Claude Opus 5.5 纯代码做动物科普竖屏短片

Finn @finnardley原推发布于 2026.11.05

在 X 上打开原推

Your daily reminder that you can make videos like this on automation i made this in full code with claude opus 5.5 n paid $0 in ai video gens enjoy my free prompt ↓ <inputs> Ask me for my animal or topic ("Why is the X the most disturbing animal in the ocean?", "Why is X terrifying?"), or pitch 3, each with 5 true facts that get worse as they go. Ask which image model I have (GPT Image 2 or any model that takes a reference image at 4K, about $0.03 an image) and which voice I have (ElevenLabs, or Kokoro for free). Then write the script and the shot list with seconds, and show me both before anything costs money. </inputs> <direction> A 1080x1920, 60 fps, 30-35 s vertical Short in the style of the big narrated animal-facts channels where a guy in a hoodie with a mic pops into every scene. Here everything is a flat 2D cartoon: thick dark brown outlines (#1f1418), flat colours, simple cel shading, big round googly eyes. The host: a cartoon young man with messy brown hair, glasses, a plain heather grey hoodie with no logo, dark jeans, white sneakers, always holding a black mic. - He is about 1000-1050 px tall on screen. His feet sit below the frame edge, so he is cut off at the shins. - He pops up from below the frame in 0.32 s with an overshoot ease. - His mouth flaps on the voice whenever he talks to camera. - His face always stays above the caption pill. The animal: one character in 9 poses (neutral, shocked, angry, plus costume poses for the facts: old with a beard and cane, a powdered wig, a teen with braces, drunk, a front view of the face, defeated on a plate). One costume per fact is the joke. A new place every 1.5-3 s. Never repeat a composition. Something always moves: - Animals swim with a tail wave and bob ±10 px. - The camera pushes in 2-4% per shot. - Pose swaps squash. - Hits shake the frame. Titles are one or two big words per beat in Luckiest Guy (Google Fonts), 104-190 px at y 430, tilted ±0.03-0.05 rad: - They wobble ±0.025 rad. - The outline is 0.22 of the size in #1f1418, with a darker copy of the colour 0.09 of the size underneath. - Colours are pairs of fill and shadow: yellow #ffd21f/#c27a10, red #ff4a3d/#8e1a14, white #ffffff/#9aa6b4, pink #ff7ab8/#a8336d, green #7dff6a/#2e9a2a. Captions: - 1-4 words in a white pill at y 1470: Luckiest Guy 66 px, dark text, a 7 px outline, pill height 1.55x the text. - The key word is yellow with a dark outline. - The pill pops from 0.82 to 1 in 0.13 s. A red marker ellipse circles the animal when its name is said. Banned: photoreal cutouts mixed into cartoon art, a real person's face, the real channel's logo or name on screen, a still frame over 1 s, the host covering the animal's face on its key word. </direction> <structure> Length: 30-35 s, about 115 words at about 3.5 words a second. Times are from a 33.5 s film; scale them to the voice. 0-1.9 s, the hook "Why is the X the most disturbing animal in the ocean?": - The animal glides in from off screen in 0.75 s while the host pops up bottom right, already talking. - On the animal's name: a red ring, the name as a title, and the host swaps to pointing at it for 0.7 s. 1.9-3.7 s: a punch-in on "disturbing". The water is graded red, the angry pose, an impact starburst, the word in red. 3.7-8.2 s, fact 1, the shocking number: - A counter rolls up to it. - Then a comparison everyone gets ("older than America") shown as a costume pose and a prop that springs in. 8.2-12.7 s, facts 2 and 3 played as gags: - One is a costume in a new place. - One is a race the animal loses. On "slower than you walk", the host strolls past it. 12.7-15.6 s: the gross detail, with a zoom into the face. 15.6-16.7 s: "But it gets worse." A red alarm grade with a pulsing vignette. The animal yells into the lens and the host swaps from talking to horrified. 16.7-21 s, the worst fact: - The host cowers behind the victim, which then vanishes in a poof. - An x-ray shot shows what is inside. 21-27.7 s: the human angle, each line with a person from the cast acting it out. The host joins in (drunk next to the drunk Viking). 27.7-29 s: someone happily eats it while the host turns green and gags. 29-33.5 s: the callback joke that ties back to the hook. The host talks to camera, then cracks up on the last word. Hold 1.15 s after the last word. </structure> <build> 1. Tools: Python, Playwright with Chromium or Edge, FFmpeg, OpenCV, Pillow, numpy, soundfile. Install what's missing and tell me what you installed. 