PROMPT LIBRARY
MiniMax H3 prompts, each with the video it made.
Each prompt is quoted word for word from the model maker's guide or a creator's post, next to the output it produced. Open a card for the full prompt.
MiniMax H3 makes video with native stereo sound from text, images, video and audio. A clip runs 4 to 15 seconds at 24 fps, with the shorter side at 768 pixels by default and 2K through MiniMax's hosted H3-Regenerate-2K, in ratios from 21:9 to 9:16. Dialogue is stable in 11 languages, among them Chinese, English, Japanese, Korean and Spanish. It has two modes: one takes zero, one or two keyframes, and the other takes references, up to 9 images, 3 videos and 3 audio clips with 12 files in total. The H3-Base weights are public, so it also runs locally; MiniMax's apps and API add its hosted prompt rewriter and 2K pass. The prompts below are ones creators shared on X and examples from MiniMax's model card, each shown with its video.
Sources: MiniMax H3 on GitHub · Model card on Hugging Face · H3 prompt writing guide
How creators write MiniMax H3, Wan 3.0, Seedance 2.5 and Kling 4.0 prompts, measured →
Updated 2026-11-03 · 310 prompts, newest first
Other model pages: GPT Image 2.6 · Nano Banana 2.1 · Seedance 2.5 · Kling 4.0 · Wan 3.0
Blue combat boots in cartoon-style closeup
Glowing thread weaves a fantasy world across tapestry
Cracked metal orb rests among rubble
Digital assembly builds a character from wireframe layers
Woman with dark bangs and pearl drop earrings
Creature encyclopedia UI with cursor selecting creatures
Painterly white-haired swordswoman with glowing blue sword
Rem scoops fallen desktop icons back into place
Carved limestone relief map reveals drama credits
Anime girl in plaid boots slides into punk MV
Living room FPV dives into TV toast scene
Live hallway footage restyled as 1970s cartoon
Anime barista makes latte art from storyboard
White clay cat copies street cat pose
Anime gamer girl character sheet with turnarounds
Laundromat turns into crayon doodle night sky
Girl rides red bicycle past summer shop
Samurai Soul fights a colossal energy giant
Anime girl in orange headphones, amber eye glance
Same crossing scene from three camera angles
Anime closeup of turquoise eyes under navy bangs
Anime heroine dances through yellow punk music video
Cherry blossoms, palace roofs and N Seoul Tower
Woman in white dress getting ready at mirror
Woman with windblown hair on a golden hour beach
Woman raises iced tea bottle under sunny sky
Cyber ice rebel with silver hair and neon splashes
Woman in white walks along tropical beach
Woman rides red dragon over snowy mountains
Woman and dog at windy green window
Golden peacock forms from light in grand hall
Silhouetted swordswoman under a giant purple moon
Boots walk past vintage convertible at desert station
Woman walks coastal path at golden hour
Letters morph into butterfly, then bloom into realism
Kaiju squid rampages through city seen from airplane
Selfie vlog through lakeside town at golden hour
Woman in white parka among snowy peaks
Smiling woman shows hair serum bottle in bedroom
Glam beauty close-up with gold pearl earrings
Black luxury sedan city ad ending with a kiss
First person firefight at a night military base
Empty lot becomes glass skyscraper in aerial drone shot
Woman on moving train roof in city
Heroine among giant marshmallows and lollipops in Candyland
Young woman showing serum bottle in bathroom
Hands tying white running shoes at golden hour
Nothing matches. Try a shorter word.
MiniMax H3 prompts worth starting from
The most liked prompt in each category on this page, in full, with what makes it work. Copy one, then change the subject.
Tom restores fallen desktop icons one by one
The prompt locks the reference desktop as the exact first frame and lists every stationary element, so only three icons move. A static camera, a timed beat sheet and icon by icon pickup steps keep the physical actions in order and stop the icons from teleporting.
Video Prompt : Create a 10-second horizontal 16:9 video using the provided Tom and Jerry macOS desktop image as the EXACT first frame and visual reference.
IMPORTANT:
Preserve the reference desktop exactly as shown, including the warm sunset living room, wooden floor, sofa, window, books, popcorn box, scattered popcorn, macOS menu bar, left-side application grid, bottom Dock, and all desktop elements.
Keep Tom and Jerry's exact recognizable classic cartoon appearance, proportions, colors, facial features, expressions, poses, and original 2D animation style. Do not redesign, modernize, photorealize, or change their visual style.
