60+ prompts. Sixty real videos made with Claude Opus 5.5, each credited to its creator and linked to the original post, with the toolchain they used and, where they published it, the prompt word for word. Opus 5.5 does not draw pixels: every clip here was rendered from code it wrote, or cut together from video models it directed.
Claude Opus 5.5 (model id claude-opus-5-5, released by Anthropic on 22 September 2026) is a language model. Anthropic's documentation lists text and image input and text output only, states that Claude cannot generate or edit images, and offers no video input: an animated GIF is read as its first frame. The videos credited to it are made in one of three ways. It writes animation code (a self-contained HTML page driven by a seek(t) function, Remotion or HyperFrames compositions, Manim scenes, Three.js and WebGL shaders, Blender Python) that a headless browser or renderer turns into frames and FFmpeg encodes into an MP4. It directs video, image and music models through APIs, writing the shot list and every prompt, then edits the results. Or it edits existing footage through FFmpeg, After Effects or DaVinci Resolve. A video billed as "one prompt" is often a brief of several thousand characters, run inside an agent with skills, API keys and hours of autonomous work, so each case below records what the creator actually supplied and what they reported spending. The 60 cases were posted between 22 September and 3 October 2026 and each was checked against the original post. On apimodels.app, Opus 5.5 is $2.40 per 1M input tokens and $12 per 1M output tokens, against Anthropic's $4 / $20 list.
make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out.
Model:
Post: "Opus 5.5 on Max effort".
Toolchain:
One sentence at max effort. Indexers describe the result as Canvas drawing cut to a 128 BPM grid; the usual route is one HTML file rendered frame by frame by a headless browser, then FFmpeg.
Clip:
15 s, 1920x1080, about 2.25M views (6 Oct 2026).
Same prompt, other runs:
@ajith_io (x.com/ajith_io/status/2103449416325890146), @kenn at xhigh (x.com/kenn/status/2103337314021937232), @himanshutwtxs with sound (x.com/himanshutwtxs/status/2103495232637882858), @thismacapital as a French product ad, "15 minutes, 2-3 revisions, all HTML/JS/CSS" (x.com/thismacapital/status/2103773635714375808).
Pitfalls:
Thousands of people ran this exact sentence, so the results rhyme with each other. A one-liner tests the setup, not an idea; use it to check your environment works.
Create a bold, dynamic 15-second motion graphics showreel that feels like the ultimate portfolio piece of an exceptionally talented motion designer. Showcase a wide range of advanced techniques: kinetic typography, smooth transitions, 2D and 3D animation, abstract geometry, fluid simulations, particles, distortion, creative masking, compositing, lighting, depth, and seamless camera movement. Keep the pacing fast, confident, and visually surprising, with every shot transitioning naturally into the next. Make it feel meticulously art-directed rather than like a random collection of effects. Push the creativity, polish, timing, and visual impact as far as possible. This should feel like the opening reel that gets a motion designer hired.
Model:
Post (Czech): all of it is code made in Opus 5.5 at max effort.
Toolchain:
Three.js, WebGL shaders and Canvas; the creator says no image, video or music generator was used.
One shape morphs through a dozen UI states on the beat
<inputs>
Ask me for: 8 to 12 UI states I want the shape to become (e.g. button, loader, player, slider, toggle, tabs, chart, command palette, toast), pure black and white or one accent color, and a royalty-free song around 120 BPM (e.g. Mixkit, free for commercial use).
</inputs>
<direction>
Dribbble-level UI motion. One shape, never cut: every state is the same element morphing its size, radius and color while its content swaps with a short blur. A cursor drives every change with real clicks and drags. Light warm-gray canvas, black and white components, one clean UI font (Geist). Springs everywhere, a tiny overshoot at most. The camera zooms so each state fills the frame. The last frame is the first frame, so it loops.
Banned: bouncy easing, particle bursts, glows, gradients on UI chrome, mismatched icon strokes, dead time, anything that looks like a template.
</direction>
<structure>
120 BPM, 7 bars, something happens on every beat.
Button → loader → check → dynamic island → music player with a play/pause morph → scrub the progress bar → it becomes a volume slider that stretches when dragged past max → a toggle flips on the beat → the knob becomes a liquid tab indicator → the tabs open into a chart that draws itself, with a tooltip on hover → it collapses into ⌘K → type to filter → enter → toast → back to the button.
</structure>
<build>
1. One HTML file, square 1440x1440. Every style is computed from time inside seek(t): no CSS transitions, no timers, no state carried between frames.
2. Springs are closed-form step responses. A value that changes target many times is the sum of one spring per change, so it stays a pure function of time.
3. The tab indicator's two edges ride different springs, so the leading edge stretches ahead of the trailing one. Same trick for the toggle knob.
4. Drags are direct manipulation: while the cursor is held, the value is computed from its position. On release it springs back from wherever it was.
5. Analyze the song with numpy for the beat grid and start on a downbeat. Place every UI sound by its measured peak.
6. Render with Playwright: 4 subframes per frame, blended with ffmpeg tmix for motion blur at 60fps.
7. Render one frame per beat before the full render. Fix anything off the grid, cramped or hard to read.
</build>
<gotchas>
Never put will-change on anything the camera scales or the text renders blurry. Text that swaps inside a morphing container needs its own enter and exit timing or it overlaps. Make the last frame identical to the first, cursor position and speed included, or the loop stutters.
</gotchas>
<start>
Ask me for the inputs, then show me the state list on the beat grid before you write any code.
</start>
Model:
Post: "opus 5.5 is f*cking cracked at motion design this entire video is code, 0 after effects". The author open-sourced the prompt and wrote "steal it".
Toolchain:
Single 1440x1440 HTML; every style computed inside seek(t); closed-form springs; numpy beat analysis of the song; Playwright renders 4 subframes per frame blended with FFmpeg tmix for 60fps motion blur; one frame per beat is checked before the full render.
Clip:
14 s, 1440x1440, about 1.04M views and 12K likes.
Pitfalls:
Needs Node, Playwright, FFmpeg and Python with numpy (a Claude Code-style environment). The gotchas block is the author's own scar tissue: will-change blurs scaled text; a last frame that differs from the first makes the loop stutter. Use music licensed for commercial use.
One-take keynote-style launch film with liquid glass and an iris transition
<inputs>
Ask me for: a one-word brand name for the wordmark (a verb works best), 9 to 12 high-res photos, a royalty-free song around 120 BPM with a drop and a quiet breakdown (Mixkit, free for commercial use), and a free stock clip of a plain wall with moving plant shadows (Pexels).
</inputs>
<direction>
An Apple-keynote launch film, 2D only, one continuous take. Every scene is made out of the previous one: nothing fades, blurs or cuts. Objects change shape instead: text rises out of a mask line, icons pop from zero on a spring, bars draw across, pages push, and a black shape floods the whole frame and contracts into the next scene. Warm off-white canvas, black UI, iOS 26 liquid glass over the photos. Archivo (wdth 125, weight 800) for the wordmark, Geist for UI. A cursor drives every change with real clicks, drags and long-presses. The camera zooms screen-studio style so each moment fills the square, and the cursor scales with it.
Banned: crossfades, blur-ins, brightness "developing", 3D flips, particles, glows, holds longer than 1s, anything that looks like a template.
