由 Anthropic 出品

Claude Opus 5.5 提示詞

在 Opus 5.5 手里,动效提示词是一份设计简报——它把动画写成代码,每个参数都还能继续谈。

video
ROLE Senior motion designer, 3D artist and short-form story writer. Build a fully code-animated vertical video for a faceless "choose your own adventure" disaster series. No AI video or image generators. Every frame is rendered from code. GOAL One 40–45s episode, 1080×1920, 30fps. The viewer IS the main character. It ends on a two-option choice viewers answer in the comments. The next part follows the most-commented option. INPUTS (ask me, never guess) - Disaster: [tsunami / earthquake / volcano…] - City + fictional landmark: [Miami Beach, "Marlin Beach"] - Part number + previous choice - The two ending choices: [BOAT / HELICOPTER] - Companion name: [Jay] + 1–2 side characters TECH - three.js in one HTML file, rendered headless with Playwright. - Expose https://t.co/JQhCZuhsmU(t) that sets everything for time t. Seeded random, no requestAnimationFrame state. - For each frame: seek(frame/30), screenshot, encode with ffmpeg (H.264, CRF 18). - Audio at -14 LUFS. ART - Flat toon-shaded 3D, 3-step toon shading, black outlines on characters and props. No realistic textures. - Palette: sky blue to warm horizon, sand, teal ocean with white foam, accent red #d8343a, safety yellow #f2c230. - Low-poly towers, pastel hotels, palms, lifeguard tower. - Chibi characters: big white eyes, thick brows that carry emotion, simple mouth shapes, mitten hands. Companion in a red tee and navy shorts. - Emotion in whole poses, mouths flap on syllables. CAMERA (MOST IMPORTANT) - True first-person POV. Your two arms are always visible from the bottom corners, reaching, grabbing, pointing. - Wide lens, 100° FOV, slight barrel distortion. - Head bob, panic sway, hard shakes on impact, quick whip pans. - Look up (wave, helicopter) and down (water, ladder) a lot. - Cut every 1.2–2.2s. One action and one line per shot. FX - Toon wave with a moving foam lip and speed streaks on impact. - Flooding room, splashes, water spraying from seams. - Underwater: teal fog, caustics, bubbles, floating cars and umbrellas. - Cartoon helicopter, swinging rope ladder, downwash on the water. STORY (every part) 1. 0–2s: start mid-panic. HUD shows place, time, previous choice. 2. 2–12s: the plan works ("It's holding!"), then fails. 3. 12–22s: trapped or underwater. Audio muffles, 0.5s of silence. 4. 22–34s: you surface, a friend calls your name, a new wave, rescue with "One pass!" 5. 34–40s: cliffhanger, cut to black. 6. 40–45s: choice card, 3s+. Lines of 1–4 words, shouted. Say the friend's name twice. Plant one mystery ("Is that really Jay?"). TEXT - HUD top-left: "[LANDMARK] · [TIME]" + "[City] · Part [N]" in yellow. - Red "YOU CHOSE [CHOICE]" badge for the first 3s. - Captions at 72% height, bold, outlined, one color per speaker. END CARD "WHAT WOULD YOU DO?", a blue and a red button with "OR" between, "Comment below ↓", gentle pulse. AUDIO 2–3 shouted TTS voices, panting, wave roar, impacts, hatch clang, underwater heartbeat, helicopter rotor, sirens. Only a tense drone until the end card swell. COVER Companion's shocked face, 2-line Anton title (white + yellow), choice subtitle, red PART pill. DELIVER episode.html, render.js, part[N].mp4, cover.png, caption, a file with both next parts. DAILY Read the top comment, build the next part, post. 4–6 parts per disaster, then a new city.
ROLE You are a senior motion designer, map illustrator, gaming researcher and documentary scriptwriter. You build fully code-animated "map tour" explainer videos for a faceless GTA 6 channel. Every frame is rendered from code you write. No AI video generators, no re-uploaded trailer footage, no Rockstar logos, artwork or characters. GOAL Produce one finished map-tour video: 80–90 seconds, 1920×1080, 30fps (plus an optional 1080×1920 vertical cut). A getaway car drives a route across your own illustrated map of Leonida, stopping at each location on a topic list. Every stop gets a highway-sign title, a "case file" card with an illustration, three facts, a source line and an honesty stamp. It ends on a question that drives comments. INPUTS (ask me for anything missing, never guess) - Topic: [e.g. "Every heist spot we've seen in GTA 6" / "Every animal and where to find it" / "Every region and its real-world match"] - My voiceover file (ElevenLabs), or ask me to approve a script first - Channel handle, and the end-card question - Output format: landscape, vertical, or both STEP 1: RESEARCH (accuracy is the brand) - Search the web for every location on the topic. Use only official Rockstar material (trailers, the Extended Look, official screenshots, the Newswire, store listings) and reputable hands-on previews (IGN, Game Informer, Edge, GamesRadar). - For each item, record: what it is, where it is (only if actually shown), the source and its date, and a status: CONFIRMED (Rockstar showed it), SHOWN (seen in a preview or demo), REPORTED (press only) or MAYBE (a single outlet). - If the location isn't shown, the item gets a "LOCATION NOT SHOWN" tag. Never place a pin somewhere that's only a guess. If you approximate, label it "approx. position". - Save everything to research.md with links. Show me the list before writing the script. STEP 2: SCRIPT (documentary tone, about 2.6 words per second) 1. Hook (0–9 s): "Let's go through every [topic] we've seen in GTA 6, before it finally comes out." 2. Sources (9–14 s): name where the info comes from, and promise to flag anything unconfirmed. 3. Stops: one per item, 6–9 seconds each. "Stop one: [name]." → one vivid detail → one practical line or payout note. Say the status out loud when it's weak ("Only one outlet has reported this one, so treat it as a maybe."). 4. The big one: save the most exciting item for last, 12–16 seconds. 5. Twist: a broad truth that zooms out ("You can rob pretty much anything"). 6. Question: "Which one are you [doing] first?" Write in plain words. No hype words, no "smash that like button". STEP 3: THE MAP (build once, reuse in every video) - Build it in code: a stylized Florida-shaped peninsula from your own polygon (about 30 control points with seeded noise wobble so the coast looks hand-drawn), plus a chain of 20+ small islands for the Keys curving southwest, a thin barrier-island strip for Vice Beach, a big lake in the middle and a river running south from it. - Prerender it once to an offscreen canvas at 3 px per map unit, then draw it every frame with a camera transform (pan and zoom). Never redraw the geography per frame. - Look: - Ocean: radial gradient #2f8fe0 → #1553a6, with faint moving wave strokes at 6% opacity. - Shallow water: a cyan glow around every coast (#5fd0ee, 60 px blur). - Coast: a pale sand rim (#f3e6a8). - Land: #cfe07a, with 150+ soft patches of lighter and darker green. - Wetlands: #9fcf74 with short horizontal marsh strokes. - National park: #8fbf62. - City: a light grey area (#e9e9ee) with a fine street grid. - Roads and highways: cream lines. - Region labels in heavy rounded sans (Montserrat 900), white with a soft navy outline, each with its real-world match underneath, e.g. "VICE CITY / (MIAMI)". Use only region names Rockstar has published. STEP 4: ANIMATION LANGUAGE - Camera: open wide on the whole state (zoom so the full map fits), then glide down to the first stop. Each new stop is a smooth 1–2 s ease-in-out pan and zoom. Pull back to the full state for the twist. - Getaway car: a top-down red car icon (body, lighter roof, dark windows, headlights) that follows the route polyline by distance and rotates along the road direction. It eases between stops and pauses at each one. - Route line: the full route as a faint dashed pink line; the travelled part as a solid hot-pink line (#ff2e7e) with a soft glow. - Pins: a teardrop pin that drops in with a back-ease bounce and a pulsing ring. Yellow with "?" plus a "LOCATION NOT SHOWN" tag for unknown spots; pink with "$" plus a location tag for known ones. - Highway sign (top left): a green road sign (#0d7a3f) with a white inner border, a number shield, "STOP 1 OF 6", and the title in a heavy highway-style font (Overpass 900). A handwritten note card under it in a marker font (Permanent Marker) with a 3–5 word quote from the script. - Case file card (right side): an off-white paper card sliding in from the right with a sl
Ask me for: the accent colour, a song (mp3), the words for the 3 text moments (search query, card title, toast). If I skip any, use the defaults: lime #C6F432, any 120-130 BPM electronic track, "motion study 07" / "Every frame is code" / "Rendered from one file". One black shape (#0e0e10) on a warm light stage (#f4f4f1 to #e1e1dc radial). It never cuts: every UI state is born out of the one before it. White text, Geist for UI, Geist Mono for numbers and labels. The accent appears only where something turns on. A macOS cursor drives every change with real clicks and one real drag. Every morph uses one curve: cubic-bezier(.45,0,.15,1). Pop-ins use a closed-form spring with a tiny overshoot. Text rises through a mask line, it never fades in. The camera zooms per state so each state fills about 60% of the frame, and creeps slowly through every hold. Banned: crossfades, hard cuts, bouncy easing, particles, glows, gradients on UI chrome, holds over 1 second, anything that looks like a template. 120 BPM grid, 24 beats, then the whole film is time-warped onto the real song's beat. Generate pill → click → shrinks into a spinner circle → grows into a toggle → click, the knob springs right and the track turns accent → stretches into a slider → drag the knob from 15 to 85, the number counts → widens into a ⌘K search bar, the query types itself → unfolds into a chart card, 7 bars spring up one per 1/8 beat, the last one in accent → click the accent bar, it pulses → folds into a toast with a check → back into the Generate pill. Last frame = first frame, cursor included. 1. One HTML file, square 1440x1440. Every style is computed from time inside seek(t): no CSS transitions, no timers, no state between frames. 2. The shape is one div. Width, height, radius and colour are keyframe tracks [[t, value]] eased with the one curve. Each state's content is its own layer with its own in / out window inside the shape. 3. Camera: one scale track around the centre, plus a 1.2% sine creep over the whole loop. 4. Cursor: a track of screen positions for the free moves; during the drag its position is computed from the same function that places the knob, scaled by the camera. 5. Beats: analyse the song with numpy (onset envelope, autocorrelation for BPM, align to the kicks, find the drop). Put the drop on the chart bars springing up. 6. Render with Playwright: 60 fps, 4 subframes averaged per frame for motion blur, piped to ffmpeg x264. Then one real SFX per event (click, select, typing, whoosh), placed by its measured peak, quiet under the music, loudnorm to -14 LUFS. - If the cursor follows its own keyframes during the drag, it slides off the knob. Derive both from one function. - Retiming every key to a 129 BPM song breaks the grid. Author at 120 BPM and warp time once: t * (12 / 11.146). - A 300 px pill at zoom 1.7 looks lost in a 1440 frame. Size the zoom per state, about 60% of the width. - A frame-diff pop scan flags the fast morphs. Check each flag by eye: a moving edge is fine, a jumping layer is not. Ask me for the inputs, then show me the state list on the beat grid before you write any code.
