How to write prompts for AI video, image and audio

Duration25.0% of video prompts

How long the clip runs.

Models hold a shot for a fixed span. Naming the length stops a model compressing a three beat idea into one, or padding a single beat with drift.

Values

6 seconds, 8 seconds, 10 seconds, 15 seconds, 30 seconds, 60 seconds

Also used in production skills, not counted

4-15s, 2-15s, 3-15s, 4, 6, 8, 10, 12, 15, 5, 4 s, ~15s, one to ten whole minutes

Fragments to drop in

• 15-second cinematic clip.
• Six seconds, one continuous take, no cuts.
• 30 seconds in three beats of roughly ten seconds each.

From production skills

• --duration 12
• --duration 5
• --duration 10
• --duration 4

Model-bound: Seedance 2.0 4-15s (12s is valid); Kling 3.0 and Kling 3.0 Turbo 3-15s; Grok Video 1.5 2-15s; Veo 3.1 accepts only 4, 6 or 8; Marketing Studio integer ≥ 4; explainer blocks are fixed 10-second units with N = duration_minutes × 6; sprite animation is fixed at 4 s; cover reveal ~5s; scroll-scrub single-shot asks for the longest single take the model supports (~15s).

Aspect ratio31.3% of video prompts

Frame shape.

Decides whether the subject can breathe. A vertical crop of a wide composition loses the sides, so the ratio has to be chosen before the composition is described.

Values

16:9, 9:16, 1:1

Also used in production skills, not counted

4:3, 3:4, 21:9, auto, 3:2, 2:3, 9:21, 4:5, 5:4, 1.91:1

Fragments to drop in

• Framed 16:9 for landscape playback.
• Vertical 9:16, composed for a phone held upright.
• Square 1:1 with the subject centred.

From production skills

• --aspect_ratio 16:9
• --aspect_ratio 9:16
• --aspect_ratio auto

Meanings given in prompt-engineering.md: 16:9 landscape, cinematic; 9:16 vertical, social; 1:1 square, profile / icon. Veo 3.1 accepts only 16:9 or 9:16. For sprite work the ratio must be read off the key-pose file and passed explicitly on every call , omitting it falls back to the tool default and crops the subject.

Shot size25.1% of video prompts

How much of the subject fills the frame.

The most reliable lever there is. Naming it removes the model's habit of defaulting to a mid shot for everything.

Values

close-up, wide shot, medium shot, extreme close-up, establishing shot, two-shot, macro shot, over-the-shoulder, closeup

Also used in production skills, not counted

one hero subject, kept centered, clean negative space, center-safe area, cover, full body in frame, empty margin above the head and below the feet

Fragments to drop in

• Extreme close-up on the texture, filling the frame.
• Wide establishing shot, subject small in a large space.
• Medium shot from the waist up, room to move on both sides.

From production skills

• One hero subject, kept centered, with clean negative space around it , that space is where the chapter copy sits.
• Full body in frame with empty margin above the head and below the feet.

The video fills the viewport and crops the edges, so essential subjects must stay away from the far left/right edges.

Camera angle6.8% of video prompts

Where the camera sits relative to the subject.

Changes who has power in the frame. Low angle makes a product monumental, overhead flattens it into a diagram.

Values

overhead, low-angle, eye level, three-quarter, top-down, profile, high-angle

Fragments to drop in

• Low-angle looking up, subject towering over the lens.
• Directly overhead, top-down on the surface.
• Eye-level, camera at the subject's own height.

Camera movement19.1% of video prompts

How the camera travels during the shot.

The single most common cause of a shot feeling wrong. If you do not name it the model invents drift, and drift reads as an accident.

Values

push-in, tracking shot, handheld, orbit, locked-off, slow zoom, truck, slow dolly, parallax, pull-back, slow pan, whip pan, crane up, dolly in, steadicam, tilt down, dolly out, pan left, tilt up, dolly forward, dolly back

Also used in production skills, not counted

zooms in, dollies left, sweeping pan, slow push, fast whip, camera slowly pulls back, camera dollies in, slow orbit, rise, fly-through, lateral track, crane, detail push, rack focus soft to sharp, light sweep, Very slow push-in, Subtle camera push-in, slow push-in, drift, scale shock, hard contrast cut, slow forward drift, gentle forward velocity, pull out, descending into, camera locked, no camera movement, no zoom

Fragments to drop in

• Slow push-in toward the subject, ending tight.
• Locked-off tripod, no camera movement at all.
• Smooth orbit around the subject, one quarter turn.
• Handheld with slight sway, documentary feel.