</direction>
<structure>
120 BPM, 54 beats, something happens on every beat.
Open: the wordmark squeezes into its own period like an accordion, the dot grows into a black pill, a label rises inside it. Click: six iris blades close over the label and snap open onto a photo. The circle becomes a square, shrinks, the grid unfolds from behind it like a paper map (center, plus, corners), reflows into a bento, and a click zooms into one tile, landing exactly on the drop.
Glass: a glass word pops in letter by letter, melts into a droplet that stretches into a glass toolbar. The adjust icon turns it into a slider. Dragging relights the photo from day to golden hour (two aligned shots), and the knob turns into a glass lens while held. It lifts into a glass orb, the next photo opens inside it as a circle, and the orb expands into a lock screen: glass clock digits, date, and a home bar that stretches into a glass music player.
Stage: the lock screen pulls back into a phone (the bezel grows out of the screen edge). The Dynamic Island stretches like liquid, pinches off, flies over and grows into a Mac window that rolls up like a blind. Long-press the wallpaper, drag it onto a Safari tab, the page pushes in, drop it and it becomes the hero of a landing page. Scroll: the hero morphs into a framed print on a product card, with the mat and molding growing out of the photo's edge. Pick a frame color (it paints across) and a size, then the nav button flies down into "Order print".
Order: one black shape keeps morphing: Ordered ✓ → Printing % → On its way (a van on a route) → Delivered ✓.
Wall: the delivered circle floods the frame edge to edge, holds black for a beat, and contracts into the framed print hanging on the real wall footage. The same iris opens onto the print, closes again, the frame floods the screen and contracts into the pill → the dot → the letters spring back out, landing on the beat return. Last frame = first frame.
</structure>
<build>
1. One HTML file, square 1440x1440. Every style is computed from time inside an async seek(t): no CSS transitions, no timers, no state between frames.
2. Springs are closed-form step responses. A value with many targets is the sum of one spring per change, so it stays a pure function of time.
3. Liquid glass: each glass element holds its own clone of the scene behind it, filtered with an SVG feImage displacement map (a rounded-rect distance field) through three feDisplacementMaps at slightly different scales for chromatic edges, plus a rim light. Glass letters: a canvas distance field per glyph gives the map, mask and highlights.
4. Goo: blur + alpha threshold, then composite the source atop it so the glass stays sharp inside.
5. Iris: 6 blades around a hexagonal aperture. Each blade is its two vertices, both edge extensions and the SHORT arc between them.
6. Wordmark squeeze: every letter moves toward the dot by the same factor and its drawn width follows (narrow the wdth axis, scale the rest), so the letters stay touching.
7. Footage: re-encode all-intra (ffmpeg -g 1), load it as a blob URL, await 'seeked' before drawing each frame.
8. Sound: a downloaded SFX for every event (Mixkit), never synthesized, each placed by its measured peak. The song starts on a downbeat: the zoom lands on the drop, the wall sits in the breakdown, the wordmark returns with the beat. Loudnorm to -14 LUFS.
9. Render with Playwright: 4 subframes per frame blended with ffmpeg tmix, 60fps. Check one frame per beat, then scan for single-frame pops (frame-difference spikes 3x their neighbours).
</build>
<gotchas>
backdrop-filter: url() misreads displacement maps in Chromium, so clone the scene instead. A flood must overscale past the corners and take about 0.3s, or half the screen changes in one frame. A child with visibility: visible shows through a hidden parent, so use inherit. Text that swaps inside a morphing shape needs its own mask. python http.server can't range-seek video, so use the blob URL.
</gotchas>
<start>
Ask me for the inputs, then show me the beat map and 4 stills (open, glass, stage, wall) before you write the full film.
</start>
Model:
Post: "OPUS 5.5 IS THE ONLY MOTION DESIGNER YOU NEED… pure code". Full prompt (5,200+ characters) published in the post.
Toolchain:
Single 1440x1440 HTML with async seek(t); closed-form springs; SVG feDisplacementMap liquid glass; real footage re-encoded all-intra and seeked per frame; licensed SFX placed by measured peak; loudness normalised to -14 LUFS; Playwright 4 subframes + tmix at 60fps; frame-difference scan for single-frame pops.
Clip:
About 29 s, 1440x1440, about 130K views.
Pitfalls:
You supply the photos, the track and a stock clip. Chromium's backdrop-filter: url() misreads displacement maps, and python http.server cannot range-seek video; both are in the gotchas.
Minimal product launch: scanning wall of vertical clips, cut on the beat
<inputs>
Ask me for: the product name and a one-line promise, 3 to 5 UI moments to show, one accent color, 10 to 20 real vertical clips I own, and a royalty-free song with a clear drop (e.g. Mixkit, free for commercial use).
</inputs>
<direction>
High-end minimal. One idea per shot, lots of empty space, one accent color, one clean sans (Geist or Inter) with tight tracking. Masked type reveals, match cuts, one smooth camera language. Real footage only, never placeholder cards. No full stops in on-screen text.
Banned: shockwave rings, particle bursts, RGB split, camera shake, lens flares, neon glows, grid floors, flashing backgrounds, bouncy easing.
</direction>
<structure>
10 bars at 120 BPM, 2 seconds each.
Bar 1: the hook lands word by word on the beats.
Bar 2: one hook word morphs into the product UI. A cursor types and clicks.
The drop: a circle opens out of the button into a dark scene.
Then one move per bar: a wall of real clips with a scan line and 3 winners, the key output as big type, a 3D carousel of real videos with floor reflections and a motion-blurred whip onto one hero clip, the hero in a phone next to a panel that flips into results, big stats on push cuts, a 3-word ticker, a logo reveal, a fade to black.
</structure>
<build>
1. One HTML file at 1920x1080. Every style is computed from time inside window.seek(t): no CSS animations, no timers, no state between frames.
2. Real video: extract clips to 30fps JPEG sequences with ffmpeg and swap img sources per frame. seek awaits the image decodes.
3. Analyze the song with numpy: tempo, beat grid, energy per bar, the drop. Calibrate the grid to the real kick hits. Every cut sits on a downbeat, every UI hit on a beat.
4. Render with Playwright: 3 subframes per frame at t minus, at, and plus 1/240s, then blend with ffmpeg tmix for real motion blur at 60fps.
5. Place each sound effect so its measured peak, not its file start, lands on the event. Keep the effects quiet under the music. Loudnorm to -14 LUFS.
6. Probe 20 or more frames before the full render. Fix anything cluttered, overlapping or hard to read.
</build>
<gotchas>
Never set opacity or filter on a preserve-3d element, because it flattens and both faces show. Fade its wrapper instead. Measure element positions at runtime for match cuts. Only use music and sound effects whose license allows commercial use.
</gotchas>
<start>
Ask me for the inputs, then show me a storyboard with every timing on the beat grid before you write any code.
</start>
Model:
Post: "this entire video is code opus picked the music, downloaded the sfx, built every frame and synced it all to the beat by itself".
Toolchain:
Single 1920x1080 HTML with window.seek(t); real clips extracted to 30fps JPEG sequences and swapped per frame; numpy tempo, beat grid, energy and drop analysis; 3 subframes of motion blur; 20+ frames probed before the full render.
Clip:
20 s, 1920x1080.
About the prompt:
X auto-inserted a stray "http://" into build step 1 of the posted text; it is removed here and nothing else is changed.
Pitfalls:
Opacity or filter on a preserve-3d element flattens it and shows both faces; fade the wrapper instead.
Posts: "using Opus 5.5 to make game result animations" and "the Opus 5.5 game motion prompt is here" (Chinese).
