How to Make a Vibe Knowledge Awards Short Video with Claude Opus 5.5: A One-Afternoon Workflow Without Opening an Editor (2026)
[//]: # (# How to Make a Vibe Knowledge Awards Short Video with Claude Opus 5.5: A One-Afternoon Workflow Without Opening an Editor (2026))
TL;DR: In one afternoon, using Opus 5.5's Cowork cloud computer, I went from no topic and no benchmark to a finished 2:10 vertical video, "The You in the Mirror Would Starve to Death on Earth". Every visual, caption, animation, music cue and sound effect was generated by code, and I never opened a video editor. The AI drew, rendered and self-checked; I only picked the direction, found the benchmark and signed off.
The video went out under the Douyin topic #vibe知识大赏 (Vibe Knowledge Awards), which already has 1.17 billion plays and 51,000 participants (21st Century Business Herald, Oct 22 (opens new window)). On Bilibili, the channel 比特新语 only started posting in early October and reached 650,000 followers in 20 days.
When I started I had neither a topic nor a benchmark. Below is the actual sequence: tools → topic → benchmark breakdown → 15-second sample → full cut → publishing. Every prompt I sent the AI is collected in the appendix and can be copied as-is.
# 01 | Tools: one account, one browser
In the whole workflow I personally did three things: picked the direction, found the benchmark, signed off. Writing code, rendering and compositing all happened in Opus 5.5's own cloud environment.
| Tool | What it does in this video | Do I need to install it? |
|---|---|---|
| Opus 5.5 (Claude) | Picks the topic, verifies facts, writes the storyboard and all the code, checks its own output | Just an account |
| HTML Canvas | Draws every frame with code on a web canvas | No |
| Playwright | Opens a browser automatically and screenshots the animation frame by frame (3,900 frames for the full cut) | No |
| Python | Synthesizes the score, hits, whooshes and countdown ticks; analyzes the benchmark's tempo | No |
| ffmpeg | Compresses images and audio into an mp4; also extracts frames from the benchmark video | No |
| CapCut (optional) | Only used for its music library when looking for BGM; this time the score was AI-generated, so there is no copyright issue | Optional |
Visuals and sound are produced separately, then merged in ffmpeg.
This uses Claude's Cowork mode: it comes with a cloud computer that already has all the tools above installed. The whole conversation ran in a browser, with nothing installed locally.
# 02 | Topic: when you have no idea, let the AI offer options by playbook
I had no idea at all, so I told the AI: "I want to make one too, but I have no idea and haven't picked a benchmark."
It first searched the current state of the category and this year's Nobel Prizes, then offered one topic for each of four viral-title playbooks, and deliberately avoided the angles that have been done to death: information theory, optogenetics and IceCube neutrinos.
| Playbook | Topic | Opening hook |
|---|---|---|
| Hot topic | This year's Nobel Prize answers one question: why does life only use "left hands"? | The you in the mirror would starve to death on Earth |
| Counter-intuitive number | It only takes 23 people for a 50% chance of a shared birthday | Your class must have two people born on the same day |
| Emotional pain point | Why do you always choke on the attempt you tried hardest at? | Choking after praise is not your fault |
| Classic theme | Camus: one must imagine Sisyphus happy | Pushing the same boulder every day — would you go mad? |
I picked the hot topic. The reason: the 2026 Nobel Prize in Chemistry went to Kagan and Soai for "nonlinear effects and autocatalysis in asymmetric organic synthesis", which answers exactly the question "why does life only use one hand".
Before starting, the AI verified the core numbers and only used figures that two outlets agreed on:
- Soai reaction (1995): starting from a deviation of just 0.00005%, several rounds of self-amplification pushed purity above 99.5% (The Paper (opens new window), Huxiu (opens new window))
- 0.00005% in plain terms: 1 extra left-handed molecule in 2 million
The key at this step: don't let the AI write the script directly. Make it offer options, make the call yourself, then have it verify every number.