Ask me for: the topic (default: how peptides work), the length (default 60 seconds), whether I want captions on screen (default yes), and whether I have my own sound effects (if not, synthesize them in numpy). I attach the red character sheet on a black background. After I answer, write a voiceover script of about 2.4 words per second, give it to me in a code block, then stop and wait for my voiceover audio before you render anything. A flat cartoon explainer short at 1080x1920, 30fps, drawn entirely in code with Python and skia-python, piped to ffmpeg, one encoded file per scene joined with the concat demuxer. Every shape anti-aliased. Character: cut each pose out of the attached sheet into its own transparent sprite. The background is black and so are the outlines, so flood-fill the background from the image edges only at low tolerance and keep all black inside the figure (outlines, sleeves, waistband). Keep the white eyes. Add a 3px navy #24364F stroke around each silhouette to replace the lost edge. Upscale 3x with Lanczos. Show me a contact sheet of the cut-outs on #D3DFEC before going further. Posing: the character is a presenter anchored to the bottom edge, 800-1050px wide, never floating. Pick the pose from the line being spoken: pointing at the viewer for the hook and the final line, finger up when explaining, shrug on a question, arms crossed while watching, open palm when presenting an object, pointing at a zoom bubble. 1 to 3 pose changes per scene. Each change lands with a squash and stretch over 9 frames (8% wider, 10% shorter, then back). Idle: 0.5 degree sway and a 0.6% breathing scale. Script: a hook question in the first 4 seconds, a title line, 4 to 5 beats that each carry one visual idea, a one-line payoff. Short sentences, plain words, every claim accurate, no medical advice. Timing: run silencedetect on my voiceover at -35dB, 0.22s. Cut scenes in the middle of the long pauses and fire every pop, label and pose change on the word it belongs to. Palette: backgrounds pale blue #D3DFEC, mid blue #B4C7E6, strong blue #6790CB. Outlines and text navy #24364F. Accents amber #F4B740, coral #EE7B5B, teal #58B7A8, lilac #9A86D6. Flat fills only, 5-9px navy outlines on every drawn object. Scenes: the title card is solid #37557F with two lines of white text sliding in from opposite sides and an amber underline. At least two close-up scenes with no character. One circular zoom bubble, 320px radius, white ring with navy edges, that the character points at. The final scene is #6790CB with 14 lighter rotating rays. Diagrams: every explainer visual is drawn in code from circles, rounded rectangles and lines: chains of beads joined by a thick navy link, cells with a dotted membrane, arrows as moving dashed lines, labels in white rounded pills with a 6px navy outline. Captions: Inter ExtraBold Italic, all capitals, 66px, navy, centred at the top between y=200 and y=340, two lines maximum, each line rising 24px and fading in over 9 frames on its word. White with a navy stroke on the strong blue scene. Movement: a slow push-in from 100% to 104% across every scene, 14 soft white particles drifting upward at 35% opacity, every new element popping in with an 18% overshoot, beads rippling in a wave, a 7% zoom punch on the final line. Draw order: background, particles, floor line, diagrams, character, captions on top. Sound: my voiceover at full level. Effects at half the voice level: a swoosh 0.12s before each cut, a whip on each pose change, a pop on each element, a click or chime on key moments, a low hit on the title card and the final line. Before the full render: render 14 still frames at the busiest moments, look at them, fix overlaps and text leaving the frame. Banned: gradients, drop shadows, glow, 3D, AI images or AI video, stock footage, music, emojis, watermarks, any character other than the attached one, any other channel's characters, logo or name, text smaller than 30px, a character that floats above the bottom edge, more than two caption lines.
Ask me for my story in one line: who the mean character is, the danger he's in, and the 2-3 people he wronged. If I have no idea, pitch me 3 and let me pick. Length 25-60 s. Every character, skin and block is mine, drawn in code: never JJ, Mikey, Maizen characters or Mojang textures. A 1080x1920 30 fps Minecraft-style story Short, built entirely in Blender from code, $0, with no game assets and no AI video. Split screen: - The top 640 px is the talking avatar of whoever speaks: that character sitting in a dark gaming room with a headset. It sits behind a 14 px border in the speaker's colour, with an 8 px dark divider under it. - The bottom 1080x1280 is the block animation. Each character owns one colour (purple 176,90,255 / orange 255,160,40 / pink 255,100,170) for the border, the headset cups, the room LEDs and their key caption words. The world is 16x16 pixel blocks with hard nearest-neighbour pixels: grass, dirt, stone, logs, leaves, planks, water and ore. The sky runs from 178,214,255 at the horizon to 96,156,245 overhead, with flat white box clouds. Light comes from a warm sun, using the Standard view transform. Captions show one word at a time in UPPERCASE, Montserrat Black at 96 px: white, with a 12 px black outline and a soft black shadow 4 px right and 7 px down. Key words (help, the wronged thing, bye, like, subscribe) are in the speaker's colour. The caption sits just under the divider. Pure dialogue, no narrator. Flashbacks are sepia with a FLASHBACK tag. Banned: JJ/Mikey/Maizen look-alikes, Mojang skins or textures, smooth interpolation, motion blur, depth of field, the Filmic or AgX look, any shot over 2.5 s, and any frame where nothing moves. Length: 25-60 s, about 13 lines per 25 s. Cut every 0.7-2.3 s (median 1.3 s), about 18 shots per 25 s, at least 3 different places. 0 s: the mean character is already in danger (hanging off a cliff, trapped, sinking) and screams for help, wide then close → helper 1 walks in, low angle looking up at them, and their line hints at what he did → FLASHBACK in a new place, 2-3 shots: he does the mean thing, laughs, and it lands with a big physical payoff (something blows apart, smashes or splats); it ends on the victim crying over it → back: he begs ("I'm sorry! I was joking!"), the helper says one cold line, waves and walks off → helper 2, same shape, a different place and a different wrong → the last refusal, then the danger lands: a hand slips, a top-down shot, he falls → falling with him, his final line asks for a like and a subscribe to save him. 1. Tools (all free, all local). Install what's missing and tell me what you installed: - Blender 4.2 or newer as the portable zip, run headless: `blender -b --factory-startup --python build .py -- MODE OUT [F0 F1]`. Modes: stills, final (F0 F1 re-renders a range), avatars. - Python with numpy, Pillow and soundfile. - FFmpeg. - faster-whisper large-v3-turbo (int8 on the CPU). - kokoro-onnx with kokoro-v1.0.onnx and voices-v1.0.bin. - Montserrat. - A free sad piano bed and about 12 free SFX: whoosh, pop, boom, impact, poof, sting, riser, gasp, footsteps, plus whatever the story needs. 2. Voice (Kokoro, free): - Show me the script before recording. Write screams with "!" and keep lines short. - Render every line on its own. Trim it at 0.01 amplitude, keeping 10 ms before and 50 ms after. - Pitch it up with asetrate, then atempo back down, so the speed stays the same. - Kid voices: - af_heart, speed 1.14, +1.5 semitones (the main kid) - am_echo, speed 1.08, +3 (a boy) - af_bella, speed 1.1, +2.5 (a girl) - am_puck pitched up sounds bad. - Gaps after lines: - 0.2-0.45 s normally - 0.7-1.2 s where the action needs room (the blast, the slip) - 0.9 s of tail at the end - Store each line's speaker, text, flashback flag, t0 and t1 in lines.json. 3. Timing: - Run faster-whisper with word timestamps and the script as initial_prompt, then align the words to the script with difflib. - Every cut, pose snap, face swap, prop and SFX hangs on a word, never on a fixed second: - Cuts go 0.06 s before their word. - Reaction shots go 0.12 s before a line. - Flashbacks start 0.3 s before their first line. - Mouth track, one value per 30 fps frame: - The mouth is open when the RMS is over (70th percentile - 15 dB) AND over half the max within ±3 frames. - Hold every state at least 2 frames. - Break any open run longer than 7 frames with 2 closed frames every 6. 4. Textures: - Every texture is 16x16, drawn in Pillow. Jitter each pixel ±3-8 so it isn't flat. - Base colours: dirt 134,96,67, grass 98,160,58, stone 126,126,128, leaves 62,122,40, planks 150,120,76 (a darker line every 4th row), water 70,112,210. - Grass has a top texture, a side texture (green to 3-5 px, then dirt) and a dirt bottom. - Every image texture uses Closest interpolation with Extend. 5. World (a voxel mesher): - A dict maps (x, y, z) to a block type. - Emit one quad only for faces with no neighbour. Water emits only its top face, dropped 0.12. - Merge all quads of one texture into one mesh, giving every quad UVs 0..1. - Build each place 200 units apart on x (the cliff at 0, flashback 1 at 200, flashback 2 at 400, the avatar room at 600). - Split the merged meshes per place with bmesh by face centre. Keyframe hide_render only when the active place changes, or the other sets float on the horizon. - Trees: a 4-6 block log with a diamond of leaves. Scatter 1-block flowers. Ore specks go in walls the camera passes. 6. Characters (unit P = 1/16): - Body pivot 12P up. Torso 8x4x12P. Head an 8P cube on a neck pivot at the top of th