From production skills

• camera dollies in
• subtle product reveal, camera slowly pulls back, ambient motion
• MOTION: Very slow push-in; the figure erodes grain by grain and particles drift sideways.
• Subtle camera push-in, no flicker, no extra text, no new objects.
• no cuts, no camera shake, slow steady motion only, locked exposure, no on-screen text
• camera locked, no camera movement, no zoom, subject stays fully in frame, plain static background

For image-to-video the movement IS the prompt; for scrubbable footage a single unbroken move is mandatory and cuts are a defect. For sprite sheets the camera must be explicitly locked or frames become unusable.

Lens and focal length3.8% of video prompts

Perspective and compression.

A focal length is a shorthand a model understands: 85mm flatters a face, 24mm bends a room outward, macro turns a surface into a landscape.

Values

35mm, macro lens, anamorphic, 100mm, 50mm, 85mm, 28mm, wide-angle lens, 32mm, 50 mm, fisheye, telephoto, 40mm, 35 mm, 85 mm, 24mm

Fragments to drop in

• Shot on an 85mm lens, compressed and flattering.
• Wide 24mm, close to the subject, edges stretching.
• Macro lens, the subject filling the frame at life size.

Depth of field11.8% of video prompts

What is sharp and what falls away.

Separates subject from background without changing the composition. The corpus reaches for shallow depth of field more than any other optical term.

Values

shallow depth of field, bokeh, rack focus, f/1.4, f/1.8, f/2.0, soft focus, f/2.8

Fragments to drop in

• Shallow depth of field, background dissolving into bokeh.
• Deep focus, everything from foreground to horizon sharp.
• Rack focus from the foreground object to the face behind it.

Lighting7.4% of video prompts

Where the light comes from and how hard it is.

Sets mood more cheaply than any other choice. Naming a setup also stops the flat, evenly lit look models fall back on.

Values

soft lighting, rim lighting, rim light, silhouette, volumetric lighting, golden hour, volumetric light, soft light, side lighting, chiaroscuro, key light, low-key, practical lighting, backlit, high-key, side light, practical lights, backlight, god rays, neon lighting, blue hour, softbox

Fragments to drop in

• Soft lighting from a large source, shadows barely there.
• Rim lighting from behind, edge of the subject glowing.
• Single hard key from the side, deep shadow on the far cheek.
• Golden hour sun, low and warm, long shadows.

Colour and grade17.0% of video prompts

The palette and contrast of the finished image.

Carries brand more than any other single element, and is easy to get consistent across a set of clips by reusing the same phrase.

Values

vibrant, pastel, film grain, black-and-white, natural color, high contrast, desaturated, crushed blacks, natural colour, low contrast

Also used in production skills, not counted

name the grade and the hexes, one visual grade, colour grading

Fragments to drop in

• Vibrant saturated grade, colours pushed but not clipping.
• Muted pastel palette, low contrast, gentle.
• High contrast with crushed blacks and a cool cast.
• Fine 35mm film grain over the whole frame.

From production skills

• Prompt "no cuts, no camera shake, slow steady motion only" and name the grade/hexes.
• Take only the visual render style and color grading of the input image(s); mix the styles if there is more than one image.

scroll-scrub.md forbids mixing models mid-chain because "their grain, color, and motion signatures create a visible seam even when position matches".

Medium and style38.1% of video prompts

What the output is pretending to be.

The most used word in the whole corpus is cinematic, which tells you models respond to a named medium. Being specific beats being evocative.

Values

cinematic, photorealistic, anime, documentary, 3d render, unreal engine, stop-motion, 35mm film, watercolor, vector, photoreal, claymation, line art, octane

Also used in production skills, not counted

flat 2D vector animation, bold clean outlines, solid vibrant flat fills, no shading, no gradients, hand-inked black marker on off-white paper, solid jet-black fills, thin white scratch highlights, marker grain, strictly monochrome, strict monochrome minimalism, black silhouettes on white void, high contrast, lots of negative space, hand-painted storybook gouache, soft textures, warm muted palette, visible brush strokes, non-photorealistic, illustrated, not a photo, no live-action, no realism

Fragments to drop in

• Cinematic, shot on 35mm film.
• Photorealistic, indistinguishable from a real camera.
• 3D render, clean studio lighting, product visualisation.
• Hand drawn 2D animation, visible line work.