Begin with Tom running dramatically toward Jerry on the RIGHT side while Jerry runs away carrying a piece of cheese.
CAMERA:
Static front-facing camera.
One continuous shot.
No camera movement.
No zoom.
No pan.
No cuts.
TIMELINE:
0–1.5s:
Tom continues chasing Jerry across the living-room floor.
Tom looks determined and slightly frustrated.
Jerry runs quickly toward the RIGHT while holding the cheese.
Jerry looks back at Tom with a mischievous smile.
Tom's arms, legs, ears, tail, and facial expression move naturally with exaggerated classic cartoon animation.
1.5–2.5s:
Jerry suddenly changes direction and runs toward the LEFT.
Tom reacts dramatically and tries to stop himself.
Tom's feet slide across the wooden floor and he accidentally kicks the scattered popcorn forward.
A sudden exaggerated cartoon air movement travels toward the LEFT side of the desktop.
The air movement accidentally knocks THREE desktop icons loose:
Gmail
Discord
Microsoft Teams
The three icons physically leave their original positions in the RIGHTMOST desktop-icon column.
They rotate separately through the air and fall into the open desktop area above the Dock near the center-left.
Each icon lands separately with a small bounce.
Their original grid positions are now visibly empty.
2.5–3.2s:
Tom freezes after realizing what happened.
His eyes widen.
Jerry also stops for a moment and looks at the fallen icons.
Tom slowly looks toward the viewer with an embarrassed expression.
Jerry looks at Tom mischievously as if amused by the situation.
Tom immediately decides to fix the problem.
3.2–7.8s:
Tom restores the three fallen icons ONE AT A TIME.
FIRST — Gmail:
Tom carefully reaches toward the fallen Gmail icon.
He picks up Gmail individually with one paw.
The icon must visibly leave the desktop surface.
Tom carries Gmail across the desktop toward its original empty position.
He carefully places Gmail back into its exact original location.
The Gmail logo and label remain unchanged.
A soft click is heard.
Only after Gmail is completely restored does Tom reach for the next icon.
SECOND — Discord:
Tom reaches down and individually picks up the fallen Discord icon.
He carries Discord toward its original empty position.
He carefully presses it back into the correct grid position.
A second gentle click is heard.
Jerry watches Tom while standing nearby with the cheese.
THIRD — Microsoft Teams:
Tom reaches down once more.
He picks up Microsoft Teams individually.
He carries it back to its original position.
He carefully releases it into the empty grid position.
A third gentle click is heard.
Show all three separate pickup-and-replace actions clearly.
IMPORTANT PHYSICAL ACTION RULES:
Every moving icon must visibly:
leave its original position
fall through the air
land on the desktop
be physically picked up
be carried by Tom
be placed back into its original position
Do NOT:
- restore icons automatically
- teleport icons
- make icons snap back from a distance
- move multiple icons simultaneously
- scoop multiple icons together
- stack icons
- duplicate icons
- morph logos
- change labels
Tom must restore Gmail, Discord, and Microsoft Teams separately.
His paws must remain anatomically consistent with the original cartoon style.
No extra paws.
No extra fingers.
No stretched limbs.
No distorted characters.
Jerry remains nearby and reacts naturally while Tom fixes the desktop.
7.8–10s:
After placing Microsoft Teams back into position, Tom quickly returns to the RIGHT side.
Jerry runs a short distance away while still holding the cheese.
Tom straightens himself and tries to regain his dignity.
He looks toward the restored icons.
Then he looks directly at the viewer with an innocent, slightly embarrassed expression.
Jerry briefly looks back at Tom with a mischievous smile.
Tom pretends nothing happened.
Hold the final pose for the remaining moment.
The final desktop must closely match the original first frame.
DESKTOP PRESERVATION:
Only Gmail, Discord, and Microsoft Teams are allowed to move.
Everything else must remain completely stationary:
Chrome
Safari
YouTube
Spotify
Slack
Zoom
Notion
Microsoft Word
Microsoft Excel
Instagram
Reddit
Netflix
Steam
Dropbox
Google Drive
The macOS menu bar must remain stationary.
The Dock must remain stationary.
The wallpaper and all furniture must remain stationary.
Do not change application logos or labels.
Do not duplicate any icons.
AUDIO:
Light playful cartoon piano and pizzicato music.
Fast playful rhythm during Tom's chase.
A short cartoon sliding sound when Tom loses balance.