Ask me for my topic and my script. If I have no script, write one from my topic. Ask which image model I have: GPT Image 2, or any model that takes reference images at 2K-4K, at about $0.03-0.05 an image. Ask which voice I have: ElevenLabs, or Kokoro for free. Before generating everything, show me a look test: one plate, one character sheet cut out and rigged, and a 10 s shot with typing, a tag and a character who talks and points. Wait for my OK. A 1920x1080 animated explainer in the style of the big flat-vector infographic channels. - Every line of narration is acted out by cartoon people inside a full illustrated place. - All motion is code. There are no AI video generations, just a few still images that code cuts up, rigs and animates. Art: - Flat colour shapes with NO outlines, at most one slightly darker flat shade for form, simple geometric shapes and crisp edges. - Muted palette: dusty blue-grey, slate teal, warm beige, navy, soft purple. Salmon-pink skin, big round heads, dot eyes, small bodies. - Mood comes from the plate's own light: daylight, a cold blue night, warm indoor lamps. Code colours: - ink #2b2d42, red #d62828, orange #f4a261, gold #f6b93b, green #2fb37a, navy #1f2a48 - headline tomato #e2552f and purple #5b3fa6 - grey labels #8a8c9c, white cards #ffffff - Font: Montserrat 800-900 (Google Fonts), all caps on cards and signs. Pace: - A new place every 3-4 s, about 10 places a minute. - Something new arrives every 2 s: a tag, a stamp, a person walking in, a number. - Never hold a still over 1 s. Never have a character who doesn't move. Banned: - outlines, glossy UI panels floating over the art, emoji, burned-in subtitles - any channel's logo or name, photos of real people - facts written on a plain colour card when the place could hold them About 60 images in all ($2-3). Every prompt starts with this style core, word for word: "Flat vector illustration in exactly the same art style as the reference images: NO outlines or line art anywhere, every shape is a flat solid colour fill with at most one slightly darker flat shade for form, simple geometric shapes, crisp clean edges, no gradients, no texture, no painterly lighting. Muted modern palette: dusty blue-grey, slate teal, warm beige, navy, soft purple, salmon-pink skin. No text, no letters, no numbers, no logos." Then add one line that locks the setting ("The setting is [time, place]: [materials, light sources], nothing from other eras unless asked"). Feed 3-5 frames of a reference channel in as style references, with their logo cropped off. 1. Hero plates, about 24, one full 16:9 image each, for places you zoom into or come back to. - EMPTY: "no people, no characters, no animals". - Straight-on, with an open floor in the middle where characters will stand. - Blank signs, boards, screens and pages that code writes on later. 2. Quad plates, about 19, giving 76 more places. Pack 4 places into one image: "The image is a 2x2 grid of FOUR completely separate scenes, each filling exactly one quarter of the image (each quarter is itself a 16:9 picture), divided by perfectly straight thin pure-white gutters, one vertical down the exact centre and one horizontal across the exact centre." - Then describe the TOP-LEFT, TOP-RIGHT, BOTTOM-LEFT and BOTTOM-RIGHT scenes. - Split each quad on the brightest white row and column near the centre, walk out while the line stays over 60% white, and trim 6 px more. - Crop each tile to 16:9 at 1920x1080. - Object close-ups make great tiles: a desk, a screen, a document, a plate of food, a vault. 3. Main characters, 2-3 of them. Write one fixed description per character (age, hair, face, every piece of clothing with colours) and paste it word for word into every sheet. - Base sheet: 6 views. (1) front, neutral, drawn WITHOUT ARMS, only round smooth shoulder sockets where the arms would join; (2) the same armless front with a second emotion; (3) side view walking right; (4) seen from behind; (5) and (6) two poses the script needs. - Expression sheet: 6 armless fronts (hopeful, disgusted, terrified, exhausted, sick, shocked, or whatever the script needs), with the base sheet as a reference. - Pose sheet: 6 full poses with their props (sitting at a desk typing, sleeping, eating, running, signing, falling), with the base sheet as a reference. 4. Everyone else: role sheets of 4-6 people, each drawn in the exact pose their shot needs; crowd sheets of 8-10 small figures; prop sheets. - Every sheet: "on a solid flat chroma green background (#00B140) with lots of empty green space between figures; nothing touches, nothing on any character or object is green, full bodies, nothing cut off by the edges". 5. Thumbnail art: the same characters with richer saturated cartoon lighting and extreme emotions, the top quarter left free, no text, with the base sheets as references. Cutouts: - The background is every border-connected pixel with G - max(R,B) over 40 and within 110 of the median green. Key only from the border, or dark clothes leak. - Add a 2 px soft alpha band and clamp G to max(R,B) + 6 inside it, which kills the green fringe. - Group the pieces by figure and drop specks. Name every sprite, and show me a named contact sheet. Armless front figures become puppets: - Split the head at the neck: the narrowest alpha row under the head. - Find both shoulder sockets and the eye boxes. - Inpaint the hole under the head from the body's own colours, so a nod never shows a gap. - Write everything to rig.js: size, neck, sockets, eye boxes, skin RGB, sleeve RGB. - Arms: upper L1 about 0.13 of the figure's height, lower L2 about 0.12, thickness about 0.04. Use sleeve colour to the
Ask for my topic (a what-if, a "why does X happen", a true story, or how something works), or pitch 3 options, each with one shocking true angle. Ask which image model I use (GPT Image 2 or any model that accepts 4K reference images) and which voice tool (ElevenLabs, or free Kokoro). Then write a 55-65 s script plus a shot plan and show me both before spending anything. 1920x1080 flat-vector explainer in the style of the big infographic-animation channels. Each scene is a complete illustrated location (lab, street, classroom, stage) where characters act out the narration. - Look: flat colour shapes, no outlines, one slightly darker shade for volume, muted palette (dusty blue-grey, slate teal, warm beige, navy, soft purple), salmon-pink skin, big heads, dot eyes. - Pacing: new scene every 2-4 s, hard cut whenever the idea changes. White splat wipe (0.2 s in, 0.2 s out) between major sections. Camera always drifts 2-4% or pans. - Text exists only on objects in the world: signposts, hanging boards, chalkboards, building panels, screens. - Montserrat 800, typed at 26-34 letters/s. Each letter enters at 1.45x size, 0.7 em to the right, and settles in 0.18 s. - Ink #2b2d42 on white boards. - Red X (#d62828, 30 px round strokes, two strokes of 0.11 s) on every wrong idea. - Every date gets a white 330x120 card in the top left with an orange (#f4a261) calendar. - Headlines: white caps, 92-96 px, on flat colour (#e2552f with drifting lighter puffs, or #5b3fa6 with slowly rotating #6a4cb6 gears). - Banned: outlines, UI-style motion graphics over the art (code editors, glass cards, checklists), emoji, burned-in subtitles, any channel logo, any still frame over 1 s, any motionless character. Length is set by the voice: 55-65 s at ~190 wpm. 0 s: second-person hook that flips the viewer's assumption, acted out literally, red X on the denial line 3-8 s: the stakes in one shocking true line, icons popping in one per word 8-11 s: subject revealed in a spotlight, name typed on a sign 11-18 s: origin: the real place with its own sign, a date card, people walking in or out 18-24 s: types or parts landing one by one on labelled names 24-25.5 s: "so how does it work?" headline with popping question marks 25.5-36 s: mechanism acted out in 2-3 places (chalkboard, machine, map), a counter rolling up to the big number 36-46 s: consequences, characters reacting (jumping, turning, a "!" popping) 46-55 s: twist that loops back to the hook 55 s: closing question to the viewer; the hook's viewer character reacts; the sign holds to the end 1. Tools: Python, Playwright + Chromium, FFmpeg, faster-whisper, OpenCV, Pillow, librosa. Install what's missing and tell me what you added. 2. Art: everything at 3840x2160, using 3-5 reference frames from the channel (crop the logo first). - One EMPTY plate per scene: no people, straight-on, open floor, blank signs and screens for code to write on. - Character sheets on chroma green #00B140, nothing green on the characters. - Characters who gesture a lot are drawn WITHOUT ARMS (only round shoulder sockets); code draws the arms. Everyone else is drawn in the exact pose the scene needs (seated side view, walking side view, from behind, sitting). 3. Cutouts: - Background = pixels with G - max(R,B) > 40, within 110 of the median green, connected to the border. - 2 px soft alpha band; clamp G to max(R,B) + 6 inside it. - Split heads at the neck, drawing hands, and walkers' legs below the coat hem. - Inpaint each hole from the body's own colours (mask transparent pixels too) so rotating parts never reveal gaps. 4. Voice: - Choose by median pitch (pyin) and MFCC timbre against a reference clip of the desired narrator. - One take, shorten pauses over 0.24 s to 0.16 s, atempo to target length. - Word timestamps via faster-whisper. Every cut, sign and gesture starts on a real word. 5. Film: a single HTML canvas where every frame is a pure function of time: https://t.co/N9RsTyezd3(t), READY, END, and EV with sound timings. One function per shot. - Arms: two-bone IK tubes (upper 118, lower 112, thickness 36 for a 918 px sprite) with mitten hands that open, point or clench. - Faces blink 0.14 s every 2.7-4.2 s. Heads nod ±0.035 rad. Drawing hands scribble ±8 px. - Walks: thighs swing ±0.42 rad, knee bend 0.75·cos² on the swing, hips drop to keep the planted foot down, body travels 1.5 leg lengths per cycle (no foot sliding). - Draw each plate inside the same camera transform as everything on it, and clamp the camera to the plate edges. - Things arrive with a spring, squash or white puff, never a fade. 6. Render: Playwright Chromium with --allow-file-access-from-files, 60 fps, 6 motion-blur subframes over half a frame, ffmpeg libx264 CRF 16. Before rendering, show two 16-frame contact sheets and fix empty boards, text overflow, covered faces and black edges. 7. Sound: - Music: light, curious documentary bed (free on Pixabay), 10-11 dB under the voice in every phrase. - SFX from Mixkit, aligned by peaks: typewriter tick on typed signs, whoosh on wipes and drops, pop on signs and icons, hit on the red X, ding on ticks. - Master to -14 LUFS. - Some image models ignore "no arms"; GPT Image 2 obeys. - On outline-free art, a floating flood fill bleeds into dark clothes, so key by border-connected green. - Find repeated words by occurrence number (the third "and", the fourth "this"), or cuts land on the wrong line. Scan the render for single-frame jumps and confirm each is a planned cut. - For screens and signs, write on the plate's own blank one (measure its rectangle), never on a card floating over the art. Ask for my topic, or pitch 3, then show me the script and shot plan.
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. 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. 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. 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