From production skills

• STYLE REFERENCE: Match the attached reference image EXACTLY. Replicate its look precisely: {STYLE tokens}. Every element below rendered in that identical style.
• non-photorealistic, illustrated, not a photo, no live-action, no realism
• Paste the same STYLE tokens into every block.

Write the render style, palette, line character and finish once, then repeat it byte-identically. scroll-scrub.md calls the equivalent a 'World grammar': one byte-identical style preamble, perspective, palette, light direction, surface finish and background behavior across all scene prompts.

Subject motion10.2% of video prompts

What moves inside the frame, and how fast.

Distinct from camera movement and frequently confused with it. Say what the subject does, or the model will hold it still and move the camera instead.

Values

reveal, float, drift, slow motion, glides, rotates, time-lapse, ripples, unfolds, hovers, cascade

Also used in production skills, not counted

the dancer spins, smoke rises slowly, ambient motion, fur ripples, stars orbit, clouds drift, train lights flicker, object assembling, exploded-view assembly, transformation or morph, a hero object emerging from darkness, smooth idle breathing cycle, subtle weight shift, full walk cycle in place, single sword slash, fast wind-up, sharp strike, follow-through, plain static background, studio black/charcoal, a soft gradient, the subject emerging from darkness, dark, seamless, low-detail, clean background, simple background

Fragments to drop in

• The lid glides open on hidden runners, slow and even.
• Fabric billows in slow motion, every fold readable.
• Steam drifts upward and dissipates before it leaves the frame.

From production skills

• SCENE: A lone black silhouette slowly dissolves at the edges, crumbling into fine drifting sand that scatters into the white emptiness.
• The character performs ONLY this action; nothing else happens.
• Use one clear action per block.
• A background the copy can survive. Dark, seamless, low-detail (studio black/charcoal, a soft gradient, the subject emerging from darkness) is the reliable choice.

"Exactly one action per video; any second verb in the prompt is a defect" (game-2d-animation.md).

Critical when HTML copy sits over the video: "A bright, busy, full-frame environment behind body copy is the single most common reason a beautiful film reads as unusable."

Mood21.9% of video prompts

The feeling the clip should leave.

A single adjective steers dozens of small decisions the model would otherwise make at random. Cheap to add, and it makes a set of clips feel related.

Values

playful, energetic, whimsical, calm, dramatic, minimal, elegant, uplifting, moody, ominous, ethereal, luxurious, serene, nostalgic, cosy, gritty

Fragments to drop in

• Calm and unhurried throughout.
• Playful and energetic, quick and bright.
• Moody and restrained, close to darkness.

Audio bed11.8% of video prompts

The sound that plays under the picture.

Video prompts in this corpus routinely describe their own soundtrack. Naming instruments and what must not be heard is more reliable than naming a genre.

Values

percussion, bells, ukulele, xylophone, piano, drums, bass, marimba, strings, synth, guitar, choir, handclaps, pads, flute, synthesizer, synthetic, harp, synthase, kalimba

Also used in production skills, not counted

ambient SFX or music only, no voice, dialogue, or narration, Low sustained drone, a soft whisper of falling sand, generate_audio false, sound on/off

Fragments to drop in

• Minimal deep electronic ambience with soft mechanical clicks.
• Warm ukulele and marimba, light percussion, no vocals.
• No music at all, only the room and the mechanism.

From production skills

• AUDIO: {ambient SFX or music only, NO voice, dialogue, or narration}.
• AUDIO: Low sustained drone and a soft whisper of falling sand, no voice.
• Keep clip audio diegetic only. Characters never speak or lip-sync.

In the explainer pipeline narration is a separate audio job; clip audio must never contain speech.

Pacing and cuts

How the time is divided.

Without it a model gives you one unbroken drift. Naming beats is what turns a clip into a sequence.