A brief airy whoosh as the three icons are knocked loose.
Three soft landing sounds as the icons fall.
Three distinct gentle clicks as Tom restores Gmail, Discord, and Microsoft Teams.
Music briefly pauses during Tom's shocked reaction, then resumes with a playful rhythm.
No dialogue.
No subtitles.
No text overlays.
VISUAL QUALITY:
Premium cinematic cartoon animation integrated into the existing macOS desktop.
Keep Tom and Jerry clearly 2D while maintaining the cinematic lighting of the environment.
Natural squash-and-stretch animation.
Smooth character motion.
Clean cartoon outlines.
Consistent colors.
Natural shadows.
Subtle floor reflections.
Realistic interaction with the environment.
FINAL FRAME:
All three icons must be restored to their exact original positions.
Tom and Jerry remain on the RIGHT side.
Tom looks embarrassed but tries to appear innocent.
Jerry remains mischievous.
No missing icons.
No duplicate icons.
No changed logos.
No extra characters.
No extra objects.
No camera movement.
No cuts.
The final frame should closely match the original reference image. Twenty shot character detail montage, face revealed last
A numbered shot list with time blocks and one abrupt camera move per shot builds three connected action bursts. Faces stay out of frame until shot 20, and hard match cuts link similar shapes without morphing.
Use @[character reference] for the character's identity, anatomy, costume, colors, materials and visual style. Adapt body-region cues to its actual anatomy, replacing absent features with existing equivalents. Invent no anatomy or accessories. Select signature inserts from its most recognizable details. Rigid parts rotate; flexible materials bend or stretch.
A 15-second character-detail montage: exactly 20 shots joined by hard cuts. An empty, seamless warm-white studio background; soft directional light reveals the surfaces. Shots 1-19 are tightly cropped body details and rear angles: keep the face, eyes and facial profile entirely outside the frame, including reflections. Shot 20 is the first and only face reveal. No full-body view.
Three connected bursts: a sharp head turn with the body bracing; a coiled arm opening into a strong sweep and torso twist; one short step into a planted pivot, finishing with a sudden look into camera. Carry body state across cuts without resetting to neutral. Existing hair, clothing and attachments react naturally. Keep the character active, with clear articulation and weight shifts.
Jump between distant body regions. Mix brief visual punches with longer moving-detail reads. Use one abrupt camera move per shot except locked-camera inserts 4 and 11, where the character supplies the speed. Vary height, angle and direction. Brake moves briefly for legibility while the subject keeps moving. Local whips stay tightly cropped; changes of body region use hard cuts. Connect selected cuts through similar shapes or motion vectors, without morphing. Real-time action, no slow motion or freeze frames.
0-3s, shots 1-5. A concealed head turn triggers the first burst:
1. Extreme upper-back detail: a crash push catches its surface shifting as a sharp head turn begins entirely above the crop.
2. Fingers: a short downward snap catches them curling into a fist as the body braces.
3. Rear three-quarter shoulder detail: a tight camera arc catches the shoulder drawing back behind the turning head.
4. Locked floor-level foot detail: the grounded support foot braces, its raised heel pressing down firmly without taking a step.
5. Rear neck detail: a short upward tilt catches the nape stretching with the turn; stop below the back of the head, keeping the jaw and face excluded.
3-9s, shots 6-13. The coiled arm opens into a forceful sweep:
6. A signature upper-body detail: a fast diagonal slide catches an existing seam, joint or distinctive contour responding as the arm draws inward.
7. Elbow: a sharp lateral track catches the folding joint tightening the arm's coil.
8. Supporting knee: hard match cut from the bent elbow to the knee's similar angle; a short camera drop catches the leg compressing under the same preparation.
9. Wrist: a quick shallow orbit catches it turning outward to initiate the release.
10. Upper chest: a short punch-in catches the chest opening as the arm drives outward beyond the crop. Let this moving surface read slightly longer.
11. Locked hand detail: fingers spring open and the hand sweeps rapidly across the fixed frame. Catch one crisp, readable phase of the spread fingers before they leave the crop.
12. Upper back: a sharp sideways slide continues that screen direction as the shoulder rolls through the sweep and the torso twists.
13. A distinctive torso or waist detail: a close diagonal move catches its physical response to the twist, such as an existing fold pulling taut or an attached piece swinging. Give the material response a clear moving read.