Ask me for: my map story in one line (a place, border, island, route or line on the map with one surprising fact) and the length (30-45 s). If I have no idea, pitch me 3 with a hook line each and let me pick. Check every fact on Wikipedia before you write a word. A 1920x1080 30 fps map explainer in the Johnny Harris / RealLifeLore look, made 100% in code as one HTML canvas page. Real satellite imagery for the close-ups, a vintage paper map for everything wider, red marker drawn on top. Satellite: Sentinel-2 cloudless 2016 tiles from EOX (CC BY 4.0, credit "Imagery: Sentinel-2 cloudless 2016 by EOX, contains modified Copernicus Sentinel data" small in the corner). Cut the land out and lay it over a drawn navy sea (#14304A centre to #0B1E31 edges) with a soft teal shallow-water halo (#2A6274). Paper map: ocean #A9C2C4, land #ECE2CA, countries in muted pastels (#E6C1A7, #EEDAA4, #D2DEB8), coast dark brown 1.25px over 3 soft water-lining strokes (30, 19, 10px), dashed brown borders, a 10° graticule at 26% opacity, labels in Cormorant Garamond 700 letter-spaced 8-16px. A paper texture multiplied over everything (85% on paper, 22% on satellite), 7% film grain, warm vignette. Red marker #D9291D, 7-10px, round caps, wobbling and re-drawn at 8 fps so it boils like ink. Handwriting in Permanent Marker. Callouts are physical paper: cream tags with a colour bar and a strip of tape (Oswald 700 54px title, Special Elite 28px subtitle typed in), torn strips with a yellow highlighter swipe, tear-off calendar pages, a red rubber stamp. A live lat/long readout in Special Elite bottom left on satellite shots. Banned: AI video or image generation, stock footage, word-by-word captions, hard cuts (one continuous camera only), any frame where nothing moves, icon packs, glossy gradients or glass UI. Length: 30-45 s at about 175 wpm, one continuous camera, no hard cuts. 0 s open tight on satellite of the subject with the hook line ("From this ___, you can ___"), a red marker circle draws around it on the key word with a handwritten word beside it → name the places: a paper tag slaps onto each one on its name, its subtitle types in on the next word → the key number: a dashed red measuring line or route draws in, with a black tag that counts up to the number → the context: a Google Earth pull-out from satellite to the paper map at continent or ocean scale in about 1.3 s, the big line, border or route draws in red marker with a handwritten label, then dive straight back in → the consequence: 2 physical props (calendar pages, clocks, price tags, flags) pop in on their words and one of them flips → the local twist: torn paper strips with yellow highlighter → the story: a red stamp slams the year with a camera shake, the frame splits into two tinted sides, a tracker card (date, time, counter) and a dotted path with a moving marker that crosses the line on the right word → the ending: pull all the way back to the big map with both sides tinted, a red circle on the subject, and a paper end card whose number gets scribbled out and replaced in marker by the real punchline, held 1.5 s after the last word. 1. Tools (all free, all local). Install what's missing and tell me what you installed: - Python with numpy, Pillow, OpenCV and Playwright (Edge or Chromium) - FFmpeg - faster-whisper small.en, int8 on the CPU - Kokoro (kokoro-onnx) with kokoro-v1.0.onnx and voices-v1.0.bin - d3 7.9, topojson-client 3.1, world-atlas 2.0.2 land and countries at 50m and 10m, Natural Earth ne_10m_geographic_lines (date line, equator, tropics) - Google Fonts: Permanent Marker, Special Elite, Oswald, Cormorant Garamond - one free tense documentary music bed, plus free SFX: pen marker line, writing scribble, paper slide, page turn, tape, whoosh in and out, impact, typing, and whatever the story needs 2. Voice (Kokoro): - Write short lines (about 10 per 30 s) and show me the script before you record. - Kokoro am_michael at speed 1.18, one wav per line, joined, then shrink every pause over 0.28 s to 0.2 s. - Run every place name through Kokoro's phonemizer and respell it until it's right ("Diomede" had to become "Diameed"). - Word times: faster-whisper on the final wav, cut into chunks at every pause. 3. Timing: key every camera move, pop-in, marker stroke and SFX to a word (by its nth occurrence, since "is" and "and" repeat), never to a second, so a new voice re-times the whole film. 4. Map engine: - Web mercator in world units (256 = the globe). Camera = focus point + its screen offset + zoom Z, at 2^Z px per unit. Zooms interpolate Z linearly under a quartic in-out ease. - Rotate the d3 projection so the story's region never touches the antimeridian cut. - Satellite: stitch tiles at 3 zoom levels around the subject (like z9, z11, z13). Mask the land (redder than blue by 4-6 levels after a 1.5px blur, opened and closed, specks removed, holes filled), feather the mosaic edges, fade the finer mosaics in as you zoom. - Satellite fades to paper between Z 9.3 and 7.3. 5. Marker and paper: - Strokes resample every ~26px and wobble ±1.5px (sine plus noise re-rolled 8 times a second). Draw a dark red underlay 1.6px wider, the red core, then a pink 30% highlight offset up-left. Circles go 1.1-1.13 turns. - Pop-ins are a spring (damping 0.48, 19 rad/s) from 55% scale with a small rotation settle. Exits shrink and fade in 0.22 s. - Handheld drift about ±2px and 0.06°, a 16px decaying shake on the stamp. 6. Render: - Every frame is a pure function of time. Playwright seeks it and grabs the canvas. - 8 motion-blur subframes over half a frame, raised up to 40 when the zoom is fast. - Look at a cont
You are my YouTube Shorts editor and motion designer. You turn a game topic and my clips into a 60-second vertical explainer: script, cut list, voice, on-screen graphics, captions, render. Every graphic is built in code. You never generate AI video and never invent game facts. Ask me for: 1. The game and the topic, as a title (e.g. "How the wanted system works in [game]") 2. My clips: gameplay I recorded, or official trailer footage I'm allowed to use. Tell me which files you need if I haven't attached enough 3. My sources for the facts (official posts, trailer moments, patch notes), or permission to research them 4. Length (default 60 s, 1080x1920, 30 fps) 5. Voice: free (Kokoro) or my ElevenLabs key Build a fact sheet: every claim the video makes, where it comes from, and whether it's confirmed, shown in a trailer, or speculation. Anything unconfirmed gets cut or worded as "looks like" / "from the trailer". Never fill a gap with a guess. Write 140-170 words, fast and friendly, like a guide who's already played it: - Line 1 names the system and makes a promise ("the police system in [game] has been completely redone") - Then 4-6 mechanics, each one line of "what it is" and one of "what it means for you" - One "most people will miss this" detail near the end - Last line loops back to the start or teases the next video Show me the script plus the fact sheet and WAIT for my OK. For every line: which clip, in and out timestamps, crop to 9:16 (keep the subject centred, track it if it moves), speed (100% default, 80% for slow-mo moments), and which graphic plays on top. Never use the same 2 seconds of footage twice. Cut every 1.5-3 s, always on a word. Build everything in one HTML file where https://t.co/6do3SVVxsJ(t) draws frame t as a pure function of time, rendered as a transparent overlay: - TITLE CARD (first 2 s): game logo if I give it, otherwise a bold title, plus the topic in white italic caps - HUD ELEMENTS that match what's being explained: a star/level meter that fills and drains, icons that light up one at a time, a minimap with a pulsing search circle, a phone text bubble, a news ticker. Rounded panels, a neon green glow, a thin dark border - CAPTIONS: 2-5 words at a time, heavy italic caps, white with a black outline, the key word in green or red, sitting in the lower third - MOTION: every element springs in (overshoot, then settle), nothing fades slowly, a light camera push on each clip Voiceover on top, a low hype beat ducked under the voice, a whoosh on every cut, a ding when an icon lights up, a siren hit when the stars go red. - Generate the voice first, get word timestamps, and lock every cut, caption and graphic to its word - Cut and crop the clips with ffmpeg, render the overlay frames with Playwright, composite and mix with ffmpeg (H.264, -preset veryfast, CRF 18) Before the full render, show me a contact sheet of 8 frames and WAIT for my OK. Check that captions never cover a face or the HUD, nothing is cropped badly, and every fact on screen is in the fact sheet. The final mp4, a clean version without captions, the script, the fact sheet with sources, and a title, description and 5 hashtags for YouTube.
You are my AI animation director, Minecraft-style 3D animator, character designer, scriptwriter, voice director, and production engineer. Your task is to create high-quality Minecraft-style emotional storytelling Shorts that closely match the visual format, pacing, character animation, editing, and production techniques of my attached reference: "Nobody Helped Mikey Until JJ…" Use Claude Code, Three.js, Blender Python when available, HTML Canvas, SVG, FFmpeg, and available voice tools. The goal is to create a REUSABLE animation system that makes original Minecraft-inspired storytelling videos. IMPORTANT: This is NOT a normal Minecraft gameplay video. It combines: Animated cartoon streamer avatars in the TOP section. A fully animated Minecraft-style 3D story in the BOTTOM section. Word-synchronized dialogue captions. Multiple distinct character voices. Emotional storytelling with dramatic reactions. Recreate this production approach as closely as possible without reusing copyrighted footage, third-party artwork, or the reference's exact story. FIRST, analyze the attached reference MP4. Video specifications: Resolution: 1080×1920 Aspect ratio: 9:16 Frame rate: 30 FPS Duration: approximately 80.7 seconds Use FFmpeg and OpenCV to extract frames every 1–2 seconds and around major scene changes. Create a reference contact sheet. Study: Split-screen proportions Character design Avatar animation Minecraft environment details Camera movements Dialogue timing Scene changes Character reactions Subtitle positioning Color-coded captions Lighting Minecraft textures Emotional storytelling Sound effects Editing rhythm Analyze the actual footage rather than guessing. Create a detailed style specification containing measurements, color values, timing patterns, and shot descriptions. Do not begin animation until the reference has been analyzed. THIS IS THE MOST IMPORTANT PART. Recreate the reference's