Also used in production skills, not counted

smooth cinematic easing, slow, steady motion, constant speed, gentle ease only at the very start and end, locked exposure and white balance, no flicker, minimal motion blur, one continuous move, no hard cuts, START state ≠ END state, 0-1.5s , scene alive, no text yet, 1.5-3.5s , text entrance, 3.5-5s , settle, pops/slides/bounces in, pixel type snaps in block by block, chrome bubble letters inflate, condensed uppercase slams down, fades up, slides in with a soft bounce, Micro-motion only

Fragments to drop in

• One continuous take, no cuts.
• Three beats: reveal, detail, hero shot.
• Cut away before the hand touches the product.

From production skills

• One continuous move , no hard cuts.
• Slow, steady motion , constant speed, gentle ease only at the very start and end. The scroll supplies the pacing.
• Locked exposure and white balance, no flicker; minimal motion blur.
• Resolves at both ends: the first frame reads as the establishing shot, the last as the closing beauty state. START state ≠ END state, or the scrub has no payoff.
• Premium 5-second motion cover reveal, smooth cinematic easing.
• All elements ease precisely into their final positions and the video ends exactly on the provided end frame, holding still for the last moments.
• Then the title "[TITLE TEXT]" [ENTRANCE MOTION matched to its typography], the tagline fades up beneath it, and the small rounded pill button slides in with a soft bounce.

Matters whenever every frame may be held as a still (scroll-scrub) , exposure pumping shimmers and heavy blur smears.

Entrance motion should be matched to the typography of the text being revealed.

Resolution / quality tierfrom production skills

Seedance 2.5 caps at 720p so 1080p/4K work stays on Seedance 2.0; Grok Video 1.5 only 480p or 720p; Marketing Studio video 480p or 720p.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

480p, 720p, 1080p, 4k, basic, high, ultra, pro, std, standard, bitrate_mode standard|high

Fragments to drop in

• --resolution 4k
• --resolution 720p
• --resolution 1080p --mode std

First/last frame anchoringfrom production skills

scroll-scrub.md warns that passing a storyboard as a start frame makes the page open on a static image; pass it in the generic reference role instead.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

--start-image, --end-image, start frame, end frame, start = end, one-shot, loop, end-frame reveal, generic image/reference role , not a start-frame role

Fragments to drop in

• `--start-image` anchors the first frame. Prompt describes motion.
• Don't redescribe the static frame , model already has it.
• Looping actions pass the SAME image as both start frame and end frame. One-shot actions (attack, death, hit, cast) pass only the start frame , forcing them back to the start pose ruins the action.
• Pass the finished cover as the END frame and describe the buildup

Ad format / mode (branded video)from production skills

Hook text is prepended to the user's prompt and does not replace it; COOKBOOK.md notes "Hooks (the prompt on --prompt) matter more than mode for performance."

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

ugc, ugc_how_to, ugc_unboxing, product_showcase, product_review, tv_spot, wild_card, ugc_virtual_try_on, virtual_try_on, hook, setting, ad reference

Fragments to drop in

• Looks like a real person filmed on phone
• Polished broadcast commercial
• Show the product itself, less presenter
• Presenter giving an opinion

Edit instruction (video editing workflows)from production skills

draw_to_video takes a source video, an edited/sketched frame, the timestamp for that frame and a short edit instruction.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

make the jacket red, sketch, timestamp

Fragments to drop in

• --prompt "make the jacket red"

Aspect ratio16.5% of image prompts

Frame shape.

Choose it first. Composition instructions that assume a landscape frame fall apart in a vertical one.

Values

16:9, 9:16, 1:1, 4:5, 21:9, 3:2, 4:3

Fragments to drop in

• Square 1:1.
• Vertical 4:5 for a feed post.
• Wide 16:9 with the subject offset left.

Composition and framing31.3% of image prompts

Where things sit in the frame.

The most common failure in the corpus is a subject cropped or crowded. Explicit composition instructions, including what must stay fully visible, fix it.

Values

close-up, wide shot, medium shot, extreme close-up, two-shot, closeup, over-the-shoulder, establishing shot, macro shot, full shot

Also used in production skills, not counted

one power-third hero, clear scale hierarchy, depth, strong subject/background separation, centered, left-anchored, asymmetric split, one focal visual, decisive negative space, generous negative space, reserve the block side, soft open negative space upper-left, single unified frame, split-screen, diagonal divide, halves/panels, extreme scale contrast, power third, centered isolated presentation

Fragments to drop in

• Centred on a single vertical axis, symmetrical.
• Rule of thirds, subject on the right third, negative space left.
• Everything fully inside the frame, nothing cropped at the edges.
• Generous negative space above the subject for a headline.