9-13s, shots 14-19. One short step and a pivot build toward the reveal:
14. Thigh of the other leg: an abrupt low tracking move catches it driving the single short step out of the twist.
15. Advancing foot: a brief downward whip catches it landing; show clear floor contact.
16. Ankle and lower-leg detail: a fast side slide catches the landed leg accepting weight as the ankle flexes.
17. Waist: a tight orbit catches the torso rotating above the planted-foot pivot. The foot supports the turn rather than sliding across the floor.
18. Back-of-head contour: a short lateral whip catches the start of the return turn from behind. Keep only the rear contour in frame, never the facial profile.
19. A signature shoulder detail: a sharp diagonal slide catches its follow-through beneath the turning head, which stays completely above the crop. Cut at the peak of this movement.
13-15s, shot 20. First face reveal:
20. Hard cut to a frontal face close-up as the character sharply completes the turn toward the lens. An explosive short push reveals the entire face clearly for the first time and stops tightening before cropping out its features. Direct eye contact lands while hair, clothing or natural body follow-through remains alive. Keep the whole face readable through the end; finish on living motion, without a held pose or another cut.
No added props, scenery, text, graphic overlays or split screens. Rem scoops fallen desktop icons back into place
A reference image locked as the exact first frame, plus a second-by-second beat sheet, keeps the desktop layout and character fixed while the icons react to her skirt. Naming the only objects allowed to interact and banning camera moves and cuts stops the scene from drifting.
Use the reference image as the exact first frame. 16:9, locked-off camera, one continuous take. Preserve Rem’s face, short blue hair, black-and-white maid outfit, and original anime art style consistently throughout. Keep the desktop layout and Japanese menu bar unchanged.
0–2s: Rem stands on the right, happily swaying gently from side to side while lifting the edges of her skirt slightly with both hands. As she turns, the hem accidentally catches the blue-purple desktop background. The background reacts like a soft fabric curtain, sending a horizontal ripple across the screen. The two folders and one screenshot icon on the left wobble, slide downward, fall to the bottom edge of the screen, and make a small playful bounce on impact.
2–3s: Rem immediately freezes. Her eyes widen and she quickly covers her mouth with both hands, staring in surprise at the fallen icons.
3–6s: She hurriedly leans toward the left and reaches out with both hands. She scoops up all three icons together and quickly pushes them back toward their original positions on the left. The icons behave like thin, lightweight magnetic tiles, snapping back onto the desktop surface. Two of them end up slightly crooked. Rem has no time to straighten them and quickly pulls her hands back.
6–8s: Rem returns to her original position on the right, folds her hands neatly in front of her, and tries to put on an innocent, nothing-happened smile. Her eyes secretly glance toward the two crooked icons on the left. The tips of her hair, hair ribbon, and skirt hem move gently in a light breeze. The desktop background settles completely, while the two icons remain slightly tilted for a subtle comedic ending.
The background music begins with a light, cute Japanese-style piano melody. When the icons fall, add a short descending slide sound and light impact sounds, then shift into a playful pizzicato rhythm.
Keep all movement fluid and expressions lively. Only the three desktop icons on the left may physically interact with Rem or the background. The top menu bar and bottom Dock must remain completely fixed. Preserve the original icon designs, filenames, and text exactly. No new characters. No camera movement, zoom, cuts, or transitions. Woman with dark bangs and pearl drop earrings
A PARAMETERS block at the top fills in the on-screen labels, and a timed shot list of thirteen cuts lays out the edit beat by beat. Naming the reference image as the sole identity authority keeps the character consistent through every cut.
PARAMETERS
CHARACTER_NAME = "[CHARACTER NAME]"
STATUS_TEXT_1 = "[ACTIVE]"
STATUS_TEXT_2 = "[MODE: ENGAGE]"
INTERSTITIAL_TEXT_1 = "[SYSTEM LINK]"
INTERSTITIAL_TEXT_2 = "[TARGET LOCK]"
INTERSTITIAL_TEXT_3 = "[FULL DRIVE]"
SUBTITLE = "[CHARACTER TITLE / CODENAME]"
PROJECT_TEXT = "[OPTIONAL PROJECT NAME]"
Create a 15-second premium cinematic character reveal trailer with exactly 13 distinct cuts, 24fps.
Use @[image1] as the sole authority for the character's identity, proportions, outfit, materials, colors and visual style. Never redesign the character. Adapt the environment, abilities, VFX and movement naturally to them.