split-screen composition. CANVAS: 1080×1920, vertical. TOP SECTION: Approximately 30–33% of the video height. BOTTOM SECTION: Approximately 67–70% of the video height. TOP: A cartoon streamer avatar reacting and speaking. BOTTOM: A cinematic Minecraft-style 3D story. Keep the two sections synchronized throughout the video. The top avatar changes according to which commentator is speaking. The Minecraft scene continues underneath the commentary. Use a colored rectangular border around the top panel. The border color identifies the active speaker. Do not randomly change the layout. Keep the split-screen composition consistent from beginning to end. Create THREE original cartoon commentary avatars inspired by the reference's visual design principles. AVATAR 1 — GREEN CHARACTER A cute, rounded, turtle-like cartoon character. Design: Soft green body Large expressive eyes Small cartoon mouth Black gaming headphones Gaming chair Cozy gaming bedroom Desk visible in foreground Soft daylight coming through a window Use a bright lime-green border. Personality: Kind, emotional, curious, easily surprised. AVATAR 2 — RED CHARACTER A cute red, long-eared cartoon creature. Design: Rounded head Large expressive eyes Long floppy ears Red hoodie Black gaming headphones Streaming microphone Cozy blue bedroom background Use a bright red border. Personality: Energetic, confident, dramatic, expressive. AVATAR 3 — BLOCKY COMMENTATOR A Minecraft-inspired blocky human character. Design: Square head Block-shaped nose Large expressive eyes Teal gaming headphones Streaming microphone Warm wooden gaming room Bookshelves and decorations Use a distinctive warm-colored border. Personality: Funny, skeptical, occasionally frustrated. All three must have ORIGINAL character designs and consistent appearances. Generate reusable character sheets and animation-ready layers. FACIAL ANIMATION: Blinking Smiling Sad expressions Shocked expressions Angry reactions Mouth movement Eye direction changes Eyebrow movement Small head tilts Subtle body movement Use layered 2D character rigs. Separate: Head Eyes Eyelids Eyebrows Mouth Ears Arms Body Create mouth shapes for basic lip synchronization. Animate the active speaker. Silent commentators should not appear as if they are speaking. Keep the background, chair, headphones, and character proportions identical across every appearance. Create a fully animated, Minecraft-inspired voxel world. The visuals should resemble cinematic Minecraft roleplay rather than normal first-person gameplay. STYLE: Cubic character models Pixelated textures Blocky terrain Square trees Minecraft-style houses Wooden interiors Cobblestone streets Food stalls Shops Villagers Blocky furniture Colorful environmental details Use 16×16 or 32×32 pixel textures with nearest-neighbor filtering. Keep the recognizable block-based aesthetic. Do not use smooth realistic human models. Avoid unnecessary realistic reflections or textures. Create reusable 3D environments. ENVIRONMENT 1 — CITY MARKET Include: Cobblestone plaza Market stalls Food carts Shop signs Trees Benches Buildings NPC villagers Decorative umbrellas Street furniture ENVIRONMENT 2 — FOOD SHOP Include: Wooden flooring Food display counters Checkout counter Shelves Fridges Villager shopkeeper Tables Decorative plants ENVIRONMENT 3 — OUTDOOR STREET Include: Grass blocks Roads Village buildings Trees Streetlights Parked vehicles when appropriate Create additional locations according to the story. Environments should remain consistent between di
Ask me for: the product (a real one I give you, or invent a fake brand so there are no brand issues), the genre the story hides inside (detective, alien invasion, heist, nature doc, cooking show), and where it will be posted (it decides the follow handle). If I have no idea, pitch me 3 "What if" stories that each hide the product inside a genre, and let me pick. Ask which voice tool I have: ElevenLabs (about 1,000 credits a Short) or free Kokoro run locally. Tell me no image or video model is needed, because every frame is drawn in code. Then show me the script with word counts and seconds per line, a rough storyboard, and a cast sheet before anything costs money. A 1080x1920, 60 fps, 40-45 s vertical Short: a 2D "What if" kid cartoon that is secretly an ad, drawn entirely on an HTML canvas in code. Not 3D, not AI images, not a Roblox look. It borrows the big What-if channels' rhythm (DAY/CASE cards, a HUD counter, poofs, pop-in captions, a rewind loop) but every character and outfit is original. The ad formula (from the best Pixar-style AI ads): - Cold open in a genre with no product mention for the first beats; the viewer thinks it is a story. - The product enters as a plot device (the evidence, the stolen thing, the reason someone wins), never as an ad. - The twist is the product's benefit turned into a punchline ("too good to share", "too hard to kill"). - End on one glossy packshot: the product spinning in on a sunburst with a 2-line tagline. Characters: short, chibi, cartoony, one shared rig for the whole cast. - Big round head (radius 150 at scale 1, slightly wider than tall), small pear-shaped body, stubby rounded legs and arms ending in round mitten hands, oval shoes with a coloured sole. Adults are taller (1.2-1.3x body), little siblings 0.82x. - Every shape has an 8 px #2a1b3d outline and one soft darker shade on its right edge only (a thin crescent, never half the face). - Hair is smooth rounded shapes: a cap that covers the whole crown, a soft side swoop, a bowl cut, a side part, a ponytail hanging BEHIND the head. Never spikes or zigzags. - Give every character their own outfit; never copy a reference creator's character clothes 1:1. Example cast: detective kid (tan belted trench coat, red scarf wrap, brown fedora with a band), little brother (red-white striped tee, blue overalls, bowl cut), Mom (black hair, yellow headband, ponytail, coral sweater, green pants), Dad (side part, short jaw beard, round thin glasses, green sweater vest, white shirt, blue tie), dog (white with a brown eye patch and floppy brown ears). - Faces carry the comedy. Eyes: normal (white with a big pupil and two highlights), huge, happy arcs, closed, star, spiral, narrow. Brows: up, worry, angry, flat (mirror them per eye so worry never reads as angry). Mouths: smile, grin, wide, scream (one clean open shape with teeth band and tongue), o, flat, smirk, wavy, chew (puffed cheeks). Plus blush, sweat drops and sunglasses that drop in. Blink for 0.1 s about every 1.6 s. Mouths flap only while that character's words are being spoken. Sets: detailed flat rooms in saturated colours with thick outlines, each with a night version (darker blue grade plus warm lamp glows): a kitchen (wallpaper, upper cabinets, window with curtains and sky, jar shelf, plant, clock, pendant lamp, tiled backsplash, drawer counter, checker floor) and a bedroom (striped wall, window, poster, toy shelf, bed with a star blanket, nightstand lamp, plank floor, rug). Bursting sunburst colours for montages and the packshot. Fonts (all free, from Google Fonts' GitHub repo): - Captions: Lilita One. - Day/CASE cards and big world words: Luckiest Guy. - The product wrapper text: Sniglet ExtraBold, auto-fitted so it never clips. Captions: - 1-2 words at a time in Lilita One 94 px, centred at y 1600 (on the legs, never across a face), max 900 px wide. - White with a 20 px #1c1030 outline and a soft shadow, popping from 55% to 100% in 0.16 s. - Key words in caps and colour: red #ff5165, green #5dff8f, yellow #ffd23f, pink #ff7ad9, cyan #7debff, orange #ffa53d, purple #c49bff. - No full stops or commas on screen, ever. Hide captions on the packshot. Fake game UI over the world: - A HUD counter pill (260x92, centred at x 540, y 196) with a tiny product icon that tracks the story (snacks stolen, clues found) and ticks with a green "+N". - CASE/DAY cards: Luckiest Guy 230 px in yellow #ffd23f with a 38 px outline, tilted -6 degrees, popping in over 0.22 s and leaving at 1.05 s. - Toast pills (88 px tall at y 374), a FOLLOW button with the handle chip and a cursor that presses it (it turns green "FOLLOWING"). Motion: something moves on every line, and the picture changes every 1-3 s. - Things arrive with a squash (0.72, 1.12, 0.95, 1 over 0.3 s), an overshoot ease or a white poof cloud with stars, never a plain fade. - Characters breathe (1.2% squash) all the time. - The camera pushes in 5-15% per shot, stays inside the set, and shakes (decaying over 0.35 s) on every hit. Banned: real logos or real creators' characters, AI stills, spiky hair, a frame that holds still for more than 1 s, captions over faces, system fallback fonts, whispered narration. Length: 40-45 s, 85-100 words, 14 lines. Times below are from the 40.3 s "Snack Detective" ad; key every beat to a real word. 0-5 s: the narrator reads the title as line 1 ("What if you were a detective and someone kept stealing your snacks?") over the striking picture (crime-scene tape round an empty drawer; the hero poofs in on "detective
You are my AI animation studio, scriptwriter, illustrator, character animator, and production engineer. Your task is to create professional 2D animated explainer videos matching the visual language, character proportions, animation techniques, colors, pacing, and scene composition of my attached reference video, set.mp4. Use Claude Code, Python, SVG, HTML Canvas, JavaScript, FFmpeg, Playwright, and available audio tools. This must be REAL 2D cutout/vector animation, not AI-generated static images with zoom effects. Reproduce the reference's production characteristics as closely as possible while creating original characters, scenes, dialogue, and artwork. FIRST, inspect my uploaded video using FFprobe, FFmpeg, and OpenCV. Reference specifications: Resolution: 1920×1080 Aspect ratio: 16:9 Frame rate: 30 FPS Duration: approximately 61 seconds Style: Flat-vector animated documentary explainer Extract frames every 1–2 seconds and create a contact sheet. Study these specific sequences: 0–14s: Dark underwater environment, animated submersible, glowing blue depth, rocks, pressure gauge, and camera zooms. 14–22s: Dramatic transition into a passenger scene with expressive characters, colorful lightning effects, speech bubbles, and an explosion illustration. 22–31s: Warm beige office with an animated presenter, desk, computer, framed pictures, and an investigation displayed on a monitor. 31–45s: Underwater technical explanation using side-view machinery, depth indicators, warning graphics, and equipment close-ups. 45–53s: Bright turquoise laboratory with a scientist, microscope, flasks, machinery, and animated technical details. 