Toolchain:
A 2D SVG version first, then rebuilt in Three.js as real 3D (MeshToonMaterial bands, normal + depth post-process outlines), a GSAP timeline, Canvas 2D particles and Web Audio synthesised effects, in one HTML page that plays on click and is screen-recorded.
Clip:
Gacha version with the prompt: about 17.6 s, 1920x1080, about 130K views. Treasure-chest version: x.com/op7418/status/2103724883301814408 (about 14.7 s, about 140K views).
Pitfalls:
The long "do not" lists are the lessons: inverted-hull outlines vary in width, CSS 3D flips of large elements project across the whole screen, and looping tweens that are not reset keep shaking after the reveal.
Pixel-art wizard casting a spell, pure Canvas code
Create a single self-contained HTML file that renders an animated pixel art wizard casting a spell, using vanilla JavaScript and Canvas 2D. No external assets, libraries, or network requests.
RENDERING
- Draw everything to an offscreen canvas at a fixed logical resolution of 128x96, then blit to a fullscreen display canvas scaled by the largest integer factor that fits the window, centered, with imageSmoothingEnabled = false and CSS image-rendering: pixelated.
- All drawing snaps to integer coordinates on the logical canvas. No sub-pixel positions, anti-aliasing, gradients, or shadowBlur.
- Fixed palette of ~24 hex colors: deep blues/purples for night sky, warm robe tones, 3-4 bright magic colors. Every pixel comes from this palette.
CHARACTER
- Build the wizard procedurally from filled rects and pixel runs, ~24x32 logical pixels: pointed hat with a bend, long beard, two-shade robe with darker outline, staff with a gem at the tip.
- Parameterize the pose (staff angle, arm raise, head tilt, robe sway). Animate parameters smoothly, then quantize to the pixel grid each frame so motion reads at an 8-12 fps pixel animation feel even though the loop runs at 60fps.
ANIMATION
- Looping state machine: IDLE (2-frame bob, beard sway) -> CHARGE (staff raises, gem flickers, sparks spiral inward) -> CAST (bright burst, projectile fires across the scene, 1-2 pixel screen shake) -> RECOVER (settle back). Ease pose parameters between keyframes.
- Pooled allocation-free particle system: preallocate and reuse. Sparks orbit the gem during CHARGE, explode outward on CAST, each particle stepping its palette index from white to magic color to dark before despawn. Snap particle positions to the grid when drawing.
- Fixed 60hz timestep update with rAF rendering. Zero object allocation inside the loop.
SCENE
- Minimal background: dark sky, a few twinkling 1px stars, moon, stone floor line. Character silhouette must read clearly.
- Subtle 1px rim light on the wizard from the gem, brightening during CHARGE and CAST.
QUALITY BAR
- Crisp pixels at any window size, seamless loop, stable 60fps, readable silhouette. Should look like a polished 16-bit sprite animation, not vector shapes scaled down.
Model:
Post: "opus 5.5, animated pixel art wizard, pure code prompt shared below" (launch day). Prompt shared in the reply.
Toolchain:
Single HTML, vanilla JS and Canvas 2D at a 128x96 logical resolution scaled by an integer factor, a fixed palette of about 24 colours, a state-machine animation and an allocation-free particle pool; the clip is a screen recording.
Clip:
About 11 s, 1080x890, about 450K views.
Pitfalls:
This one runs a live requestAnimationFrame loop rather than deterministic seek(t), so a screen recording depends on machine speed. For a stable MP4, rewrite it in the seek(t) style of the UI-morph template.
Colourful remix of the UI-morph template that exposed the default render pipeline
Summary
Ran @twoclipping's "one shape, a dozen UI states" template unchanged, with one extra instruction: make the video more colorful.
Model:
Post: "opus 5.5 is really amazing one shot".
Toolchain:
Per the author: Opus put all the code in one index.html, rendered it frame by frame in a headless Playwright window via a seek function plus eval, then assembled the MP4 with FFmpeg. It prefers zero dependencies and writes everything from scratch, and did not reach for Remotion or HyperFrames even though they were available.
Clip:
About 15.4 s, 1440x1440.
Pitfalls:
This post is the clearest evidence that the single-HTML route is the default. If you want a framework, name it in the prompt.
Three-step HyperFrames recipe: install the skill, reuse the showreel line, point at your project
make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out.
Model:
Post (Portuguese): "model used: opus 5.5 (effort: max)".
Toolchain:
1) Install the HyperFrames skill. 2) Use the one-line showreel prompt above. 3) Reference your project folder in the prompt. 4) Wait. The author says it used about 5% of a five-hour usage window. Relays a tutorial by @gabrielbuzziv.
Adopt the role of an expert motion designer. Build a 30-second animated explainer for my business as a single HTML page. 5 scenes. The customer's problem, what I do, how it works in 3 steps, one proof point, and my name at the end. Bold text, smooth transitions, my brand colours. My business [DESCRIBE WHAT YOU SELL, WHO IT'S FOR AND YOUR COLOURS]
Model:
Post: "Claude Opus 5.5 can make an explainer video for your business".
Clip:
34.5 s, 1920x1080.
Pitfalls:
The author's advice: direct it. Give it the scenes, the length and the pace; fill the bracket with what you sell, who it is for and your colours.
Arabic vertical promo for an AI coding-tutor app (Remotion + TTS)
Summary
Asked Opus 5.5 (medium to high effort) to build an Arabic, right-to-left vertical promo for an AI coding-tutor app in Remotion so it could be previewed quickly, limiting sound effects to camera shutter and lens sounds, with narration generated by a separate text-to-speech model.
Model:
Author's follow-up (Arabic): used Claude with Opus 5.5, medium to high effort.
Toolchain:
Remotion for previews; voice from Gemini 3.8 Flash TTS; about 1.5 hours of conversation.
Notion column-permissions trailer in Remotion: 2 days, 33 versions
Prompt not published
No prompt published. The trailer was built in Remotion (React and TypeScript) with Claude Code over two days: about 90 messages through an orchestration bot, 33 rendered versions, with the whole process (including rejected cuts) tracked in a public Notion document.
Model:
Author's note: "with @claude's opus 5.5 [xhigh]".
Reported cost:
About $400 of Claude credit at API rates; the author says a $100-a-month plan would also cover it.
Links:
Process document: notion.notion.site/column-permissions-trailer
Who posted it:
Made by a Notion employee for a Notion feature; not an apimodels.app customer case.
Clip:
About 31.9 s, 1920x1080.
Pitfalls:
A real brand film is far from one-shot: 33 versions before sign-off.
Shotbase launch video with HyperFrames, under 20 minutes
Prompt not published
No prompt published. The creator says Opus 5.5 produced the launch video for the Shotbase editing app in HyperFrames in under 20 minutes, in one attempt.
Model:
Post: "Claude Opus 5.5 1-shotted the launch video of @shotbaseapp using @HyperFrames_ in less than 20 mins".
Cosmos moodboard-app promo, two prompts in HyperFrames
Prompt not published
Prompt not yet retrieved (the author says the two prompts are in the replies). Built in HyperFrames for the Cosmos moodboard app with two prompts and no references.
Model:
Post: "Opus 5.5 is probably the biggest jump in output quality I've seen lately. Built this in @HyperFrames_ for Cosmos, 2 prompts only. No references at all."
Post title: "Chinese Civilization by Opus 5.5" (Chinese).