The overall presentation should feel like a premium AAA action-game character introduction. Motion-design sections use a bold editorial language: oversized condensed typography, clean white or pale-grey fields, huge circular color blocks, thin technical diagrams, concentric rings, halftone textures, tiny interface markings, scan lines, geometric accents, sweeping brush or energy strokes and strong foreground/background parallax. The main palette shifts between black and white, icy cyan, electric blue, neon yellow and warm golden amber. In text-based graphic screens, the typography must appear clearly and be fully readable first, then the character, a cropped close-up or a silhouette may slide in front of it, overlap it or stand beside it. In some graphic-card backgrounds, also use cropped character close-ups or enlarged facial details as integrated design elements.
Use the text parameters above exactly.
CUT 01 | 0.00-1.10s | FACE
Extreme cinematic close-up. Calm focused expression, part of the face naturally obscured by hair, hand, mask or accessory. Shallow depth of field and subtle push-in.
CUT 02 | 1.10-2.20s | ACTIVATION
A signature effect ignites around the character's eyes, hands, body or equipment: glowing rings, symbols, particles, energy filaments, holographic nodes or character-specific effects. The gaze lifts as the effect rapidly peaks.
CUT 03 | 2.20-3.20s | ACTIVE CARD
Hard graphic-match into a bright editorial freeze-frame. Large vivid circular field behind the character, oversized black condensed typography, thin technical circles, tiny UI text, halftone detail and sharp diagonal graphic streaks. Show the typography clearly first, then let the character overlap it or enter beside it. A cropped facial close-up or upper-body close-up may be integrated into the background design.
Typography:
STATUS_TEXT_1
CUT 04 | 3.20-4.10s | ENGAGE CARD
Snap to a high-contrast black character silhouette over a bold horizontal accent-color typography band. The text should read clearly before the silhouette crosses in front of it or locks beside it. A large luminous brushstroke or energy ribbon sweeps diagonally across the foreground with strong motion blur.
Typography:
STATUS_TEXT_2
CUT 05 | 4.10-5.20s | WORLD
Smash cut into a cinematic environment suited to the character. Wide dramatic perspective as they immediately dash, leap, land or burst forward. Fast whip-pan or reactive tracking camera.
CUT 06 | 5.20-6.20s | DETAIL
Fast close or medium shot of one defining ability, weapon motion, gesture, transformation or acrobatic action. Sharp character, strong foreground streaking.
CUT 07 | 6.20-6.90s | INTERSTITIAL 01
Hard cut to a text-led graphic design screen. Huge condensed typography, layered circles, technical lines, scan lines, halftone textures, tiny data text and aggressive graphic wipes. The typography must be fully readable first, then bring in the character as a smaller integrated figure, cropped close-up or silhouette.
Typography:
INTERSTITIAL_TEXT_1
CUT 08 | 6.90-7.90s | IMPACT
Return to action from a new aggressive angle. Attack, dodge, ability release or interaction with a simple secondary threat. Low angle, side track, whip-pan or short orbit.
CUT 09 | 7.90-8.60s | INTERSTITIAL 02
Another editorial text card, cleaner and punchier, built from bold typography, diagonal slashes, accent-color bars, fine UI markings and strong parallax. Let the typography dominate first, then place the character beside it or pass a cropped close-up in front for depth.
Typography:
INTERSTITIAL_TEXT_2
CUT 10 | 8.60-9.60s | VELOCITY
New perspective continues the momentum. Character rushes toward camera, crosses close to lens, launches upward or rapidly changes direction while the environment stretches into controlled speed blur.
CUT 11 | 9.60-10.30s | INTERSTITIAL 03
Final text-based graphic card before the hero beat. Bold oversized typography, sweeping energy stroke, geometric overlays, concentric diagrams and a fast silhouette or cropped character insert. Ensure the text is legible before the character layer overlaps or frames it.
Typography:
INTERSTITIAL_TEXT_3
CUT 12 | 10.30-12.30s | HERO MOMENT
The character performs their most iconic movement, ability, strike, landing or confident stop. Briefly stabilize at the visual peak, then collapse the shot into a bright smear, silhouette frame or graphic-match transition.
CUT 13 | 12.30-15.00s | CHARACTER ID
Premium final identity card. Clean pale background with enormous translucent concentric circles, fine technical arcs, network nodes, subtle halftone texture, tiny graphic markings and restrained chromatic glints. Place the character in a dynamic full-body hero pose overlapping huge condensed typography. The name should register clearly before the character overlap becomes dominant. A large cropped close-up can also be subtly embedded into the background composition.