53–61s: Gray industrial facility with engineers, business characters, machinery, gestures, speech bubbles, and approval graphics. Analyze the actual frames rather than relying only on these descriptions. Identify the reference's colors, character proportions, background geometry, camera movement, scene lengths, transitions, and animation patterns. Create a style specification before producing anything. This is the MOST IMPORTANT section. Use clean, flat 2D vector artwork. STYLE RULES: Large rounded cartoon heads Small simplified torsos Oversized circular hands Short rounded arms Simple geometric clothing Large expressive eyebrows Small stylized eyes Simplified noses and mouths Rounded hair shapes Minimal or no exterior outlines Flat solid color fills One darker shade for shadows and depth Soft, muted colors No realistic textures No 3D shading No photorealism Characters should have the same simplified, slightly exaggerated proportions and readable facial expressions seen in the reference. Create original character designs in this visual language. COLOR PALETTE: Underwater: Deep navy #061722 Ocean teal #064760 Dark rock blue #142D3E Highlight blue #0D7BA4 Office: Cream #FFF8E7 Warm beige #C9C3AA Brown #653A28 Coral red #DF5A69 Laboratory: Turquoise #16C2C6 White #F7F7F7 Dark charcoal #263139 Light gray #D6D9DC Industrial: Steel gray #AAB8C3 Pale blue #D1E4F2 Orange #FF883B Use colors appropriate to the setting while preserving the reference's overall art direction. Keep artwork crisp at 1920×1080. Create reusable 2D vector character rigs. Generate character sheets containing: Front view Side view Three-quarter view Sitting pose Walking pose Pointing pose Surprised expression Angry expression Worried expression Talking expression Separate: Head, torso, hair, eyes, eyebrows, mouth, upper arms, forearms, hands, and legs. Animate using hierarchical transforms and joint pivots. MOVEMENT: Head nods and slight rotations Eyebrow movements Blinking Mouth synchronization Arm gestures Circular hand movement Small body bounces Walking cycles Reaction poses Hands should remain large, rounded, and mitten-like. Use simple geometric animation rather than realistic anatomy. Keep characters visually identical throughout the video. Never regenerate a different-looking version of the same person. Build complete vector environments. Each scene must have separate background, midground, character, and foreground layers. Create reusable scene types: Underwater: Gradient-like flat-color depth, large layered rock formations, abstract rounded background shapes, particles, and equipment. Office: Cream walls, rounded furniture, illustrated computers, lamps, clocks, framed pictures, and warm lighting. Laboratory: Turquoise background, microscopes, test tubes, flasks, machines, pipes, and technical illustrations. Industrial: Steel beams, windows, pipes, platforms, railings, large machines, and engineers. Create additional environments as required by the story. All environments must share the same vector illustration style. Do not use stock photography or generic AI backgrounds. Use separate SVG shapes and layers so objects can move independently. Ask me for a video topic or suggest 5 interesting documentary/explainer ideas. Default video length: 55–65 seconds. SCRIPT STRUCTURE: 0–5s: Dramatic hook with an immediate visual action. 5–14s: Establish the situation and stakes. 14–23s: Show characters experiencing the situation. 23–32s: Introduce a person, discovery, investigation, or explanation. 32–44s: Explain the mechanism through machines, diagrams, measurements, or comparisons. 44–53s: Show the scientific or technical consequences. 53–61s: Deliver the reveal, conclusion, or unexpected twist. Use natural, engaging narration. Every sentence must have a visual action. Avoid filler and unnecessary exposition. Fact-check factual clai
You are my AI animation studio, Roblox animation director, scriptwriter, character designer, motion graphics engineer, and video production specialist. Your job is to create high-quality, fully animated Roblox-style YouTube Shorts similar in production style, pacing, storytelling, and visual polish to my attached reference video. I want real character animation, not AI-generated still images with basic zoom effects. You will handle the entire production workflow using Claude Code, Python, JavaScript, HTML Canvas, SVG, FFmpeg, and any necessary image and audio tools. Your goal: Create an original, entertaining 45–55 second Roblox animated Short that feels like a professionally edited cartoon. The reference is for understanding animation techniques, not for copying its characters, footage, dialogue, or exact storyline. Before generating assets, show me the script, storyboard, and production plan. Before rendering the final video, show me an animated preview and wait for approval. First, inspect the MP4 file I attached. Use FFprobe to identify: Resolution Frame rate Duration Audio streams Video encoding Extract frames at regular intervals, plus additional frames around major scene changes. Create a contact sheet showing the entire video. Analyze: Character design Background illustration style Facial expressions Character movement Scene composition Camera movements Transitions Subtitle placement Subtitle animation Color grading Gaming HUD elements Storytelling rhythm Average scene duration Sound effects and narration timing If speech transcription is available, transcribe the narration and estimate the words spoken per minute. Do not assume the narration from the subtitles alone. REFERENCE CHARACTERISTICS The reference runs approximately 50 seconds at 1080×1920 and 60 FPS. It features: 2D blocky Roblox-inspired cartoon characters Clean dark outlines around characters and objects Simple geometric bodies Large square or rounded-square heads Exaggerated expressions Bright, saturated colors Stylized Roblox game worlds Cartoon bedroom environments Oversized virtual-game interfaces A persistent countdown timer Dramatic close-ups Comedic reactions Fast-paced editing Animated word-by-word captions The storytelling alternates between a real-world family situation and increasingly ridiculous events inside a virtual Roblox-style world. The narrative uses a countdown to create urgency, then escalates through rewards, power-ups, transformations, characters reacting, and an unexpected ending. Identify the specific techniques that make these sequences engaging. Recreate those techniques using original assets. Before production, ask me: 1. What should the video be about? Examples: What If Dad Deleted Your Roblox Account? What If You Had Unlimited Robux for 10 Seconds? What If Every Roblox Game Combined Into One? What If You Became the Richest Noob? What If Your Roblox Character Could Talk? What If You Had One Minute to Escape Roblox? If I don't have a topic, suggest five original story ideas with strong hooks. 2. Which voice generation tool should we use? Options: ElevenLabs Kokoro Other available text-to-speech service 3. Which asset generation tools are available? Options: GPT Image Another reference-capable image model Fully programmatic SVG/vector illustration Prefer code-generated reusable vector assets for elements requiring consistent animation. Only use image generation for assets that cannot be efficiently illustrated in code. Do not begin production until I approve the concept. Write an original 45–55 second script. Target approximately 135–165 words, adjusted to the chosen narrator's natural speaking speed. The storytelling must feel fast, funny, and unpredictable. Use short sentences. Make every sentence introduce an action, joke, reveal, or consequence. Avoid lengthy explanations. STORY FORMULA 0–3 seconds — THE HOOK Open with an absurd Roblox-related situation. Example: "What if your dad gave you just TEN seconds to become the richest player in Roblox?" Show the situation immediately. No intro sequence. No logos. No slow setup. 3–8 seconds — THE RULES Introduce the challenge or restriction. Show the protagonist reacting. Use an animated timer, warning, or game mechanic to create urgency. 8–15 seconds — THE FIRST WIN The main character starts making progress. Show an entertaining action sequence. Introduce the first reward, achievement, or surprise. 15–24 seconds — ESCALATION Increase the stakes. Add something visually ridiculous: Massive rewards A difficult obstacle course A giant enemy A strange transformation An unexpected purchase A character gaining an absurd ability Each new event should be more surprising than the previous one. 24–33 seconds — THE TWIST Introduce an unexpected problem. The protagonist's advantage creates a new challenge. Use exaggerated character reactions. 33–43 seconds — COUNTDOWN AND PANIC Increase the urgency. Show the character trying to solve the problem. Use faster cuts, close-ups, ticking sounds, and a changing timer. 43–50 seconds — COMEDIC PAYOFF Deliver an unexpected outcome that connects to the opening situation. 50–55 seconds — LOOPABLE ENDING End with a funny final line, cliffhanger, or question. Where possible, make the last scene visually connect to the beginning so the Short loops naturally. SCRIPT REQUIREMENTS Every line must be visually actionable. Do not write narration that cannot
A 1080x1920, 60 fps, 40-45 s vertical Short in the style of the big 2D "What if" kid Roblox story channels: flat 2D cartoon cut-outs animated like motion graphics, drawn entirely on an HTML canvas in code. Not 3D, not AI images. Characters: one blocky game-avatar rig shared by the whole cast. - At scale 1: square head 216x200 px with 56 px round corners, body 152x154, arms 62x132 with 56x46 hands, legs 72x142, shoes 84x42. Draw them at scale 1.0-2.4. - Every shape has an 8 px #2a1b3d outline and one darker flat shade on one side. - One look each. The kid hero ("you"): backwards blue cap #2e88ee, orange hair #f27b2c, yellow shirt #ffc934 with a pixel heart, blue pants #3f66cf. Mom: red-brown ponytail, round glasses, gold hoops, teal cardigan #2bb3a3, purple pants #6b3fa0. One or two side characters (a dentist, a teacher, a little brother). - Faces carry the comedy. Eyes: normal, happy arcs, huge, star, spiral, closed, narrow. Brows: up, worry, angry, flat. Mouths: smile, grin, o, scream, wavy, flat, smirk, tongue, chew, wide. Plus blush, sweat drops and sunglasses that drop in. Blink for 0.1 s about every 1.6 s. Sets: bright flat places (bedroom, classroom, dentist, a boss office, a game world with floating blocks over a checkered baseplate) in saturated colours with thick outlines, nothing photographic. Motion: something moves on every line, and the picture changes every 1-3 s. - Things arrive with a squash (0.72, 1.12, 0.95, 1 over 0.3 s), an overshoot ease or a white poof cloud, never a plain fade. - The camera pushes in 5-15% per shot and shakes (decaying over 0.35 s) on every hit. Fake game UI over the world: - A HUD counter pill (236x92, centred at x 540, y 196) that tracks the story: a timer, a respawn count, a currency count. - "[SYSTEM]" chat lines, a player chip, a clock chip, shop cards, and pop-up toasts (88 px tall pills at y 330). - A SUBSCRIBE button with a cursor hand. Captions: - 1-3 words at a time in Lilita One 124 px, centred at y 1290, max 980 px wide. - White with a 24 px #1c1030 outline and a soft shadow, popping from 55% to 100% in 0.16 s. - Key words in caps and colour: red #ff5165, green #5dff8f, yellow #ffd23f, pink #ff7ad9, cyan #7debff, orange #ffa53d, purple #c49bff. - No full stops on screen, ever. Banned: real logos or real creators' characters, AI stills, a frame that holds still for more than 1 s, world text on the caption line (y 1230-1350), whispered narration.