Clip:
About 217 s, 1920x1080.
Pitfalls:
A model Chinese director brief: the score is the clock, the beat grid and shot table come before rendering, dates are checked before delivery, and real modern people are kept out of frame.
Explaining recursion in nine radically different video styles
A video explaining recursion, where every explanation about recursion has a radically different video style, make this self-referential & clever & fast moving.
Model:
Post by Ethan Mollick: "I had Claude Opus 5.5 make…".
Clip:
About 75 s, 1920x1080, about 70K views.
Pitfalls:
A concept plus one structural rule (every section changes style) works better than a pile of adjectives.
We're going to try a little test. Do you think you could render a recipe motion graphic animation using javascript or html (w/e you think will produce the best) to show the full recipe from start to finish (empty glass to completed cocktail) - Explainer video style - Showing the recipe ingreidents + measurements as they're going into the cup. Should be a 30s video.
Model:
Post: "Opus 5.5 rendered this in HTML from (1) image".
Toolchain:
One reference photo uploaded, then an HTML/JS animation, then recorded. Typos in the prompt are the author's and kept as published.
Clip:
30 s, 1080x1080.
Pitfalls:
The typical image-in, animation-out use: Opus 5.5 can read a photo, not a video.
Interactive Raptor 3 rocket engine you can take apart (screen recording)
Summary
Asked Opus 5.5 to explain how a rocket engine works by building an interactive Raptor 3 in the browser that you can take apart: cut it open, trace oxygen and methane through the two turbopumps, and throttle it to watch the shock diamonds move. The video is a screen recording of that page.
Model:
Post: "I asked Claude Opus 5.5 to explain how a rocket engine works by building an interactive Raptor 3".
Links:
Live page: airsup.ai/rocket-engine. Same author, fusion-reactor version: x.com/konstantinsaifo/status/2104216976801587629
Twin-paradox explainer in Japanese, paper-cutout style, 1,395 Canvas frames
Prompt not published
Prompt not published beyond the method: in Claude Code the author gave Opus 5.5 the URL of a reference post plus brief instructions, and it produced a Japanese version of a twin-paradox explainer in a paper-cutout style, entirely in code, including BGM and sound effects generated in JavaScript.
Model:
Post (Japanese): "made a Japanese version with Claude Code x Opus 5.5".
Toolchain:
Canvas rendered 1,395 frames and converted them to MP4; rendering took about 2 minutes.
3Blue1Brown-style explainer on variational autoencoders
Prompt not published
No prompt published (the author calls it "one short prompt"). Manim for the animation and equations, a model actually trained on MNIST so every digit on screen is real, narration from a cloned voice with Qwen3-TTS on MLX, and original background music.
Model:
Post: "Opus 5.5 can now make animated explainer videos end to end, in a style similar to 3Blue1Brown".
Clip:
About 5.6 minutes, 1920x1080 (low reach, about 1.5K views).
No prompt published. A three-minute history of AI made 100% in code in Claude Code: about 7,400 lines of React and TypeScript in Remotion, every image drawn in SVG or Canvas, open-source TTS narration and a score composed in Python.
Model:
Post: "This 3-minute film was made 100% in code by Claude in Claude Code", with Opus 5.5.
Black-hole lensing lab polished by reviewer and fixer agents overnight
Prompt not published
First prompt not published. One prompt produced the first version of an interactive gravitational-lensing lab; then three rounds of three agents each (two reviewers listing 9-10 issues ranked P0/P1/P2, one fixer), focused in turn on materials, lighting and UI; camera interaction and science design; and a design-director sign-off with visual QA and performance. Each round started from a backup, ended with screenshot checks, and rolled back anything it could not fix.
Model:
Post: "I asked Opus 5.5 to explain gravitational lensing…".
Reported cost:
5 hours 28 minutes in total, about $90 at API prices.
Clip:
About 21.8 s, 3840x2160.
Pitfalls:
The most complete public example of the review-and-fix loop.
Infinite zoom through vintage collage worlds, generated assets composited in code
Build a looping "infinite zoom" animation, After Effects style: the camera
travels from landscape to landscape by flying through vintage objects.
LOOK: vintage collage realistic photo landscapes + black & white newspaper
cutout objects (halftone, white paper border, soft shadow). Film grain,
vignette, light flicker.
WORLDS (loop): snowy mountains → pocket watch (swinging on its chain) →
sea cliffs → box camera lens → desert dunes → magnifying glass →
misty lake → hand mirror → back to start.
Extras: floating hat, phone, umbrella, gramophone, key; a 1950s man walking
toward the watch; a whale swimming across the cliffs sky.
HOW:
- Generate everything via the Magnific MCP (Seedream 5 Pro landscapes,
GPT 2.5 transparent cutouts, depth maps, Kling 2.5 animation, Lyria 3
music). List the generations + credit cost and wait for my OK first.
- Split each landscape into 3 depth layers from its depth map and fill the
hidden areas. Parallax: layer scale = camera^Z, Z between 0.45 and 1.22.
- Each portal's glass holds the next world; cut seamlessly when it fills
the frame.
- Constant speed: exponential zoom to a fixed point, each segment's duration
proportional to log(zoom). Verify the cuts frame by frame.
- Man & whale: generate on pure green (#00B140), animate in place with
Kling, key out every frame, build a seamless loop (ping-pong if needed).
- Cutouts animate at 15 fps (on twos).
- 20 s loop, 1920×1080, 30 fps. Music: 96 BPM, cut to exactly 8 bars = 20 s.
DELIVER: an interactive artifact (viewer, AE-style timeline, music +
MP4 download) and a rendered MP4 with music under 30 MB.
Show me screenshots before each expensive step.
Model:
Post: "this animation made with Opus 5.5". The author shared the prompt ("I'll leave you the prompt").
Toolchain:
Generation through the Magnific MCP (Seedream 5 Pro landscapes, GPT 2.5 transparent cutouts, depth maps, Kling 2.5 animation, Lyria 3 music); Opus handled the depth layers and parallax, the exponential-zoom maths, green-screen keying, frame-by-frame cut checks, and delivered an interactive artifact plus an MP4.
Clip:
20 s, 1920x1080, about 73K views.
Pitfalls:
Note the cost gate: list the generations and credit cost and wait for an OK first, and show screenshots before each expensive step.
Short film made through the Krea MCP in 10 minutes for $8
Prompt not published
No prompt published. Opus 5.5 planned the short and generated it through the Krea MCP, calling image and video models rather than drawing any pixels itself.
Model:
Post: "Opus 5.5 has genuine creative taste. it one-shotted this video for $8 in 10 minutes using the Krea MCP".
Japanese vertical review short: Opus diagrams, TTS voice, Seedance 2.5 opening
Prompt not published
No prompt published. Division of labour: Opus 5.5 made the diagram animations, Gemini 3.8 Flash TTS produced the (cloned) narration, and Seedance 2.5 generated the opening face and voice. The creator thinks this trio covers most short-form videos.
Model:
Post (Japanese): diagram animation by Claude Opus 5.5, voice by Gemini 3.8 Flash TTS, opening face and voice by Seedance 2.5.
Eight-stage siege game in 2 hours, then a 60-second trailer from video models
Prompt not published
No prompt published. Opus 5.5 wrote the site and an eight-stage siege game in about two hours, deployed it and recorded gameplay; the 60-second trailer footage came from Seedance 2.5 and MiniMax H3 and was cut in CapCut.