Behind them, add a large black circular ink-brush or emblem shape containing a ghosted visual motif associated with the character: creature silhouette, symbol, power form, weapon motif or abstract representation of their identity.
Primary typography:
CHARACTER_NAME
Secondary:
SUBTITLE
Optional:
PROJECT_TEXT
Allow the character to overlap the typography for depth.
EDITING:
Very aggressive premium game-trailer rhythm. Hard cuts, graphic matches, whip pans, freeze frames, impact flashes, directional smears, foreground wipes, silhouette transitions and short speed ramps. Every cut must introduce a clearly different composition or visual mode. The text-based interstitial cards should break up the action rhythm and keep the sequence from feeling like uninterrupted combat footage. No slow dissolves or long continuous choreography.
BGM DESIGN:
Create a fast, character-driven AAA game-trailer score tightly synced to the edit. Adapt genre and instrumentation to the character, building from restrained tension into punchy percussion, bass and rising intensity, with clear hits for cuts, typography, transitions and action beats. Peak at the hero moment and finish with a strong final reveal stinger. Integrate subtle UI ticks, whooshes, energy swells and sub impacts into the music. Anime girl in plaid boots slides into punk MV
A reference image is set as the identity authority, and every fixed design detail (eye colour, hair clip, plaid boots, outfit) is spelled out so the character holds across the whole clip. Each time segment is tied to a beat, and transitions are driven by body movement such as a boot skid or stomp, so the camera travels without random cuts.
Create a 15-second horizontal 16:9 high-energy 2D anime pop-punk music video featuring the exact heroine from the provided character reference. The entire sequence should feel like one continuous, tightly choreographed MV passage, with each movement naturally motivating the next transition. CHARACTER LOCK: Use the reference image as the absolute authority for identity, proportions, outfit, colors and illustration style. Preserve her short messy bubblegum-pink hair with sharp outward tufts, high side ponytail, pale cyan and mint lightning-shaped hair accents, large red cross hair clip with lavender backing, huge yellow eyes with thick dark outlines, small expressive mouth and warm blush. Keep her black cropped school-style jacket with oversized white cuffs, white shirt, loose bright-red necktie, red plaid mini skirt with white triangular hem accents, black belt, bare legs and enormous red plaid boots with dark straps and heavy rounded soles. Maintain her original stylized proportions: oversized head, compact torso, long simplified legs, huge boots, chunky hands, bold black outlines, flat saturated colors and minimal cel shading. Her personality is loud, cheeky, rebellious, playful and hyperactive. 0–2.5s — Start immediately with motion. She slides rapidly into frame from the left, one boot skidding across a clean white graphic floor. Her body leans forward from momentum while her ponytail, necktie and oversized cuffs trail behind. She stomps the second boot down, looks up and flashes a mischievous grin. The impact instantly transforms the white background into huge red-and-black plaid blocks as bold “PINK PANIC!” typography slams into the composition. 2.5–5.0s — Without stopping, she rebounds from the landing into a small hop and sharp half-turn. Camera swings around her from front three-quarter view into a fast side-tracking shot as she runs two exaggerated cartoon steps. Her giant boots hit the floor on alternating beats, leaving hand-drawn stars, crosses, tape strips, scribbles and comic impact marks behind each step. Her pink hair and red tie bounce wildly with the movement. 5.0–7.5s — She kicks one leg sideways on the beat and uses the momentum to spin. The red plaid pattern from her boot stretches outward into a giant rotating graphic plane, seamlessly pulling the entire frame into an abstract 2D punk world made from black brush strokes, red plaid, white paper shapes, pink lightning bolts and rough photocopy textures. She continues directly into a bouncy dance phrase: head whip → shoulder hit → side step → small kick → fast turn. Keep every pose exaggerated and cartoonishly expressive. 