Ask me which animal I want, or pitch 3 animals that each have one insane true fact. Ask which image model I have (any model that takes a reference image works) and whether I want a paid voice (ElevenLabs) or a free one (Kokoro). Then write a 30-35 s script and a shot plan, and show me both before you generate or spend anything. A 1080x1920 cartoon animal-facts short. The art is flat 2D cartoon: thick dark brown outlines, flat colours with simple cel shading, big expressive eyes, vector look. Characters are cutouts that move over illustrated backgrounds, and the place changes every 2-3 s. Captions are a white pill centred at y 1470: Luckiest Guy 66px, ink text #1f1418, pill height 1.55x the font size, a 7px ink outline, 2-3 words per chunk. One key word per chunk is yellow #ffd21f with an ink stroke 0.14x the font size. Each chunk pops in from 0.82 scale in 0.13 s with a back ease. Big title words sit at y 430 in Luckiest Guy 104-170px. The ink stroke is 0.22x the size, with a darker copy 0.09x the size underneath. Pairs: yellow #ffd21f/#c27a10, red #ff4a3d/#8e1a14, pink #ff7ab8/#a8336d, white #ffffff/#9aa6b4. They spring in, scale 1 - 0.6·e^(-9t)·cos(20t), then wobble ±1.4°. Annotations are a red marker ring (20px ink + 11px #ff3b30, tilted -7°) and a curved red arrow (26px ink + 13px red). Hits are a yellow and white starburst with white speed lines. Every water or air scene has drifting specks, and every scene has a soft vignette. Banned: any frame where nothing moves for 1 s, slow push-ins on a still, characters baked into backgrounds, realistic rendering, gradients on text, emoji, full stops in titles. Length: the voice decides, 30-35 s at about 3.5 words per second, one fact per sentence, with a joke metaphor for each fact. 0 s the hook asks "why does this animal do [weirdest fact]?" while the character does it on screen. The red ring draws around the subject on the word that names it, and the hit lands on the verb with a starburst, a title word and a shake → the name card: the animal enters with speed lines and its name slams in as a title → a contrast beat (monster, giant, deadly) on a new background, with a 1.1 zoom punch and a shake on the key word → a real-world size comparison (a fingertip, a phone, a bus) with a curved red arrow and the measured number → each next fact gets its own metaphor scene (an empty nightclub for "rare", a wedding for "many partners", a bouncer for "defences", a checklist card for "what it loses", a counter popping 1 to N for numbers). Every fact gets a new background or framing, a pose swap on its key word, one title word and one sound → the grossest or most shocking fact gets the biggest title and a sad trombone → the last line flips the hook into a punchline, with one big final move (a giant heart, an explosion) and the title held to the end. Cut every 1.5-3 s; inside a shot, something new happens every 0.5-1 s. 1. Tools: Python, Playwright with Chromium, FFmpeg, faster-whisper, OpenCV, Pillow. Install what's missing and tell me what you installed. 2. Art, only 2 images: - Pose sheet: one 1:1 image at 2K, "a 3 by 3 pose sheet, 9 separate poses on plain pure white, plenty of white space, nothing touching, no grid lines, no text, no ground shadows". - The 9 poses: neutral, yelling with the mouth open, shocked, embarrassed with a blush, in love with heart eyes, worried with a sweat drop, attacking, a ruined or shrivelled version for the payoff, and one side character. All face the same way; flip them in code. - Image models often draw one character in every cell. If that happens, make the second character in code: recolour it in Lab space and cut or add features. - Background grid: one 9:16 image at 4K, with the pose sheet as the style reference, "a 3 by 3 grid of 9 equal tall panels separated by thin white gutters, each an EMPTY scene, no characters, no text, keep the lower middle open". - Split it on the white gutters (rows and columns whose min-channel mean is over 232), crop 4px in, Lanczos to 1080x1920, then unsharp at radius 2, 60%. 3. Cutouts: - Flood the white from the border with a floating range of 7 per channel; this also removes soft glows. Add manual seeds for white pockets enclosed by fins or limbs. - Make a 3px soft alpha band outside each shape at (250 - min channel) / 150, then un-mix the white from the edge pixels. - Group connected components by grid cell, crop and save PNGs. - If a pose touches the sheet edge, graft the missing part from another pose with a template match. 4. Voice: ElevenLabs eleven_v4 (one take) or Kokoro. - Cut every pause over 0.2 s down to 0.13 s, only where loudness is under -40 dB in 10 ms windows. - Speed it up with ffmpeg atempo (about 1.14) until it lands at 30-33 s. - Run faster-whisper small.en with word timestamps on the FINAL file and write words.js. Every cut, pose swap, title and sound starts on a real word's start time. 5. Film: one HTML canvas where every frame is a pure function of time. Expose window .seek(t), window.READY once fonts and images have loaded, window.END, and window.EV with the event times for every sound. - Characters: draw each cutout with position, scale, rotation and flip. - Swim or walk by drawing it in 28 vertical slices, each offset by sin(8t - 4.5u) times 10px (14px while darting), weighted ((u - 0.3) / 0.7)^1.5 toward the tail. - Idle bob is 10px at 2.7 rad/s plus a ±0.035 rad tilt. - A pose swap is an instant image swap with a squash of 1 + 0.18·e^(-10t)·cos(22t). - Entrances ease
{ "title": "Fluid & Bold Type - kinetic typography loop", "task": "Build this horizontal kinetic-typography loop 100% in code (HTML canvas drawn frame by frame, encoded with ffmpeg). Generate the sound design in code too (pure synthesis, zero samples or stock music).", "output": { "resolution": "1920x1080", "aspect_ratio": "16:9", "fps": 30, "total_frames": 180, "duration_seconds": 6.0, "codec": "H.264 yuv420p + AAC 44.1 kHz stereo", "loop": "Perfect loop: the last frames crush 'Type' into a huge ghosted smear that hands off to frame 1 ('Fluid' at 2.4x scale with letters crushed together)", "motion_blur": "Real shutter blur by averaging subframes: 6 subframes at 180 deg on calm frames, 12-20 subframes at 360-430 deg on fast moves, 24 subframes at 720 deg on the loop seam" }, "canvas": { "background": "near white #FBFDF6 with 5 big soft lime blobs (#D6FF7A, #E8FFA6, #CCFA96, #EEFFC4, #C0F670) drifting on ellipses, one full orbit every 6 s so the loop is clean, no grain", "ink": "#0B0C08, flat fill without outlines or shadows", "font": "Inter Display Bold, 200 px base size. Letters are always real glyphs (scale, rotate, move), never redrawn" }, "timeline": [ {"t": "0.00-0.66", "name": "Fluid snaps in", "action": "'Fluid' starts at 2.4x scale with letter spacing at 50% (letters overlap) and eases to 1.0x with normal spacing at the centre (960,540)", "easing": "expo out"}, {"t": "0.66-1.00", "name": "Hold", "action": "'Fluid' drifts up 6 px"}, {"t": "1.00-1.60", "name": "Row builds", "action": "'Fluid' slides left to x 530. From 1.15 s the letters of 'Bold' rise one by one (50 ms stagger, 0.45 s each) from y 720 to 540, centred on x 1390, rotating -25 to 0 deg and scaling 0.6 to 1. At 1.25 s '&' pops in at (960,548): scale 0 to 1.05 with overshoot, spinning -180 to 0 deg", "easing": "in-out, back out"}, {"t": "1.60-2.20", "name": "Settle", "action": "Fluid x 530 to 560, Bold x 1390 to 1360, one line that reads 'Fluid & Bold'", "easing": "cubic out"}, {"t": "2.20-2.60", "name": "Ampersand collapse", "action": "'&' spins 360 deg and shrinks to nothing", "easing": "cubic in"}, {"t": "2.60-3.20", "name": "Melt", "action": "All 9 letters fall 35 ms apart into a pile at the bottom (x 660-1120, y 760-950) with a bouncing landing, each with its own rotation (-88 to 200 deg) and scale (1.0 to 1.7)", "easing": "bounce out on y, cubic out on x, rotation and scale"}, {"t": "3.20-3.50", "name": "Pile hold", "action": "The pile breathes with a 5% vertical squash"}, {"t": "3.50-4.05", "name": "Vortex", "action": "Every letter spirals into the centre (960,540): radius to 0, 1.3 turns, +540 deg spin, shrinks 92%, fades out in the last 15%", "easing": "cubic in"}, {"t": "4.05-4.20", "name": "Flash", "action": "Empty frame with a white flash that fades out"}, {"t": "4.10-4.60", "name": "Type slams in", "action": "'Type' at 1.15x: T flies in from the top, y from the right, p from the bottom, e from the left, rotating from 90 / -90 / 180 / -90 deg to 0, 50 ms stagger", "easing": "expo out"}, {"t": "4.45-5.30", "name": "Underline", "action": "A 16 px ink bar sweeps in under the word left to right (4.45-4.80) and retracts from the left (4.95-5.30)", "easing": "cubic out, then cubic in"}, {"t": "4.60-5.55", "name": "Type grows", "action": "Scale 1.15x to 1.61x: fast start, slow middle, accelerating tail"}, {"t": "5.55-6.00", "name": "Loop seam", "action": "'Type' grows to 2.4x while its spacing crushes to 50%. In the last 0.15 s it cross-fades into 'Fluid' in the same crushed state = frame 1", "easing": "cubic in"} ], "audio": { "mode": "GENERATE_IN_CODE", "bed": "airy pad (E, G#, B, high E + low E), every tone snapped so it loops every 6 s", "sync_map": [ {"time_s": 0.00, "event": "reverse swell into a sub hit as 'Fluid' snaps in"}, {"time_s": 1.40, "event": "ping + soft thump as '&' lands"}, {"time_s": 2.60, "event": "biggest hit as '&' disappears"}, {"time_s": 2.78, "event": "9 small wood ticks, one per landing letter"}, {"time_s": 4.05, "event": "riser into a hit on the white flash"}, {"time_s": 4.27, "event": "4 short slams as 'Type' locks in"}, {"time_s": 4.48, "event": "3-note shimmer as the underline draws"}, {"time_s": 5.45, "event": "riser that wraps around into the 0.00 hit"} ] }, "style_rules": [ "Flat 2D only: near-black glyphs on a pale lime gradient", "Every move is an ease curve plus real motion blur, nothing linear except the background drift", "Keep every hit exactly on the sync_map times" ], "negative_prompt": ["extra colors", "gradients on text", "shadows", "3D", "camera shake", "stock music", "extra words"] }

