One Afternoon, No Video Editor: Making a Douyin Knowledge Short with Opus 5.5
In one afternoon, Opus 5.5 and I went from "no idea, no benchmark" to a 2:10 vertical short called Your Mirror Self Would Starve on Earth: every frame, subtitle, animation, score and sound effect generated in code. I never opened an editing app.
The niche is hot right now. Douyin's #vibe知识大赏 (Vibe Knowledge Awards) topic has 1.17 billion views and 51,000 participants (21st Century Business Herald, October 22); a Bilibili channel called 比特新语 only started posting in early October and hit 650,000 followers in 20 days.
What follows is the actual order I did things in: topic → benchmark teardown → 15-second sample → full cut → publish. Everything I said to the AI is in the appendix, ready to copy.
01 | Setup: You Only Need to Talk
Across the whole process I did three things by hand: pick the direction, find the benchmark, make the calls. The code, rendering and compositing all happened inside Opus 5.5's own cloud environment.
| Tool | What it did on this video | Do I install it? |
|---|---|---|
| Opus 5.5 (Claude) | Picked the topic, verified facts, wrote the storyboard, wrote all the code, reviewed its own output | Just an account |
| HTML Canvas | Drew every frame in code on a web canvas | No |
| Playwright | Opened a browser and screenshotted the animation frame by frame (3,900 frames for the full cut) | No |
| Python | Synthesized the score, hits, whooshes and counting ticks; analyzed the benchmark's tempo | No |
| ffmpeg | Encoded frames and audio into mp4; also pulled frames out of the benchmark | No |
| CapCut / Jianying (optional) | Only its music library, for a BGM; this video used an AI-generated score, so no rights issues | Optional |
Picture and sound were built on two separate tracks and joined in ffmpeg at the end.
The key point: this runs in Claude's Cowork mode, which ships a cloud computer with all of the above already installed. The entire conversation happened in a browser.
02 | Topic: If You Have No Idea, Let the AI Pitch Options
I started with nothing, so I told the AI plainly: "I want to make one too, but I have no idea and no benchmark picked out." It searched the current state of the niche and this year's Nobel Prizes, then pitched one topic per viral-title format, and proactively avoided information theory, optogenetics and IceCube neutrinos — all done to death.
| Format | Topic | Opening hook |
|---|---|---|
| Ride the news | This year's Nobel solves a riddle: why does life use only the "left hand"? | Your mirror self would starve on Earth |
| Counterintuitive number | 23 people is enough for a 50% chance of a shared birthday | Someone in your class shares a birthday |
| Emotional pain point | Why do you always choke when it matters most? | Falling apart after praise isn't your fault |
| Classic motif | Camus: one must imagine Sisyphus happy | Pushing the same rock every day — would you go mad? |
I picked "ride the news." The reason is simple: the 2026 Nobel Prize in Chemistry went to Kagan and Soai for "nonlinear effects and autocatalysis in asymmetric organic synthesis," which answers exactly why life uses one hand.
Before starting work, the AI verified the core numbers and only used them when two reports agreed:
- The Soai reaction (1995): an initial bias of just 0.00005%, amplified over several rounds to over 99.5% purity (The Paper, Huxiu)
- 0.00005% in plain terms: out of two million molecules, the left hand is ahead by exactly one
The key here: don't let the AI write the script directly. Have it pitch options, make the call yourself, then have it verify every number.
03 | Tear Down the Benchmark: Hand the Video Straight to the AI
On the topic page I found a viral video about the same Nobel, uploaded the mp4, and said only "analyze and learn from this." It pulled one frame every two seconds into contact sheets, sampled pixels for the palette, analyzed the audio for tempo, and returned a visual spec I could apply directly.
| Item | What the benchmark did |
|---|---|
| Spine | Frank's 1953 "three conditions" checklist; it returns to tick one off after each scientist |
| Bookends | Opens with "raise your hands," ends on the same pair of hands, landing on the word "chirality" |
| Visual metaphors | Particle hands, lock and key, a curve with "it, bent", counters ×1 → ×512, big numbers along the bottom |
| Palette | Deep blue-black ground (roughly RGB 9,12,20) with an elliptical vignette; gold = left hand, cool blue = right hand, red reserved for "when things go wrong" |
| Type | Bold Song/serif for titles (weight), monospace for numbers and axes (precision), key words tinted individually |
| Pacing | Score around 112 BPM, steady loudness; one subtitle per 2–3 seconds, dissolve transitions |
| Weaknesses | 16:9 landscape, a flat opening |
The visual spec the AI extracted, condensed into one card. When building the sample, these parameters went into the prompt verbatim.