7.5–10.0s — Her turn accelerates and the camera follows closely around her. The environment constantly recomposes around the choreography: checker patterns slide sideways, hand-drawn crosses rotate, torn-paper strips snap open and oversized comic typography appears briefly on drum hits. She lands hard with both boots apart. A huge hand-drawn “BAM!” bursts underneath her feet and physically shakes the surrounding graphic elements. She immediately rebounds upward with a cocky smile rather than holding a static pose. 10.0–12.3s — Follow her continuing movement into a connected sequence of character details. As she whips her head sideways, track past her bright yellow eye → red cross hair clip → swinging pink hair → bouncing red tie → plaid skirt → massive plaid boot. Do not present these as disconnected beauty shots. Each close-up should be motivated by the same continuous body movement and fast camera travel. The boot hits the floor at the end of the sequence, sending a red plaid shockwave across the frame. 12.3–15.0s — Ride the plaid shockwave back into a full-body view. She takes two confident bouncing steps forward, abruptly pivots and finishes facing camera with her weight shifted onto one leg, shoulders slightly forward and a cheeky rebellious grin. Her ponytail and necktie continue swinging from the turn while the background erupts into layered red plaid, black brush X marks, white stars, pink lightning and rough hand-drawn graphics. The elements rapidly organize themselves into a bold 2000s punk magazine-cover composition. Large “PINK PANIC!” typography appears behind her silhouette on the final beat. End on her strong recognizable character shape without changing her design. VISUAL STYLE: Strictly match the reference artwork: early-2000s-inspired 2D cartoon/anime aesthetic, thick clean dark outlines, flat saturated colors, minimal cel shading, simplified anatomy, exaggerated proportions and bold graphic expressions. Keep the intentionally handmade TV-animation feeling rather than modern polished anime rendering. Use red plaid as the main graphic motif, combined with black-and-white manga shapes, pink lightning, crosses, stars, scribbles, torn paper, photocopy grain, tape graphics and rough ink marks. MOTION: Fast, bouncy and highly rhythmic. Use strong key poses, limited-animation timing, overshoot, smear frames, exaggerated anticipation and cartoon impact frames. Her giant boots should create the strongest beat accents. Hair, ponytail, tie, skirt and cuffs should react continuously to her movement. Transitions must grow naturally from character actions: boot skid → plaid expansion → kick → rotating graphic plane → spin → camera orbit → stomp → shockwave. Avoid random transitions or disconnected pose montages. AUDIO: Synchronize tightly to fast Japanese pop-punk vocals mixed with punchy electronic drums and distorted guitar. Add boot skids, stomps, paper snaps, cartoon impacts, guitar hits and short graphic whooshes. IMPORTANT: Keep the exact face, yellow eyes, pink hairstyle, cyan/mint hair accents, red cross clip, black-and-white school-punk top, red tie, plaid skirt and oversized plaid boots throughout. No hand reaching toward camera, no palm covering the lens, no pointing into the lens, no outfit transformation, no hairstyle transformation, no extra characters, no realistic environments, no photorealism, no 3D CGI, no modern glossy anime redesign, no cyberpunk holograms, no slow fashion posing, no facial drift, no random accessories, no excessive jump cuts, no disconnected movements, no uncontrolled morphing, no unreadable typography. Creature encyclopedia UI with cursor selecting creatures
Each reference image is mapped to a labelled card in the prompt, so every creature stays consistent when the selection changes. Locking the camera and limiting motion to the cursor, title and selected creature keeps the interface stable across the whole sequence.
Use Image 1 as the exact UI/layout/style reference for the creature encyclopedia screen.
Use Images 2–9 as the exact creature references.
Map them like this:
Image 2 = card A = LUMI HARE
Image 3 = card B = CLOUD WISP
Image 4 = card C = EMBER FENNEC
Image 5 = card D = TIDE BEHEMOTH
Image 6 = card E = PETAL VULPIN
Image 7 = card F = ORCHARD EYE
Image 8 = card G = MEADOW CROWN
Image 9 = card H = FROST GLIDER
Create a 15-second 16:9 video. Keep the camera locked. Keep the interface, layout, typography, panels, icons and overall composition stable, elegant and readable. The UI should feel like a modern minimal digital creature encyclopedia, similar to a sleek pokedex. No scene cuts, no extra text, no extra buttons, no UI distortion.
Sequence:
0–2.5s: Cursor clicks card A. Main creature becomes LUMI HARE. Title changes to “LUMI HARE”. Creature blinks and rotates slightly.
2.5–5s: Cursor clicks card C. Main creature becomes EMBER FENNEC. Title changes to “EMBER FENNEC”. Cursor drags to rotate it left and right.
5–7.5s: Cursor clicks card E. Main creature becomes PETAL VULPIN. Title changes to “PETAL VULPIN”. Cursor pokes it a few times. It reacts, annoyed.
7.5–10s: Cursor clicks card G. Main creature becomes MEADOW CROWN. Title changes to “MEADOW CROWN”. Cursor taps near the face/horns. It recoils slightly.