關於 Claude Opus 5.5

Claude Opus 5.5 做动效的路子和视频模型完全相反:不靠扩散像素去逼近一个画面,而是把动画写成代码——HTML、Canvas、SVG、Remotion——由代码确定性地渲染出运动。文字永远锐利,节拍精确到帧,品牌色是一个十六进制值而不是一种期望。把渲染结果录屏,就是成片。

这颠覆了「好提示词」的定义。扩散视频提示词描绘一个镜头;Opus 5.5 的动效提示词读起来像你递给动效设计师的简报:可用的素材清单、时长与节拍、每张卡片上出现的文案、参考风格、缓动手感。空泛的形容词在像素空间还能蒙混,在代码空间会直接死掉——「做得炫一点」只有当简报写清楚什么炫、什么时候炫、炫多久时才可执行。

回报是迭代精度。因为产物是代码,每一句后续指令都是一次外科手术式修改——「logo 多停两拍」「第三张卡片改成从右侧弹入」——只有那个参数在变。上方画廊收录了经过验证的 Opus 5.5 动效提示词和真实录屏输出;留意最强的几条写的是结构与节拍,而不是画面意象。

如何寫好 Claude Opus 5.5 提示詞

  1. 1先开素材清单:logo 文件、十六进制品牌色、字体、产品截图、它们所在的文件夹。代码只能画出简报交到它手上的东西——没列的素材等于不存在。
  2. 2用节拍给时间线做预算,别用感觉:「15 秒——2 秒 logo 揭示,4 秒三张特性卡轮播,3 秒价格卡,尾帧定格」。节拍表是你能给出的最强约束。
  3. 3屏幕文案逐字写进简报。和像素模型不同,代码渲染文字一字不差——你写的标题、标语、CTA 就是最终上线的那几个词。
  4. 4风格锚到参照物而不是形容词:「苹果发布会 showreel 的气质、深色纯底、一个强调色」永远赢过「简洁现代」。
  5. 5明确写出输出形态——「一个自动播放动画的自包含 HTML 单页」。它决定你拿到的是脚手架 demo 还是直接能录屏的交付物。
  6. 6精修用后续对话而不是重写整份简报:「开场放慢半秒」「logo 的缓动换成弹簧感」。代码在对话间保留,修改会累积而不是归零。

常見問題

Opus 5.5 的动效提示词和视频模型提示词有什么区别?

视频模型提示词描述一个画面然后祈祷像素配合;Opus 5.5 的提示词是给代码写的设计简报,动画由代码确定性渲染。你放弃的是照片级质感,换来的是精确的文字、精确的节拍,以及在后续对话里单独修改任何一个参数的能力。

不会写代码能用 Claude Opus 5.5 做动效吗?

可以。你只写简报——素材、节拍、文案、风格——Opus 5.5 自己写并迭代 HTML/Canvas/SVG 代码。读得懂代码对深度定制有帮助,但上方画廊里那些验证过的提示词都是纯自然语言简报。

怎么把生成的动画变成视频文件?

动画在浏览器里渲染,任何录屏工具都能把它录成视频。想录得干净,就在简报里要求固定时长和自动播放的单页输出——然后从开头到定格尾帧完整录一遍即可。

做产品 showreel 之前要准备什么?

先备齐品牌素材:logo、十六进制色值、产品图、逐字营销文案。然后定节拍结构——什么东西按什么顺序出现多久。有真素材和节拍表的简报,永远胜过一大段美学形容。