# 03 | Benchmark: hand the video straight to the AI
On the topic page I found a viral video about the same Nobel Prize, uploaded the mp4 to the AI and said only "analyze and learn from it". It sampled one frame every 2 seconds into contact sheets, sampled pixels for the palette, analyzed the audio for the tempo, and returned a visual spec that could be applied directly.
| Dimension | What the benchmark video does |
|---|---|
| Story spine | Frank's 1953 "three conditions" checklist; after each scientist it cuts back and ticks one off |
| Opening and ending | Opens with "raise your hands"; ends on the same pair of hands, landing on the word "chirality" |
| Visual metaphors | Particle hands, a lock and key, a curve chart "It bent.", a counter ×1 → ×512, big numbers along the bottom |
| Color | Deep blue-black background (about RGB 9,12,20) with an elliptical vignette; gold = left hand, cold blue = right hand, red only for "something goes wrong" |
| Type | Bold serif headings (weighty), monospace for numbers and axes (precise), keywords colored individually |
| Rhythm | Score around 112 BPM with steady loudness; each caption stays 2–3 seconds, cross-fades |
| Weaknesses | 16:9 landscape, slow opening |
A visual spec card the AI extracted from the benchmark; these parameters were pasted straight into the sample-video prompt.
Then an unexpected find: my first sample overlapped with the benchmark — even the mirrored-hands opening idea was the same. The AI's advice: keep the topic, change the angle.
- Learn its form: gold-and-blue palette, serif plus monospace type, counters, bookend opening
- Add what it doesn't cover: "mirror bacteria". In December 2024, 38 scientists published a warning in Science against building them (ScienceNet (opens new window))
- Switch to vertical, and put the hook in the very first frame
The key at this step: a benchmark is not something to copy. The form can be learned; the content has to be clearly different.
# 04 | Make a 15-second sample first, then the full cut
The sample is the first 15 seconds of the full video. Its job is to confirm three things at the lowest possible cost: whether the hook grabs people, whether the look is right, whether the rhythm flows. I made two versions.
Version one ("Why Does Life Only Use Left Hands"): glowing cyan and magenta particles. Technically fine, but after the benchmark breakdown it overlapped with the reference, so it was dropped.
The four shots of version one. The visuals were not bad, but they overlapped with the benchmark.
Version two ("Mirror Life"): using the benchmark's gold-and-blue palette and type, in five shots:
| Time | Caption | Visual |
|---|---|---|
| 0–2.6s | The you in the mirror would starve to death on Earth | Two mirror particle hands, gold and blue |
| 2.6–5.6s | Nothing you eat lines up | Food particles fall toward the blue hand and bounce off |
| 5.6–9.6s | But mirror bacteria don't care | Bacteria feed on chirality-blind nutrients and divide ×1 → ×8 |
| 9.6–12.4s | Your immune system may not recognize it | Immune cells scan and show NO MATCH |
| 12.4–15s | They say: don't build it. | A counter rolls to 38 scientists |
The five shots of version two, left to right, in the same order as the table above.
After rendering, the AI pulls frames to check its own work. On this version it found and fixed three problems:
- The fingers of the particle hand stuck together and it didn't read as a hand → widened the gaps and increased particle density
- One caption was too long and touched the edge of the screen → reduced the font size
- A label was blocked by falling particles → moved it
Before and after: only once the finger gaps were widened and the particles denser did it read as two hands.
The key at this step: adjust the direction using only the sample. A 15-second sample takes only a few minutes from prompt to file, far cheaper than finishing the whole video and then throwing it away.