From production skills

• Composition: one power-third hero, clear scale hierarchy, depth, and strong subject/background separation.
• Scale + placement: "waist-up, subject fills two thirds of the frame, positioned right-of-center" , and reserve the block side: "soft open negative space upper-left on a warm white wall". Negative space = a real surface (wall, sky, backdrop), NOT a black void.
• single unified frame , no split-screen, no diagonal divide, everything blends smoothly and organically across the same continuous shot

A split layout is used only when the user asks for split, before/after, versus or side by side; "A topical phrase such as `X vs Y` does not itself require a split."

Shot size11.4% of image prompts

How close the camera is.

Same lever as in video and just as reliable. It also controls how much detail the model has to invent.

Values

close-up, wide shot, medium shot, extreme close-up, two-shot, closeup, over-the-shoulder, establishing shot, macro shot, full shot

Also used in production skills, not counted

chest-up, medium-close, about 40-60% of the frame, waist-up, subject fills two thirds of the frame, 55-75 % of the frame, Full body, Three-quarter, Waist up, Closeup on product area, head shot, head-and-shoulders, full body, large, foreground-dominant, rendered LARGE

Fragments to drop in

• Extreme close-up, texture filling the frame.
• Full body shot, head to feet, room above and below.
• Head and shoulders portrait.

From production skills

• Subjects: large, foreground-dominant, chest-up or medium-close, filling about 40-60% of the frame. End with `All faces crisply sharp as the anchors of the shot.`
• waist-up, subject fills two thirds of the frame, positioned right-of-center

app-cover.md kills a candidate whose subject is under ~50 % of frame. Framing is an explicit interview option in product-photoshoot: `[Full body / Three-quarter / Waist up / Closeup on product area]`.

Camera angle4.6% of image prompts

Viewing position.

Three-quarter is the most used angle in the corpus for a reason: it shows form without the stiffness of a straight-on view.

Values

three-quarter, profile, eye-level, low angle, overhead, high angle, top-down, bird's-eye, dutch angle

Also used in production skills, not counted

low, low wide angle, side-view, flat frontal, three-quarter isometric view, front, 3/4 left, 3/4 right, slight up/down, wide establishing view, multi_angle

Fragments to drop in

• Three-quarter view, turned slightly off axis.
• Straight-on profile, subject facing frame left.
• Shot from directly above, flat lay.

From production skills

• angle (low, overhead)
• consistent side-view perspective across all assets
• three-quarter isometric view

game-stylization.md makes the perspective word genre-bound: platformer/runner to side-view; top-down to top-down; puzzle/clicker to flat frontal. For image-to-3D concept images the perspective word is replaced by three-quarter isometric view because it shows top plus two sides.

Lens and focal length5.0% of image prompts

Perspective and compression.

Also the cheapest way to signal photographic realism, because a real photograph was taken at some focal length.

Values

35mm, 85mm, 50mm, 110 mm, 16mm, anamorphic, 24mm, 70mm, 48mm, fisheye, 600mm, 450mm, 40mm, 297 mm, 650mm, 20mm, 135mm, 100mm, wide-angle lens, 105 mm, 220mm, 110mm

Also used in production skills, not counted

low wide angle, close microphone (audio analogue), 35mm film

Fragments to drop in

• 85mm portrait lens, natural compression.
• 35mm, mild environmental context around the subject.
• Macro at life size, the surface reading as landscape.

From production skills

• **Camera**: lens (35mm, 85mm), angle (low, overhead), motion (dolly in, tracking shot)
• bright soft key with punchy shadows, low wide angle, commercial editorial grade

app-cover.md notes the lens block is replaced by style words such as "glossy 3D render / claymation diorama / painterly still" on the graphic route. Vector/logo prompts must never contain lens language.

Depth of field6.0% of image prompts

Focus falloff.

Separates a subject from clutter you did not describe, which is most backgrounds.

Values

shallow depth of field, bokeh, f/1.4, f/7.1, soft focus, f/2.8, deep focus, f/2.0, f/4, f/2, f/1.8, f/5.6

Fragments to drop in

• Shallow depth of field at f/1.8, background soft.
• Everything sharp front to back, deep focus.
• Only the eyes critically sharp.

Lighting6.8% of image prompts

Direction, quality and source of light.