There was a surprise after the teardown: my first sample had already collided with it — even the mirror gimmick in the opening was the same. The AI's advice: keep the subject, change the angle.
- Borrow the form: gold-and-blue palette, serif plus monospace, counters, matching bookends
- Add what it didn't cover: "mirror bacteria." In December 2024, 38 scientists published a warning in Science against creating them (ScienceNet)
- Go vertical, and put the hook in the very first frame
The key here: a benchmark is not for copying. Learn the form; the content has to diverge.
04 | Build a 15-Second Sample First, Then the Full Cut
The sample is the first 15 seconds of the finished video. Its job is to confirm three things cheaply: whether the hook grabs people, whether the visual style is right, whether the pacing works. I made two versions.
Version one ("why does life use only the left hand"): cyan and magenta glowing particles. Technically fine, but after the benchmark teardown it turned out to collide, so it was scrapped.
Four shots from version one. The visuals weren't bad — they just collided with the benchmark.
Version two ("mirror life"): the benchmark's gold-and-blue palette and type, in five shots:
| Time | Subtitle | Visual |
|---|---|---|
| 0–2.6s | Your mirror self would starve on Earth | Gold and blue mirror-image particle hands |
| 2.6–5.6s | Nothing it eats lines up | Food particles fall toward the blue hand and bounce off |
| 5.6–9.6s | Mirror bacteria are different | Bacteria eat nutrients that have no handedness, splitting ×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, matching the five rows above.
After rendering, the AI pulls frames to check its own work. On this version it found and fixed three problems by itself:
- The particle hand's fingers were fused together and didn't read as a hand → widen the gaps, add more particles
- One subtitle was too long and touched the screen edge → reduce the font size
- A label was hidden behind falling particles → move it
The key here: tune direction on the sample alone. A 15-second sample goes from prompt to file in minutes, far cheaper than finishing the whole cut and then tearing it up.
05 | The Full Cut: One "Go Ahead" Later, a 2:10 Video
Once the sample was approved, I replied with one line: "Looks good, keep going." The AI extended it into nine sections forming a loop: hook → Nobel puzzle → back to mirror bacteria.
| Time | What this section covers | Visual treatment |
|---|---|---|
| 0:00–0:15 | Hook: your mirror self would starve; mirror bacteria wouldn't | Same as the sample |
| 0:15–0:24 | A mirror — why would it frighten scientists? | Large gold type: "Nobel Prize in Chemistry" |
| 0:24–0:40 | Molecules have hands too; amino acids left, sugars right | 3D molecules rotating, failing to superimpose |
| 0:40–0:51 | The cost of the wrong hand: thalidomide | Data panels, counters rolling 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 plus a 50/50 progress bar |
| 1:01–1:15 | 1986, Kagan: "It, bent." | A dashed straight line against a gold curve |
| 1:15–1:32 | 1995, Soai: one extra molecule in two million → 99.5% | Molecules self-replicating, a dot matrix turning gold round by round |
| 1:32–1:42 | Maybe there was one such accident four billion years ago | A single drop of primordial ocean, gold diffusing |
| 1:42–1:58 | Mirror bacteria sit on the other side; counterarguments and a pause recommendation | Immune cells NO MATCH plus a debate panel |
| 1:58–2:10 | The other hand — should it be built? | Back to the opening hands, landing on "chirality" |
Fourteen frames pulled during the full cut's self-review, from the hook at top left to the ending at bottom right.
Roughly, the production went like this:
- Write the picture: each section is a function that takes a time
tand draws that frame. The sametalways draws the same frame, so any second can be pulled for inspection at any moment. - Pull 24 stills first: this caught the Soai dot matrix overlapping the subtitle, which was then shrunk and rearranged.
- Render frame by frame: 3,900 frames, two processes in parallel, about 3 minutes.
- Score the music: 112 BPM to match the benchmark. Chord changes per section, drums pulled out where suspense is needed. Hits, whooshes, counting ticks and alarms all keyed to specific frames.
- Encode and self-review: ffmpeg to mp4, then 14 frames pulled from the finished file to check for overlapping or overflowing subtitles.
The problem found at step 2: the dot matrix was too tall and sat on the subtitle. After the fix, the three rounds of amplification also read more clearly.
The score from step 4. Each gold line is a hit point aligned to a cut or a key line; the grey stretches drop the drums to leave room for suspense and the ending.
