what does it do (one verb), and what does the viewer understand after watching it. Also mention which storyboard frame it corresponds to. Keep it short; do not modify any files.

Once the original image was enlarged, two problems became obvious: the thickness column crosses the word "Moon," and the ending "fold" title sits on the white paper's top edge.

I asked it to move them, then checked again to make sure nothing else was damaged before continuing.

Step 4: Output the finished video and verify with numbers

Only after the sample passed did it output the full video. A 30-second clip finishes rendering in just over a minute, with visuals and sound together.

Then I asked it to do two purely verification tasks:

Continue this video and do two things with measured numbers: first, check whether the same frame drawn twice is exactly identical; second, verify the finished video's resolution, duration, whether both picture and sound are present, and the file size. Also, in plain language, explain why we can produce a video without using an AI video platform at all. Keep it in short bullet points.

The reply: the same frame drawn twice is exactly identical. The finished video is 1080p horizontal, 30 seconds, just over 6MB, with picture and sound both present. The duration matches, so this one passed.

Step 5: Switch art style and make another one

One wasn't enough. The same skill, this time with the warm flat-illustration style, and the topic "loss aversion": you lose 100 yuan on the street, then a while later find 100 yuan. Net zero, but you still feel awful, because the loss side feels heavier.

Using the same skill, make another video with a different art style: look at the folder for the second video, tell me which art style it uses, what it's about, and how it differs from the first one in color, characters, and subtitles, as well as the finished video's length and specs. Finally, in one sentence, explain why switching art styles isn't as simple as changing a name. Keep it short.

It replied that the two videos' visuals were made entirely separately, even the light intensity and sound layering were swapped as a whole set.

I followed the same QC process as the first one. The first version had three issues: the characters "lost/picked up" were covered by a banknote; the hand wasn't recognizable as a hand; and the light-colored text at the end was hard to read against the warm white wall. After fixing those and re-rendering, I finalized it.

This one is shorter, 24 seconds. The finished product is the second one shown at the beginning.

How to choose, and my advice

Suitable for: people who want to make knowledge shorts but don't want to appear on camera or learn editing; content teams that already have many topics and need batch output; science writers who want to break down "how to make pictures explain an idea."

Not suitable for: people who want realistic, live-action quality; people who expect a one-line prompt to produce a video with zero confirmations in between; people who don't want to install anything on their computer.

Three reminders:

Check the sample first, then output the full video. The collage gives the overall view, the individual original-size frame gives the details. Subtitles covering the picture, labels overlapping, and light-colored text on a light background are all invisible in thumbnails. In the first video, both issues I found were only visible after enlarging the original frame.

Don't expect one sentence to produce the finished video. You'll need to confirm several times along the way; one confirmation saves you from redoing the entire video later.

Numbers must have sources, and analogies must be labeled. "42 folds to the Moon" is a thought experiment; in reality, you can only fold paper a few times, so the video has to mark "assuming you can keep folding." Also, it only outputs horizontal video; if you want vertical, you have to reformat, not just crop.

Closing

Two days, two videos, and a set of editable files.

Anything wrong can be fixed on the spot: if text is covered, move it; if a hand is unrecognizable, replace it; if text is hard to read, change the background. Then rerun once, and everything else stays untouched.

If you also have a pile of topics that you can explain but can't visualize, this approach is worth a try.

The above is the complete hands-on record; the 7 prompt texts can be reused directly. And if you codify the storyboard and per-frame QC flow, then to switch topics you only need to change one line of copy, and you have your own knowledge-video production line.