2. Art, 4 images in total (about $0.12). Every prompt starts with "Match the exact art style of the reference image: flat 2D cartoon, thick dark brown outlines, flat colours with simple cel shading, big expressive googly eyes, clean vector look". Feed the first sheet you like back in as the style reference for the rest. - Animal sheet: 1:1 at 4K, "a 3 by 3 pose sheet of the SAME cartoon X, 9 separate poses on plain pure white, plenty of white space, nothing touching, no grid lines, no text, no ground shadows", every pose a full side view facing right except the front face. - Cast sheet: 1:1 at 4K, 9 different characters and props for the facts, same wording. - Background grid: 9:16 at 4K, "a 3 by 3 grid of 9 equal tall vertical panels separated by thin white gutters, each panel an EMPTY scene, no characters, no text, keep the lower middle open". - Host sheet: 9:16 at 4K, 9 full-body poses at the same size: (1) facing the viewer, mic below the chin, mouth CLOSED, (2) EXACTLY the same pose with only the mouth WIDE OPEN, (3) talking and pointing to the side, (4) horrified with a hand on the cheek, (5) crouching and cowering with sweat drops, (6) running in side view facing right, (7) laughing and leaning back, (8) gagging with a green face holding a toothpick of food, (9) drunk with spiral eyes. 3. Cutouts: - Flood the white from the border (seeds every 64 px around the edge) with a floating range of 7, so soft glows go too. - Feather a 3 px band outside the shapes with alpha = clamp((250 - min(R,G,B)) / 150). Un-mix the white from that band: colour = (c - 255(1 - a)) / a. - Group the pieces by the grid cell their centre falls in and drop pieces under 150 px. - Split the background grid on its white gutters (rows and columns whose minimum channel averages over 232). Crop 4 px in, cover-fit to 1080x1920 with Lanczos, unsharp mask (radius 2, 60%). 4. Host lip-sync: - Align pose 2 onto pose 1 by template-matching the glasses and eyes. Paste only its mouth onto pose 1 through a feathered ellipse (sigma 4). Now open and closed are the same drawing, so the outline never wobbles. - Mouth track at 60 fps from the voice: RMS per frame, smoothed over 3 frames. Open when it is over (70th percentile in dB - 15) and over 0.55x the loudest frame within ±0.1 s. - Hold every state at least 3 frames. Break open runs longer than 14 frames with a 3-frame close every 12 frames. That gives about 8 flips a second. - Anchor every host pose by its feet and its head centre x (the alpha centre of the top 28% of the cutout), so pose swaps never jump. Measure that per pose. 5. Voice: ElevenLabs eleven_v4 with the timestamps endpoint, a calm deep documentary narrator, stability 0.5, similarity 0.75, one take (about 300 credits for 600 characters). - Cut pauses (10 ms frames under -40 dB, longer than 0.2 s) down to their first and last 0.06 s, with 8 ms fades. - Then atempo 1.15 and a 0.15 s lead-in. - Rebuild word times from the character alignment through the same cuts and tempo. When a start lands after the next word's start, pull it back. - Free route: Kokoro, then whisper for the word times. 6. Film: one HTML canvas where every frame is a pure function of time: window .seek(t), READY, END, and EV with the sound times. One function per shot. - Sprite tail wave: draw the cutout in 28 vertical slices. Each slice moves dy = sin(8t - 4.5u) * wig * k, where k = (max(0, u - 0.3) / 0.7)^1.5 and wig is 3-14 px. - Pop: 1 - 0.6 e^(-9Δt) cos(20Δt). Pose-swap kick: 1 + a e^(-10Δt) cos(22Δt), with a = 0.1-0.22. Shake: A e^(-7Δt) sin(61Δt), with A = 22-34 px. - Props drawn in code: a rolling number counter, a waving flag, a flipping calendar, green fumes, buzzing flies, confetti, red rings, arrows, an impact starburst, poof clouds, bubbles. - Red alarm grade: grayscale(.5) sepia(1) hue-rotate(-35deg) saturate(3.2) brightness(.55). - X-ray: grayscale(1) invert(1) sepia(1) hue-rotate(150deg) saturate(3.5) on #071a2e with a 60 px cyan grid and a scan line. - Look up repeated words by their occurrence (the third "and"), or cuts land on the wrong line. 7. Check, render, sound: - Contact sheets of 12-15 frames at key words before rendering. Fix covered faces, the host over the animal's face, and titles off screen. - Render 60 fps with 8 motion-blur subframes over half a frame, then libx264 CRF 19. - Music: a quirky piano mystery bed, free on Pixabay, with the voice 11-14 dB over it in every phrase. - Mixkit SFX, one per event, placed by their peaks: a pop on every title, a whoosh when the host or the animal enters, a hit on the big beats, a bite, a boing, a sad trombone on the fails, a poof, bubbles. - Mix to -14 LUFS. </build> <gotchas> - A plain hue tint for the red alarm turned blue water green. Sepia first, then hue-rotate. - Long titles overflow at 160 px. Split them into two beats ("OLDER THAN", then "AMERICA" in red). - The host walking past on the key word hid the animal's face. Time his pass after the key word, keep him lower and smaller (0.62 scale), or draw him behind. - Photoreal sticker animals with code eyes read as slop next to a real presenter. One flat cartoon style for everything fixes it. - If the two talking poses differ by more than the mouth, swapping whole images makes the outline wobble. Paste the mouth only. </gotchas> <start> Ask me for my animal or topic, or pitch 3, then show me the script and the shot list. </start>

这条是引用着另一条写的：x.com/finnardley/status/2117997239125585942