Model:
Post: "Code: Claude Opus 5.5 Video: Seedance 2.5 & MInimax H3 @capcutapp_jp".
"Claude Pop" music video: a dictated brief and a 12-hour autonomous run
I've included an MP4 file and an original link to a video that is called "Claude Pop." It's a pop song that is about increasing rate of progress and the experience of the singularity approaching.
I want you to independently do an end-to-end complete pass on making an updated version of this video. Use the exact same audio track and think and feel very deeply about what is the best way to visually represent all of the lyrics on screen. You do not need to anchor to the current style, you can do truly anything that you think might best let you visually express yourself, including abstract motion graphics.
……
I think after that, what might be fun is if you use your visual reasoning skills and your ability to build animations in JavaScript, and then reconstruct the video from scratch as sort of an overlay, so that the visual continuity of the base is really there. It's like that animation technique where you shoot first in traditional film and then draw over top of it.
……
You're going to want to watch the entire video multiple times, take screenshots at individual parts, and think about if something is really up to the bar of quality that we need here.
……
make no mistakes.
Model:
Post: "I made this with one prompt using Opus 5.5. I spoke to my computer for 5mins, claude worked for 12 hours".
About the prompt:
Excerpt. The full brief is about 9,500 characters, in the post's replies; a full transcript is at claudevideo.org/videos/pixel-flower-illustrates-ai-time-horizon-progress-data. The "……" marks are cuts.
Toolchain:
Five minutes of dictation became the brief; Opus read a local reference library and skills; generated character sheets and scenes through an image API; used Seedance 2.5 for base footage with people and physics, split by lyric with lip-sync attempts; ElevenLabs for sound effects; then redrew the whole film in JavaScript over the footage ("shoot first, draw over"), watching it repeatedly and critiquing screenshots.
Reported cost:
A Max plan, a generation budget capped at about $2,000, and an ElevenLabs key; the brief told the agent to spend all available usage.
Clip:
About 141.5 s, 1920x1080, about 3.91M views.
Pitfalls:
"One prompt" here means a long brief plus a full environment. The brief has been copied widely; one analysis found 87% vocabulary overlap with another popular template.
1990s demoscene demo in C/C++ and OpenGL, synced to an S3M track
I would like you to create a kickass, impressive 1990s style demoscene demo using this S3M track as the music. Prior to getting to work, please see how to best play and analyze the S3M track. This will be important since synchronization of effects and appropriateness of scenes to the musical vibe (and this song does have a variety of sections with different attitude) is essential to a great demoscene demo. Also prior to getting started, please research what kinds of effects, physics, etc. would be good to use in combination. Please use C/C++ -- with appropriate demoscene/embedded/console gamedev sensibilities to avoid needlessly bloating compile times -- with OpenGL. In some cases, blitting to surfaces in realtime may be appropriate.. while in others, using OpenGL or even shaders is appropriate as well. It should be planned and polished to be spectacular and shocking. This is Claude's way of really showing off to the world that it kicks ass.
Model:
Post: "Opus 5.5 just one-shot this 90s-style demoscene demo for me… genuine one-shot (in colloquial X post sense - single prompt)".
Toolchain:
C/C++ with OpenGL, analysing and playing the S3M module for sync; the clip is a screen recording.
Source material:
Music: "Second Reality" by Purple Motion (Future Crew), supplied by the creator; copyright belongs to its authors. Shown here only as the creator's own demo, via the original post.
"I'm Upping My P(Doom)" music video painted with p5.brush and parallel subagents
Summary
Gave Claude Code the lyrics and the audio. Pass one (medium effort): use the Clawd character and give every lyric line an interesting visual and transition, with no scene ideas specified. Pass two (xhigh): switch to p5 brush strokes, make the scenes more interesting and have each scene flow into the next.
Model:
Post: "Claude Opus 5.5 has the best visual design of any model I have tested so far"; repo README: first generation Claude Opus 5.5 (Medium), second generation Claude Opus 5.5.
Toolchain:
Opus wrote an ANIMATION_GUIDE.md for parallel subagents (reported as seven) and a STORYBOARD.md; nine chapters, one file each; studio.html draws with p5.js and p5.brush; render.mjs renders frames in headless Chrome and encodes with FFmpeg. All animation took about 45 minutes.
Reported cost:
About 10% of a Max 5x weekly allowance (as relayed by a forum post).
Source material:
Music: the song comes from a 2024 YouTube video and is not the creator's own (README: "as far as I could find"); copyright belongs to its original authors.
Links:
YouTube: youtu.be/8j-hR4fJywU. Code: github.com/JohnHeibel/PDoomVideo (no licence stated, link only). Starter template by the same author: github.com/JohnHeibel/ClaudeAnimationBase (MIT).
The same brief re-run with Midjourney and a moodboard
Summary
Re-ran @donaldjewkes' "Claude Pop" brief (see the card above) almost unchanged, giving Opus 5.5 access to Midjourney and a moodboard, and let it work for 12 hours.
Model:
Post: "Gave Opus 5.5 donald's prompt, Midjourney, and a moodboard 12 hours later, woke up to this".
Clip:
About 306 s, 1920x1080, about 19.75M views (the most viewed case here).
Pitfalls:
The same brief with different reference material produced a completely different film.
"Stroke of a Pen" Bitcoin music video storyboarded by an agent swarm
Summary
Told Opus 5.5 to read the creator's Bitcoin and monetary-history wikis and make a music video with code only; the song was generated separately. Two revision rounds followed: the first because the characters looked like crude stick figures, the second to add Matrix-style code elements.
Model:
Post: "I told Opus 5.5 to read my Bitcoin & monetary-history wikis & make a music video with code only".
Toolchain:
Song generated with ElevenLabs; a swarm of agents storyboarded and coded about 75 shots on the beat for a 3 min 23 s vertical video. In round one it rewrote the skeletal rig.
Clip:
About 202.7 s, 1080x1920, about 190K views.
Pitfalls:
Character animation is a weak spot of code animation and needs its own iteration on the rig.
Mid-Autumn paper-collage short: p5.brush frames over generated backdrops
Prompt not published
No prompt published. The creator supplied the script and the music; Opus 5.5 drew the animation frame by frame in JavaScript with p5.js and p5.brush for a hand-drawn texture, Nano Banana Pro generated the background plates and paper textures, and Opus synthesised the sound effects with a Node.js program.
Model:
Post (Chinese): "a small animation made with Opus 5.5".
Clip:
About 40 s, 1920x1080.
Pitfalls:
Person on script and music, model on animation: a steady split of work.
Two-minute sand animation of 250 years of U.S. history
Make a 2-minute sand animation that tells the story of 250 years of U.S. history. Keep it lively, engaging, and tasteful. Add appropriate background music and sound design.
Model:
Post (Chinese): Opus 5.5, one prompt from start to finish with no edits.
Clip:
120 s, 1920x1080, about 76K views.
Pitfalls:
The author stresses that neither Blender nor Three.js was used.
Post (Chinese): Opus 5.5, "Pelican riding a bicycle" theatrical cut, generated from one prompt.
Toolchain:
No 3D models, textures or audio assets: Opus wrote its own GPU ray-marcher, and the pelican, bicycle, pier, sea, sky and score are all computed in code; 1,140 frames at 40 samples per frame.
About the prompt:
The original uses full-width Chinese quotation marks around the title; they are shown as straight quotes here. Wording unchanged.
Clip:
38 s, 1920x1080.