10–12s: Cursor clicks card D. Main creature becomes TIDE BEHEMOTH. Title changes to “TIDE BEHEMOTH”. Cursor keeps poking it.
12–15s: TIDE BEHEMOTH gets angry, opens its mouth very wide, lunges forward, and swallows the cursor. Then it returns to idle. Title stays “TIDE BEHEMOTH”.
Rules:
- When a card is selected, both the main creature and the main title must update.
- Only animate cursor, selection state, title change, and the selected creature.
- Only one cursor.
- Keep motion subtle and clean until the final swallow.
- No cuts, no camera move, no UI distortion, no extra text.
Audio:
soft UI click sounds, subtle hover sounds, tiny creature reaction sounds, then a sharper aggressive creature sound and one comedic swallow gulp at the end. How to write prompts for this model
Write MiniMax's three fields
MiniMax's guide puts every prompt in three fields: integrated_multimodal_description for picture, action, shots and dialogue along the timeline, overall_soundscape for ambient and action sound, and non_diegetic_music for music only the audience hears. Only 21 of the 307 creator prompts here use these fields, so most people are writing free text. Source: H3 prompt writing guide · AgentCraft: MiniMax H3 and Wan 3.0 prompt data
Number shots and give each cut a time
Start with [Shot 1] and no timestamp; begin each later shot with its cut time, such as [Shot 2] At 00:03.500, the camera cuts to. A cut should bring new information; for a small change of distance or angle, move the camera instead. 117 creator prompts here time their shots, mostly in their own format. Source: H3 prompt writing guide · AgentCraft: MiniMax H3 and Wan 3.0 prompt data
Fit the story into 4 to 15 seconds
MiniMax asks you to match the total duration of the description to the requested length, which is 4 to 15 seconds. The median creator prompt here is 2,096 characters, about 290 words, for at most 15 seconds of video, so cut beats before you add detail. Source: H3 prompt writing skill · AgentCraft: MiniMax H3 and Wan 3.0 prompt data
Describe camera moves by type, range and speed
The guide defines a camera move by its type, such as push in, pull out, pan or zoom, then its amplitude and its speed, and says to leave out medium amplitude and normal speed. 217 prompts here use camera terms. Source: H3 prompt writing guide · AgentCraft: MiniMax H3 and Wan 3.0 prompt data
Tie keyframes to the timeline
With a first frame, the prompt opens with a line saying <Picture 1> is fully referenced at 0.00 seconds, then describes what moves next. With a first and last frame, describe the path between them, usually in a single shot; with only a last frame, infer a plausible opening and land on the image in the final shot. Source: H3 prompt writing guide
Label references the same way everywhere
Call uploads <Picture 1>, <Video 1> and <Audio 1> and keep each label identical across every section. Of the 307 creator prompts, 61 number their uploads but only 23 use these labels. Source: H3 prompt writing skill · AgentCraft: MiniMax H3 and Wan 3.0 prompt data
English for direction, the original language for speech
Write the sections in English but keep dialogue, lyrics and on-screen text in their own language. Prefer concrete visual and audio details over words like cinematic or beautiful. Source: H3 prompt writing skill
Questions about this model
What is MiniMax H3?
MiniMax's omni-modal generative model. It understands text, images, video and audio together and generates video with native stereo audio, up to 15 seconds and up to 2K. Source: MiniMax H3 on GitHub
Can I run MiniMax H3 locally?
Partly. MiniMax released the H3-Base weights as two checkpoints, one for text and keyframes and one for references, producing 768p. The prompt rewriter H3-Context-IR and the 2K pass are hosted only, available through MiniMax's API. Source: MiniMax H3 on GitHub
How long can a MiniMax H3 video be?
4 to 15 seconds, at 24 frames per second with 32 kHz stereo audio. Source: MiniMax H3 on GitHub
How many references can H3 use?
In reference mode, up to 9 images, 3 videos and 3 audio clips, each clip 2 to 15 seconds and 15 seconds in total per type, with at most 12 files overall. Source: MiniMax H3 on GitHub
Which languages can characters speak?
MiniMax lists stable support for 11: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish, with others to varying degrees. Source: MiniMax H3 on GitHub
Is there an official skill for writing H3 prompts?
Yes. Install it with npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing. It makes no external calls, so it works in Claude Code, Codex, Cursor and other agents that read a SKILL.md. Source: MiniMax H3 on GitHub