Silhouette and rim light are the two most used named setups in the corpus, both of which the model will never choose on its own.

Values

silhouette, rim light, soft light, soft lighting, backlit, key light, golden hour, high-key, rim lighting, backlight, softbox, low-key, volumetric light, bounce light, practical lighting, three-point lighting, fill light, volumetric lighting, practical lights, chiaroscuro, side lighting, god rays

Also used in production skills, not counted

neon glow, moody backlight, strong key light, soft dreamy fill, back light, hair light, bright soft key with punchy shadows, flat ambient light, flat even lighting, soft studio reflections, soft contact shadow, warm hard light, restrained lighting, no shadows, no vignette, soft vignette and edge falloff, warm amber torchlight

Fragments to drop in

• Large softbox front left, soft shadows falling right.
• Backlit into silhouette, subject dark against a bright field.
• High-key, white on white, almost no shadow.
• Single hard light, sharp shadow edges.

From production skills

• **Lighting**: rim light, neon glow, moody backlight
• signature YouTube thumbnail lighting rig , strong key light sculpting the face, soft dreamy fill lifting shadows, and defined back light plus hair light tracing a clean bright rim around hair, shoulders and silhouette.
• bright soft key with punchy shadows, low wide angle, commercial editorial grade
• Flat even lighting, no shadows, no vignette.
• flat even lighting, no perspective, no objects, photographed-flat appearance

In the thumbnail rig "Only the rim may use a colored accent." Tileable textures need flat lighting so the tile has no focal light. Brandkit mockups explicitly ban "Excessive bloom, haze, depth of field, or cinematic lighting".

Colour and palette10.4% of image prompts

Palette, saturation, contrast.

Keeps a set consistent. Reuse the identical phrase across a batch rather than paraphrasing it each time.

Values

vibrant, pastel, film grain, black-and-white, desaturated, high contrast, natural color, natural colour, monochrome, low contrast, sepia

Also used in production skills, not counted

vivid, bright, glossy, poster-punchy, deep blacks, crisp highlights, rich saturated colors, cohesive as one image, commercial editorial grade, duotone, graded atmospheric photo, monochrome / force grayscale, muted cinematic movie still (banned), name 2-3 colors, exact hex values, locked primary tone, locked accent, locked background, background, text, primary, accent, support, environment in deep violet stone with charcoal shadows, hero in warm coral-cream tones contrasting the surroundings, hazards and pickups marked with acid-green glow, deep-navy #243A5E, signal-red #D23B2E, olive #7A7D3C, sky-blue #A9CFF4, peach #F7DDB9, cream #F1EEE6, dusty-rose #E7B7B0, warm-gray #D6D3CE, acid lime #D4FF3F, #FF00FF, #00FF00, #0000FF

Fragments to drop in

• Vibrant, fully saturated, clean whites.
• Muted pastel palette, soft contrast.
• Black and white, deep blacks, bright highlights.

From production skills

• Grade: vivid, bright, glossy, poster-punchy, deep blacks, crisp highlights, rich saturated colors, cohesive as one image. Restrain it only for an explicit calm/premium/muted brief.
• for a monochrome look force grayscale in the prompt AND on export
• Palette, committed and saturated: name 2-3 colors. Avoid lime/acid yellow-green.
• Use role language such as "locked primary tone," "locked accent," and "locked background."
• Exact hex values are passed separately through the Recraft request's `colors` and `background_color` params. Never put hex, RGB, Pantone, or other color codes in the prompt.

asset-system.md: a mixed-grade asset kit reads cheaper than no assets; re-generate off-grade pieces "with the hexes named harder".

Two opposite conventions co-exist: website/game prompts NAME the hexes in the prompt; Recraft logo prompts must never contain a colour code and pass hexes as params. Logo palettes use one, two or three colours , "three is a maximum, not a target or default". Game palettes are specified BY ROLE, not as one global gamma.

Medium and style26.6% of image prompts

Photograph, render, illustration or something else.

Ambiguity here produces the uncanny middle ground that looks like neither a photo nor an illustration.