Pitfalls:
Give it a generous time budget ("take a whole day if you need").
Post (Chinese): Opus 5.5 is strong; I let it pick its own topic for an interactive page and a video.
Toolchain:
As the model described it: one WebGL2 shader computes every pixel with no images; sun and moon paths from Nanchang's latitude; the pavilion facade in Canvas 2D; all sound synthesised live in Web Audio, the guzheng via a Karplus-Strong string model; also delivered an interactive page (scrub the timeline, float lanterns, record video in-page).
Porco Rosso-style sunset dogfight recreated in Three.js, from one line
用 web3D 复刻宫崎骏《红猪》日落海面上的经典空战：白绿红涂装战斗机、紫色积云、海面反光、吉卜力运镜，画面写 just breathe.
Model:
Post (Chinese): Opus 5.5 used Three.js to recreate the classic dogfight from Hayao Miyazaki's Porco Rosso. The author says this one line was the whole prompt.
Source material:
Homage: Porco Rosso, Hayao Miyazaki / Studio Ghibli. The film and its designs belong to their rights holders; shown here only as the creator's fan recreation, via the original post.
Links:
Repost with the video: x.com/NFT_Chen/status/2103882415299350878 (about 14 s).
Austerlitz, 2 December 1805: a five-minute film made entirely in code
Prompt not published
Prompt not yet retrieved (the author says "Code + prompt below" in the replies). A five-minute film of the morning of Austerlitz, all code: SRTM terrain on a 38 m grid and Natural Earth battle maps; WebGL2 rendering of terrain, sprites, particles, fog, volumetric light and a Kuwahara oil-paint filter; armies and effects written as pure functions of time so any frame renders alone; sound events derived from the picture with distance delay at 343 m/s and panning; offline Kokoro TTS narration; a score and mix built in Python; headless Chromium to FFmpeg in resumable chunks.
Model:
Post: "OPUS 5.5 IS CRAZY… I just had it turn that morning into a complete 5 minute film. All code."
Reported cost:
90 minutes to set up, 4 hours to render, about $40 of cloud-agent credit.
Links:
Code: github.com/WinterArc21/Battle-of-Austerlitz-Film (no licence stated, link only).
Clip:
About 301 s, 1920x1080, 24fps, about 640K views.
Pitfalls:
Long films need layers: narration beats, then shot timing, then sound derived from the picture.
No prompt published. A low-poly 3D retelling of the folk tale "The Taoist of Laoshan" made in Claude Code with Opus 5.5, with no written storyboard and no ready-made film skill; the only external asset is the author's cloned voice for narration.
Model:
Post (Chinese): Opus 5.5 made the 3D film directly; this time run in Claude Code with Opus 5.5.
Links:
Earlier film by the same author, "The Feast at Hong Gate": x.com/wshuyi/status/2103101793584796090
Market Street, San Francisco, the afternoon before the 1906 earthquake, in Blender
Recreate Market Street, San Francisco as it stood on April 17, 1906, the afternoon before the earthquake, in Blender.
Scope: the Ferry Building up Market to Fifth Street, including the Palace Hotel, the Call Building, the Chronicle Building, Lotta's Fountain and the Emporium.
Before modeling anything, build a source file from: the 1899-1905 Sanborn fire insurance maps (footprints, heights, materials, occupants), the Miles Brothers film "A Trip Down Market Street" (April 1906), period photographs from OpenSFHistory, the Library of Congress and the David Rumsey collection, and USGS topography. Record every building with footprint, height, facade material, occupant, and the source for each fact with a confidence level.
Build everything in Blender Python. No downloaded meshes, textures or HDRIs. Write reusable generators (Victorian commercial facade, mansard roof, bay windows, awnings, painted signage, gas and electric street lamps, cable car, horse-drawn wagon, early automobile) and assemble the street from the source file, so every building traces back to data.
Provide a 10 second video up the street.
Model:
Post: "I've been on a Blender kick with Opus 5.5. Its better 3D modeling and vision mean you can build an entire world from a single prompt."
Who posted it:
Posted by an Anthropic employee.
Clip:
About 8 s, 960x680, about 160K views.
Pitfalls:
Needs local Blender with Python and web access. The pattern to copy: build a sourced data file first, then reusable generators, so every building traces back to data.
Same procedural Blender shot: Opus 5.5 vs GPT-6 Astra
Summary
One prompt given to both Opus 5.5 and GPT-6 Astra: build a 10-second shot in Blender only, fully procedural, render it, and record a timelapse of the build.
Model:
Post: "First Opus 5.5 vs GPT-6 Astra test is 3D".
Reported cost:
Opus 5.5: 35 minutes, 199.6K output tokens, about $13.3. GPT-6 Astra: 28 minutes, 56.6K output tokens, about $14.5 (author's figures).
Pitfalls:
Opus 5.5 writes many more output tokens at a lower unit price; budget for output.
Octopus modelled, textured, rigged and animated in Blender in about 2 hours (YouTube)
Summary
Asked Opus 5.5 to model, texture, rig and animate an octopus in Blender from scratch; it took about two hours. The video compares the result with GPT-6 Astra.
Model:
Video description: "Claude Opus 5.5 just dropped… It modelled, textured, rigged and animated an octopus in Blender from scratch in about two hours".
About the prompt:
The prompt is shown in the video and was not transcribed word for word; this card summarises it.
GPU fluid-simulated campfire film in JavaScript, self-critiqued about 40 times
Summary
Asked Opus 5.5 for the most realistic fire it could build using only JavaScript, then had a vision model critique each rendered frame like a VFX supervisor, fixed the issues and repeated, about 40 rounds, before rendering a scripted shot offline frame by frame.
Model:
Post: "I asked opus 5.5 in @agnt_gg for the most realistic fire it could build using ONLY javascript".
Toolchain:
A 3D fluid simulation on the GPU with flame colour from blackbody radiation; the review loop caught the fire dying out, ring artefacts on the ground and embers that looked like cheese slices; sound driven by per-frame flame brightness; one HTML file, one conversation.
Hand-scribbled "Opus5" turned into a 3D boxing animation, packaged as a skill
Prompt not published
No prompt published. The creator handwrote "Opus5" on a notepad, Opus 5.5 turned it into a 3D boxing animation, and the creator packaged the workflow as a reusable skill.
Model:
Post (Chinese): when Opus 5.5 finished this animation, I literally stood up.
Source material:
The animation depicts real public figures (labelled Dario Amodei and Sam Altman) in a boxing match. It is the creator's own work, shown here via the original post.
Links:
Skill: github.com/limin112/min-skill (no licence stated, link only).
Windmill modelled, rigged, textured and animated in Blender via MCP in 15 minutes
Prompt not published
No prompt published. A windmill modelled, rigged, textured and animated in Blender in 15 minutes by Opus 5.5 working through Higgsfield's MCP connection to Blender.
Model:
Post: "Claude Opus 5.5 + Higgsfiled built this animated windmill in Blender in 15 minutes".
Who posted it:
The vendor's own promotional post.
Links:
Longer tutorial: YouTube, "Claude Opus 5.5 + Blender: The Ultimate 3D AI Pipeline (Higgsfield)", youtube.com/watch?v=Xeq-BwMVBUA (Aaron Randall).