# 05 | Full cut: one "go ahead", and a 2:10 video
Once the sample was approved I replied with just "No problem, go ahead." The AI extended the sample with nine segments into a loop from hook to Nobel puzzle to mirror bacteria.
| Time | Segment | Visual device |
|---|---|---|
| 0:00–0:15 | Hook: the you in the mirror would starve; mirror bacteria wouldn't | Same as the 15-second sample |
| 0:15–0:24 | Why would a mirror scare scientists? | Large gold type "Nobel Prize in Chemistry" |
| 0:24–0:40 | Molecules have hands too; amino acids left, sugars right | 3D molecules rotate and fail to overlap |
| 0:40–0:51 | The cost of picking the wrong hand: thalidomide | Data panel, numbers roll to 40+ and 10,000+ |
| 0:51–1:01 | In a test tube it's always 50/50 — how did life pick a side? | Flask + a 50/50 progress bar |
| 1:01–1:15 | 1986, Kagan: "It bent." | Dashed straight line vs. gold curve |
| 1:15–1:32 | 1995, Soai: 1 extra molecule in 2 million → 99.5% | Molecules self-replicate, a 1,040-dot grid turns gold round by round |
| 1:32–1:42 | Maybe such an accident happened four billion years ago | A drop of primeval ocean, gold spreading |
| 1:42–1:58 | Mirror bacteria on the other side; the counterargument and the pause recommendation | Immune cells NO MATCH + a controversy panel |
| 1:58–2:10 | Should the other hand be built? | Back to the opening pair of hands, landing on "chirality" |
The 14 frames pulled during the full-cut self-check, from the hook at top left to the ending at bottom right.
The production ran in five steps:
- Write the visuals: each segment is one function that takes a time
tand draws that frame. The sametalways produces the same frame, so any second can be pulled at any time - Pull 24 still frames first: this caught the molecule grid in the Soai segment pressing onto a caption, so it was shrunk and re-laid out
- Render frame by frame: 3,900 frames, two processes in parallel, about 3 minutes
- Synthesize the score: 112 BPM, matching the benchmark. Chords change per segment and drums drop out where the video builds suspense. Hits, whooshes, countdown ticks and alarms are all aligned to specific frames
- Encode and self-check: ffmpeg compresses to mp4, then 14 frames are pulled from the finished file to look for overlapping or out-of-bounds captions
The problem found while pulling frames in step 2: the grid was too tall and covered a caption. After the fix, the three rounds of zoom also read more clearly.
The score from step 4. Every gold line is a hit aligned to a cut or a key line; the grey stretches drop the drums for suspense and the ending.
The key at this step: put the counterargument inside the video. The line "some scientists think the risk is overstated" preempts the most likely objection in the comments.
# 06 | Title, copy and publishing
I made the title and the first frame say the same thing: what viewers see after clicking is exactly the suspense promised in the title. The final title was:
Vibe知识大赏｜The You in the Mirror Would Starve to Death on Earth
The AI also offered three alternatives, one per playbook:
- Hot topic: This year's Nobel Prize answers one question: why does life only use "left hands"? (better for Bilibili)
- Counter-intuitive number: 1 extra molecule in 2 million ended up winning the whole Earth
- Suspense: 38 scientists signed a warning: never build this bacterium
Publishing involved four things:
- Lead the title with 「Vibe知识大赏」 and include #vibe知识大赏 in the description — that is the topic campaign's rule
- Use the opening hook frame as the cover; the AI exported a 1080×1920 image directly
- Post one comment yourself and pin it, listing the source for every number and the counterargument, then end with "Which side are you on?"
- Publish between 7 and 10 p.m.
The cover uses the opening hook frame and says the same thing as the title.
# 07 | Pitfalls
| Pitfall | How to avoid it |
|---|---|
| Working in isolation and not knowing your topic collides with a viral hit | Find a benchmark before making the sample; if it collides, change the angle, not the topic |
| Letting the AI improvise "premium style" and getting something that looks like a 2015 deck | Break the benchmark's palette, type and tempo into hard parameters and feed them in |
| Overselling the hook and getting corrected in the comments | Explain the reasoning behind claims like "would starve" in the video; use "may" in captions for contested points |
| Repeating common misconceptions as fact | For example, "thalidomide is safe if you use only the other hand" is false (the two enantiomers interconvert in the body); the video only says the effects differ |
| Cramming too much into one video | Every shot serves the final line, "Should the other hand be built?" |
| Using a trending BGM track and getting throttled or taken down | Have the AI synthesize the score to the beat, or use commercially licensed tracks from the CapCut music library |
| Discovering out-of-bounds captions only after the full render | Check still frames before rendering, then pull frames from the finished file to re-check |
| Believing campaign rules that were never verified | I could not find an official source for "traffic support before Oct 31", so check the topic page before publishing |
# Appendix: copy-paste prompts
The messages I actually sent were all short. The prompts below are generalized versions; send them in order.