Values

cinematic, photorealistic, documentary, anime, 3d render, line art, watercolor, vector, 16mm, 35mm film, photoreal, kodak vision3, isometric, flat illustration, claymation

Also used in production skills, not counted

oil painting, photograph, soft cel shading, glossy 3D render, claymation diorama, painterly still, flat vector cartoon with soft gradients, chunky pixel art, 32x32 grid feel, soft hand-painted gouache, low-poly faceted, smooth rounded blobby, flat_vector, monoline, vector_gradient, hand_drawn_vector

Fragments to drop in

• Photorealistic editorial photograph.
• Clean 3D product render on a seamless backdrop.
• Flat vector illustration, limited palette, no gradients.
• Loose watercolour with visible paper texture.

From production skills

• **Style/medium**: oil painting, watercolor, photograph, anime, 3D render
• transform into anime style, vibrant colors, soft cel shading

Brandkit treatment vocabulary is a closed set with defined meanings: flat_vector = solid fills, clean SVG paths, no surface effects; monoline = uniform stroke weight, rounded caps, no fills; vector_gradient = vector-safe linear, radial or duotone gradient with the locked stop count; hand_drawn_vector = intentional stroke variation. "Dimensional/3D treatment is forbidden."

Materials and finish8.5% of image prompts

What surfaces are made of.

Materials are where a render either convinces or does not. Naming the finish is more effective than naming the object.

Values

glossy, matte, ceramic, concrete, leather, velvet, satin, brass, linen, frosted, marble, chrome, walnut, oak

Also used in production skills, not counted

print, emboss, debossing, foil, laser engraving, one-color printing, screen printing, embroidery, spot-color, folds, grain, perspective, occlusion, reflections, scale, manufacturing limits, believable materials, light kraft paper, natural cardboard, pale fabric, curved packaging, surface wear

Fragments to drop in

• Brushed metal with a satin finish, no mirror reflections.
• Matte ceramic, slightly uneven glaze.
• Smoked glass over dark walnut.

From production skills

• [PHYSICAL APPLICATION] Render the logo using <credible print/emboss/foil/engraving/ink behavior>. Respect folds, grain, perspective, occlusion, reflections, scale, and manufacturing limits.
• Prefer believable materials, restrained lighting, purposeful negative space, specific environments, and one focal branded application.

Use the generator only when branding must interact with fabric folds, curved packaging, embossing/debossing, foil, print texture, reflections, surface wear, occlusion or realistic perspective; otherwise place the logo deterministically.

Background36.9% of image prompts

What sits behind the subject.

Named in 37% of image prompts, the highest of any positive aspect. An undescribed background is where models put clutter, text and logos.

Also used in production skills, not counted

vivid high-contrast color field, soft vignette and edge falloff, 2-4 supporting story props/layers, credible setting, specific environments, no readable screens, clean background, solid uniform bright <KEY COLOR> background, pure flat white, no shadow, no ground, Studio clean, Outdoor natural, Street style, Editorial, Home cozy

Fragments to drop in

• Seamless mid-grey studio backdrop, no horizon line.
• Plain white, no props, no shadow on the background.
• Softly blurred interior, unreadable, no recognisable objects.

From production skills

• Background: vivid high-contrast color field or environment, soft vignette and edge falloff; do not divide it unless a split layout was explicitly requested.
• **World + props (story density)**: name 2-4 supporting elements that fill the frame around the subject. Add "no readable screens" whenever screens/props could sprout text.
• on a solid uniform bright <KEY COLOR> background, no shadows cast on the background, no ground plane, nothing cropped at the edges

app-cover.md: "One person on an empty backdrop reads as AI slop even when bright , stage a world, not a portrait." For image-to-3D the background is instead "pure flat white, no shadow, no ground".

Mood9.9% of image prompts

The feeling.

One adjective, placed early, quietly resolves dozens of choices in a consistent direction.

Values

elegant, playful, minimal, dramatic, calm, energetic, whimsical, moody, luxurious, nostalgic, uplifting, ethereal, cosy, gritty

Also used in production skills, not counted

shock, hype, rage, awe, laugh, fear, smug, charisma, confusion, determination, disgust, neutral, smiling, talking

Fragments to drop in

• Elegant and restrained.
• Playful and bright.
• Moody, close to darkness.

From production skills

• If the user gives an emotion count without names, use this ladder: shock, hype, rage, awe, laugh, fear, smug, charisma, confusion, determination, disgust.

Emotion is one of the allowed surgical-edit scopes; expression changes are made as an edit, not a full regeneration.

Text handling18.4% of image prompts

Whether any lettering appears.