Opus 5.5 driving After Effects to post-process generated clips (two tests)
Summary
Test 1: gave Opus 5.5 a 15-second clip generated by MiniMax H3 and had it drive After Effects to retime the cuts to the beat and add one effect per section (particles, light rays, kaleidoscope, datamosh). Test 3: Suno wrote the track, MiniMax H3 made a character dance to it, and Opus extracted full-body skeleton data frame by frame from the dance and fed it into After Effects to drive text, light and camera.
Model:
Post (Japanese): had Opus 5.5 operate After Effects.
Toolchain:
Author's conclusions: rough instructions still give decent results, and the output is saved as an .aep project a person can edit afterwards; skeleton data removes manual tracking, but deciding when to show the tracked effect still takes a human.
Links:
Test 1: x.com/aicreataro/status/2102656273112326609 (15 s, 1440x1280).
A full YouTube video edited by Opus 5.5 through a DaVinci Resolve MCP
Prompt not published
Prompt not transcribed (the author says it is in his video and article). Raw footage recorded in Tella; DaVinci Resolve 21.1 driven through an MCP server; HyperFrames for B-roll and animation; Opus 5.5 as the editor of a full nine-minute YouTube video.
Model:
Post: "Claude Opus 5.5 edited my entire YouTube video".
Grok-generated dance clip recut and packaged by Opus 5.5
Prompt not published
No prompt published. A raw dance clip generated by Grok was recut and packaged (captions and on-screen text) by Opus 5.5; the post shows before and after.
Model:
Post (Chinese): video 1 is the Opus 5.5 edit, video 2 is the raw Grok output.
Clip:
Two clips of about 15 s each (1920x1080 and 960x960).
Opus played its own game and cut the best moments into a demo reel
Prompt not published
No prompt published. The creator built a full game with the Unity CLI and Opus 5.5; Opus then played it with several characters and spliced the best moments into the demo video.
Model:
Post: "I turned this into a full game with unity cli + opus 5.5, also opus did this full demo video, it played the game with multiple characters and spliced together the best bits".
Demo reel for a 110K-rigid-body WebGPU physics solver
Prompt not published
No prompt published. From a single text prompt in Claude Code, Opus 5.5 put together a demo reel for an open-source WebGPU physics solver running 110,000 rigid bodies.
Model:
Post: "Opus 5.5 even put together this whole demo reel video just from a single text prompt in Claude Code".
1. One HTML file rendered frame by frame (its default)
Opus writes a self-contained HTML page (Canvas, SVG, CSS or WebGL) that exposes window.seek(t), so every pixel is a pure function of time. Playwright or another headless Chromium captures each frame, and FFmpeg encodes the frames and mixes the audio. Left without instructions, this is the route Opus 5.5 picks on its own, with zero dependencies.
Strengths
Nothing to install beyond a browser and FFmpeg; change one line and re-render; beat-accurate cuts and real motion blur (render several subframes and blend them) are easy; the same file plays live in a browser.
Limits
An MP4 needs an agent that can run commands. In a plain chat you get the HTML and have to screen-record it, which drops frames on a busy machine. Long films bloat into one very large file.
Prompt template (our synthesis, not any creator’s original)
Make a [length]-second, [width]x[height] video about [topic] for [audience]. Output one self-contained HTML file with no network requests. Expose async window.seek(t); compute every style from t (no timers, no requestAnimationFrame loop, no CSS transitions, no unseeded randomness). Style: [one named style], one accent colour [hex], font [name]. Banned: [the default looks you do not want]. Structure: [each shot by second or by beat]. Render with Playwright at [fps] fps and encode H.264 with FFmpeg. Before the full render, show me one still every [N] seconds.
Ask for Remotion (React and TypeScript) or HyperFrames (HTML compositions) by name and install the matching agent skill. Point it at your product site or project folder so it pulls real screenshots, logos and colours. Most of the product-launch cases here took this route.
Strengths
A precise timeline with live preview; versions and templates are easy to keep; long pieces stay structured; fits team review and brand work.
Limits
Needs Node and the framework installed; Opus will not pick a framework unless you name it; Remotion has its own licence terms for commercial teams, so check them.
Prompt template (our synthesis, not any creator’s original)
Use [Remotion / HyperFrames] in this project to make a [length]-second, [aspect ratio] launch video for [product]. Take real screenshots, the logo and the brand colours from [URL] yourself; it must have music. Structure: hook (2 s) → the problem → three features, [N] s each → one proof point → logo and call to action. Give me the shot list with durations first; after I approve it, build each scene and render one still per scene to check it.
3. Math and science explainers in Manim
Say "use Manim" and name a text-to-speech engine (edge-tts, Kokoro or a cloned voice) for narration. A one-line request is enough to start; the model plans the lesson itself.
Strengths
Equations and geometry look right by default, in the familiar 3Blue1Brown idiom; prompts can be very short.
Limits
Manim and LaTeX must be installed locally; length runs away without a limit (a one-line request in this library came back at 7.6 minutes); narration and animation need an explicit sync rule.
Prompt template (our synthesis, not any creator’s original)
Use Manim to make a video explaining [concept] for [audience], under [N] minutes. Keep it plain, with concrete examples and a question that makes the viewer think. Derive each formula step by step, with a matching change in the picture at every step. Narrate with [TTS engine], aligned sentence by sentence with the animation. Output a 1920x1080 MP4 and an SRT subtitle file.
4. 3D and real-time graphics: Three.js, shaders, Blender, game engines
Have it build the scene in Three.js or raw WebGL, write its own shader or ray-marcher, script Blender in Python (or drive Blender through an MCP server), or play a game it built and record the best moments. The strongest briefs make it gather sources into a data file before modelling anything.
Strengths
The highest visual ceiling in this library, and the area where Opus 5.5's graphics and visual reasoning get the most praise.
Limits
Slow and token-hungry (one Blender shot took 35 minutes; one shader film ran for 16 hours); character animation is weak until you make it rebuild the rig; needs local software.
Prompt template (our synthesis, not any creator’s original)
Rebuild [scene] in [Blender / Three.js]. Before modelling, build a source file from [references]; record size, material and source for every object. Generate everything procedurally in [Blender Python / code]: no downloaded models, textures or HDRIs; write reusable generators, then assemble the scene. Deliver a [length]-second, [resolution] shot: [camera move]. Show me three key frames before the final render.
5. Opus as director: a shot list, then video models
Opus writes the logline, character sheets, a timed shot list and every generation prompt, calls video, image, voice and music models through their APIs or an MCP server, then cuts the results and adds titles, transitions or a code-drawn layer on top. On apimodels.app the video models in these cases (Seedance 2.5, MiniMax H3) and others such as Kling V3 and Wan 3.0 run on the same API key as Opus 5.5.
Strengths
Real people, motion and physics that code alone cannot draw; a consistent look comes from the code layer and the edit.
Limits
The least predictable cost: one widely copied brief told the agent to use up the whole plan; copyright and likeness questions in the generated footage carry into the final film; it needs an explicit cost gate.
Prompt template (our synthesis, not any creator’s original)
You are directing a [length]-second video. Tools: [image model], [video model], [voice / music model]; the keys are in .env. 1) Write a one-line logline, a character sheet and a shot list timed to the second, and wait for my approval. 2) List every generation call with its estimated cost and wait for my OK. 3) Generate the footage with [video model]; if a shot needs lip-sync, pass the matching audio segment as a reference and check sync afterwards. 4) Add titles and transitions in code, assemble with FFmpeg and normalise loudness to -14 LUFS. 5) Watch the whole cut, take stills at [N] timestamps, fix the three worst problems, then deliver.