The key here: put the counterargument inside the film. The line "some scientists think the risk is overstated" pre-empts the comment-section pedants.
06 | Title, Copy and Publishing
The title and the first frame saying the same thing works best: what viewers click on is exactly the suspense from the title. Mine ended up as:
Vibe知识大赏 | Your Mirror Self Would Starve on Earth
The AI also offered three alternatives, each on a different format:
- Ride the news: This year's Nobel solves a riddle: why does life use only the "left hand"? (better for Bilibili)
- Counterintuitive number: One extra molecule in two million won the entire planet
- Manufactured suspense: 38 scientists signed a warning: do not build this bacterium
Four things were done at publishing time:
- Open the title with 「Vibe知识大赏」 and put #vibe知识大赏 in the caption — that's the topic's rule
- Use the hook frame as the cover; the AI exported a 1080×1920 image directly
- Post a comment from your own account and pin it, listing the source for every number plus the counterargument, and end with "which side are you on?"
- Publish between 7 and 10 p.m.
The cover is the hook frame from the opening, saying the same thing as the title.
07 | Pitfalls
| Pitfall | How to avoid it |
|---|---|
| Working in isolation and colliding with a viral video without noticing | Find the benchmark before building the sample; if it collides, change the angle, not the subject |
| Letting the AI free-associate "high-end" and getting a PowerPoint from a decade ago | Turn the benchmark's palette, type and tempo into hard parameters and feed them in |
| Over-promising in the hook and getting picked apart in the comments | Ideas like "would starve" need the logic spelled out in the cut; use "may" in subtitles on contested points |
| Passing common myths off as fact | For example, "thalidomide was safe with only the other hand" is a myth (the two hands interconvert in the body); the video only says "different effects" |
| Stuffing too much into one video | Every shot serves the closing line: "the other hand — should it be built?" |
| Using a trending track and getting throttled or taken down later | Have the AI synthesize a score to the tempo, or use commercial tracks from the CapCut library |
| Discovering subtitle overflow after the whole cut is done | Check stills before rendering; pull frames from the finished file afterwards |
| Trusting an event rule without checking it | I couldn't find an official source for "traffic boost before October 31" — confirm it on the topic page before publishing |
Appendix | Prompts You Can Copy
What I actually sent the AI was short. Below are the same messages tidied into reusable form; send them one at a time, in order.
① Find a topic
I want to make a knowledge short for Douyin's #vibe知识大赏 topic, but I don't have a subject yet.
Search the current viral videos in this niche and this week's news,
then pitch one topic per format — counterintuitive number / ride the news / emotional pain point / classic motif:
title, first-2-second hook, spine, closing question.
Avoid subjects that have been done to death. Once I pick one, cross-check every number against two sources.② Tear down the benchmark (send the benchmark video as an attachment)
This is the benchmark video I picked. Pull frames, sample the palette, analyze the audio, and tell me:
1. How the storyline and the bookends are designed
2. What visual each abstract concept got
3. Palette (with color values), typefaces, subtitle rhythm, score BPM
4. Its weaknesses, and whether my subject collides with it and how to differentiate③ Build the sample
Using the subject we settled on and the benchmark's visual spec, build the first 15 seconds as a sample:
vertical 1080×1920, 30fps, frames drawn on an HTML Canvas, screenshotted with Playwright,
score and sound effects synthesized in Python, muxed to mp4 with ffmpeg.
Put the hook in the very first frame — no fade in from black.
Pull keyframes before rendering to check for overlapping or overflowing subtitles, and fix problems first.④ Produce the full cut
The sample is good — go ahead with the full version, around 2 minutes.
Show me the storyboard first, then write it section by section. Put the counterargument in for contested points,
return to the opening image at the end, and close with a question anyone can answer.
After rendering, pull frames from the finished cut to review, then 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 on a different format),
one caption with the topic hashtag, and one pinned comment that gives the source for every number.FAQ
Do I need to know editing or coding to make a knowledge short with Opus 5.5?
What is the point of a 15-second sample?
How do I tear down a benchmark video?
What if my first sample collides with the benchmark?
Will the soundtrack run into copyright problems?
Summary
The AI does the drawing, the rendering, the scoring and the self-review. You supply a direction, a benchmark, and a yes or no. That split is what makes a full vertical short possible in one afternoon without ever opening an editor.
To confirm which entry point gets you Opus 5.5, see Claude Opus 5.5 Access Channels; for the same code-rendered pipeline applied to a music video, see How to Make a Music Video with Claude Opus 5.5; for installing and configuring Claude Code, see the Claude Code Guide.