① Find a topic
I want to make a knowledge short video for the Douyin #vibe知识大赏 topic, but I don't have a topic yet.
First search for the current viral videos and recent hot news in this category,
then give me one topic for each of these four playbooks:
counter-intuitive number / hot topic / emotional pain point / famous person or classic theme,
with a title, a first-2-second hook, a story spine and a closing question for each.
Avoid topics that have been done to death. After I pick one, cross-verify every number with two sources.
② Break down the benchmark (attach the benchmark video)
This is the benchmark video I picked. Sample frames, extract the palette and analyze the audio, then tell me:
1. How the story spine and the opening/ending are designed
2. What visual device is used for each abstract concept
3. Palette (with color values), type, caption rhythm, score BPM
4. Its weaknesses, whether my topic overlaps with it, and how to make mine different
③ Produce the sample
Using the topic we settled on and the benchmark's visual spec, make the first 15 seconds of the video as a sample:
vertical 1080x1920 at 30fps, draw every frame with HTML Canvas, screenshot with Playwright,
synthesize music and sound effects in Python, and assemble the mp4 with ffmpeg.
Put the hook on the very first frame; no fade-in from black.
Before rendering, pull key frames to check for overlapping or out-of-bounds captions, and fix them first.
④ Make the full video
The sample is fine, go ahead with the full version, around 2 minutes.
Show me the storyboard first, then write it segment by segment. Include the counterargument for contested points.
End by returning to the opening visual, and close with a question anyone can answer.
After rendering, pull frames from the finished file to self-check, and export a cover image.
⑤ Write the title and copy
Write 3-4 Douyin titles for this video (starting with 「Vibe知识大赏｜」, each using a different playbook),
one description paragraph with the topic hashtag, and one pinned comment that cites the source of every number.
# Frequently Asked Questions
Q: Can Claude Opus 5.5 generate a short video directly?
A: No. It does not use a video generation model. Every frame is drawn by HTML Canvas code, screenshotted automatically with Playwright, mixed with Python-generated music and sound effects, and finally compressed into an MP4 with ffmpeg.
Q: What software do I need to install to make a video like this?
A: None. Claude's Cowork mode comes with a cloud computer that already has HTML Canvas, Playwright, Python and ffmpeg installed. You only need a Claude account and a browser.
Q: Can I do this without knowing how to code?
A: Yes. In the whole workflow the human only does three things: picking the direction, finding a benchmark, and signing off. Writing code, rendering, mixing and self-checking all happen in Opus 5.5's cloud environment.
Q: How long does it take to render a 2:10 video?
A: The full cut is 3,900 frames; running two processes in parallel takes about 3 minutes. Music synthesis and the ffmpeg encode take extra time, and the final step is pulling 14 frames from the finished video to self-check.
Q: What if my sample video overlaps with the benchmark?
A: Keep the topic, change the angle. Learn its form (gold and blue palette, serif plus monospace type, counters, bookend openings), add content it does not cover, switch to vertical, and put the hook in the very first frame.
Q: Is there any copyright problem with the music?
A: No. This video does not use any stock track. The score is synthesized in Python at 112 BPM, and hits, whooshes and countdown ticks are aligned to specific frames.
# Summary
The AI draws, renders, scores and self-checks; I set the direction, supply the benchmark, and say yes or no. That division of labor is why one afternoon was enough — and why I never opened a video editor.
To see which entry point to use for Opus 5.5, read UPGPTs Claude Opus 5.5 channel comparison (opens new window); to apply the same code-rendering pipeline to a music video, see How to Make a Music Video with Claude Opus 5.5 (opens new window); for installing and configuring Claude Code, see the Claude Code guide (opens new window).