Give it raw footage (a talking-head take, generated clips, a gameplay capture) and have it cut, caption, add graphics and music with FFmpeg, or drive After Effects or DaVinci Resolve through scripts or an MCP server.
Strengths
Aims at the most common job (short-form edits); the After Effects route leaves an editable project file a human can finish.
Limits
It cannot watch the footage, only sampled frames and a transcript; bridging into professional apps takes setup; pacing calls still need a person.
Prompt template (our synthesis, not any creator’s original)
Cut the raw talking-head footage at [path] into a [length]-second vertical short. First extract one frame per second with FFmpeg and transcribe the audio, then give me a timecoded summary and proposed cut points. Remove pauses and fluffs; add word-by-word captions ([font], [position]) with key words enlarged; put a B-roll shot or graphic every [N] seconds; add music licensed for this use, ducked under the voice. Output a 1080x1920 MP4.
How to write great Claude Opus 5.5 Video prompts
1Ask for the plan before the code. The strongest briefs in this library end the same way: show me the shot list, the beat grid or the list of states, then wait. Fixing a storyboard costs one message; fixing a rendered film costs a re-render and often a rewrite.
2Make every frame a function of time. Have it expose window.seek(t) and compute every style from t: no timers, no requestAnimationFrame loop, no CSS transitions, no unseeded randomness. Then a headless browser can render any frame on demand and FFmpeg can encode a steady 60fps. A live animation captured by screen recording drops frames whenever the machine is busy.
3Name the framework, or it will not use one. Left alone, Opus 5.5 writes one dependency-free HTML file from scratch even when Remotion or HyperFrames is installed. If you want a React timeline, Manim's equation animation or Blender's renderer, say so in the first line and install the matching skill.
4Ban the defaults by name. Anthropic's own prompting guide for Opus 5.5 notes that asking it to avoid a generic look only swaps one default style for another. The briefs that stand out list exactly what is banned (centered text on a gradient, everything fading in, particle bursts, neon glows, bouncy easing, lens flares) and pair that with one accent colour and one typeface.
5Render stills and make it critique them. It can read images but not video, so the review loop runs on frames: one still per beat, or twenty sampled frames, inspected before the full render. Several creators here ran that loop dozens of times or split it into reviewer and fixer agents, and credit it for the difference between a good clip and a forgettable one.
6Hand it the soundtrack, then let it do the maths. Give it a track whose licence covers your use and ask it to measure tempo and beats (numpy is enough), cut on downbeats, place each sound effect by its measured peak rather than where the file starts, and normalise loudness to -14 LUFS. With no track, it can synthesise music and effects in Web Audio or Python on the same timeline.
7Put a cost gate in front of every paid call. When it directs video or image models, make it list every generation with its estimated cost and wait for your OK, and show you stills before each expensive step. Without that line, an agent running for hours will happily spend whatever the keys allow.
8Match effort and time to the job. Creators report xhigh or max effort on the pieces that travelled and medium for small fixes; Anthropic suggests raising effort only where you measure a quality gain. Budget wall-clock time too: a five-minute history film here took 90 minutes to build and four hours to render, and one music video ran overnight.
Prompt formula
Director brief = the inputs it must ask you for + direction (one style, one accent colour, one typeface, an explicit banned list) + structure (every shot timed on a beat grid) + build contract (resolution, fps, seek(t), renderer, how to check frames) + known gotchas + a start line that asks for the plan before any code
Example: <inputs> Ask me for the app name, three screenshots and a 100 BPM track I hold a licence for. </inputs> <direction> 20 seconds, 1080x1920, one continuous take. Off-white background, one green accent, Inter. Banned: gradients, particle bursts, glows, bouncy easing. </direction> <structure> Eight bars: the name types on; a screenshot rises into a phone; a cursor taps through three screens on the beat; a delivery route draws itself; logo and the line "Fresh beans every Friday". </structure> <build> One HTML file, every style computed inside window.seek(t). Render with Playwright at 60fps, encode with FFmpeg, check one frame per bar first. </build> <start> Ask for the inputs, then show me the bar-by-bar plan before writing any code. </start>
Frequently asked questions
Can Claude Opus 5.5 generate a video by itself?
Not as pixels. Anthropic's documentation lists text and image input and text output for its current models and states that Claude cannot generate, produce or edit images. Every video on this page was either rendered from code Opus 5.5 wrote (HTML and Canvas, Remotion, HyperFrames, Manim, Three.js, Blender Python and so on) by a headless browser or renderer plus FFmpeg, or made by video models it directed through an API and then cut together. In a chat window with no shell you get the animation as an HTML page you can open and screen-record; producing an MP4 directly needs an agent environment that can run commands.
Can Opus 5.5 watch a video I give it?
No. It has no video input, and an animated GIF is read as its first frame only. The workaround every editing case here uses is to have it extract frames (and transcribe the audio) with FFmpeg and look at those stills. It reviews its own renders the same way.
Do the "one prompt" videos really come from one sentence?
Some do: the 15-second showreel prompt is a single sentence. Many of the widely shared cases ran on briefs of thousands of characters, with skills installed in advance, API keys for voice, music or video models, and runs of one to twelve hours, sometimes with dozens of review rounds. Costs on these cards are what each creator reported, from about $2 for a 26-second launch clip to around $400 of credits across 33 versions of a product trailer. Every card is marked as an original prompt, our summary of a method the creator described, or a case whose prompt was never published.
What do I need installed to reproduce these?
For an MP4: an agent that can run commands (Claude Code or a similar tool), Node.js, Playwright or another headless Chromium, and FFmpeg. The UI-motion templates also use Python with numpy to analyse the soundtrack; Manim cases need Manim and LaTeX; Blender cases need Blender; music-video and director cases usually need API keys for voice, music or video models.
Which approach should I start with?
For a product clip or a UI animation, start with the single-HTML route or a framework such as Remotion; both are on this page with full prompts. For a lesson, use Manim. For anything with real people, natural motion or physics, let Opus direct a video model and keep a cost gate in the brief. For footage you already shot, use the editing route and expect to make the pacing calls yourself.
How much does Opus 5.5 cost for this on apimodels.app?
$2.40 per 1M input tokens and $12 per 1M output tokens (Anthropic lists $4 / $20), billed per token with failed calls free, through /v1/messages or the OpenAI-compatible /v1/chat/completions, so Claude Code and other Anthropic-compatible agents connect with a base URL and a key. Video work is output-heavy: one creator's Blender test wrote about 200K output tokens, roughly $2.40 of output at our rate. Our own single-HTML test through apimodels.app was one call of 31,786 output tokens, billed $0.368, for a 12-second 1280x720 promo. If Opus directs video models, those calls are billed separately per model; Seedance 2.5, MiniMax H3, Kling V3 and Wan 3.0 run under the same key.
Can I reuse these prompts and videos?
The videos belong to their creators. Clips from X play straight from X's own servers, every card links to the original post, and we do not host or re-upload any of them. Prompts are quoted with credit and a link: several authors explicitly invited reuse, others simply published theirs, so check the post before using one commercially. Some cases use third-party music, pay homage to an existing film or show real people, and each of those cards names the source material. If you made one of these and want the card removed or the credit corrected, email us.
Run these prompts on apimodels.app
Claude Opus 5.5 is live here alongside 130+ other image, video, audio and language models — one API, one key, pay as you go, with $0.10 of trial credit on sign-up.