Models add text nobody asked for. 18% of image prompts in the corpus explicitly forbid it, and that is the single highest value instruction you can add.

Also used in production skills, not counted

No text, no readable UI labels, no watermark, BAKED UI, reading exactly, massive, ultra-legible sans-serif with a clean outline/glow treatment, 2-4 words max, ALL CAPS, no extra text

Fragments to drop in

• No text, no lettering, no numbers anywhere in the image.
• No logos, watermarks, captions, borders or frames.
• Leave the upper third clear for a headline added later.

From production skills

• Text: default `No text, no readable UI labels, no watermark.`
• TEXT: bold thumbnail headline text baked into the image, reading exactly "<TEXT>" , massive, ultra-legible sans-serif with a clean outline/glow treatment, placed where it never covers the subject's face. No other text, no watermark.
• **The image model NEVER renders text. Not the title, not the wordmark, not labels, not UI. It renders ONLY the scene.**
• Never bake the final logo or exact copy into a generated image.

The dominant convention across the repo is: keep text out of the generation and set type deterministically afterwards. GPT Image 2 is the one model reached for when readable text must be inside the image.

Subject + setting + style (the opening clause)from production skills

"Higgsfield models reward concrete, sensory prompts." app-cover.md wants the product's core verb staged as a physical, photographable moment.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

a red fox curled in a snowy pine forest, golden hour, cinematic, subject + action, concrete and physical

Fragments to drop in

• a red fox curled in a snowy pine forest, golden hour, cinematic
• a film editor in a bold red bomber jacket mid-leap, cutting a giant arc of 35mm film with oversized chrome scissors

Typography named in the prompt (when text is generated)from production skills

text-overlay-bake.md gives the MrBeast headline stack as concrete values: font Anton, ALL CAPS, stroke 8-14% of cap size drawn under the fill, hard dark offset shadow, colour white / yellow to orange to red gradient / acid lime #D4FF3F, tracking -0.01...-0.02em, line-height 0.9, cap height 12-18% of frame height.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

Anton, Anton SC, Bebas Neue, Oswald, Archivo Black, Montserrat, Poppins, Roboto Condensed, Inter, Barlow Condensed, Playfair Display, Cormorant Garamond, DM Serif Display, Fraunces, Sacramento, patrick, caveat, marker, anton, clean grotesk (Satoshi-like), refined grotesk (Neue-Montreal-like), expressive display (Cabinet/Clash-like), compressed statement (Monument-like), editorial serif + sans pairing, Swiss rational sans with hard hierarchy

Fragments to drop in

• The prompt must state: Exact literal copy; Display/body font family names and which text uses each
• Name the exact display/body families in the prompt; never infer typography from the logo or palette.

Identity lock / subject fidelityfrom production skills

Up to three referenced identities; face photos are passed first in character order, then the logo, and a manifest line must open the prompt when two or more references are attached.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

IDENTITY LOCK, photographic identity match, same bone structure, eye shape, nose, lips, jawline, skin tone, hairline and hair texture, Do not beautify, average, or restyle the face, IMAGE REFERENCES manifest

Fragments to drop in

• CHARACTER N: the person from attached face reference #K , IDENTITY LOCK: reproduce this exact person with a photographic identity match , same bone structure, eye shape, nose, lips, jawline, skin tone, hairline and hair texture. Do not beautify, average, or restyle the face. Expression: <emotion phrase>.
• IMAGE REFERENCES: image 1 = CHARACTER 1 face reference; image 2 = brand logo.
• All faces crisply sharp as the anchors of the shot.

Aesthetic / seasonal presetfrom production skills

These are the labelled options the skill offers for restyle and style/mood questions; they become the short intent line handed to the backend enhancer.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

Clean girl, Cottagecore, Quiet luxury, Dark academia, Y2K, Christmas, Valentine's, Halloween, Black Friday, None, Clean studio, Lifestyle, Conceptual, With a model

Fragments to drop in

• Christmas version, quiet-luxury aesthetic
• vertical pin for my candle brand, cottagecore mood

Product-photo genre / modefrom production skills

For both skills the user-facing prompt is a short intent line only , "Backend assembles the final prompt; never freehand it." `--count 3` makes the backend vary preset, lighting, angle and palette across variants.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

product_shot, lifestyle_scene, closeup_product_with_person, moodboard_pin, hero_banner, social_carousel, ad_creative_pack, virtual_model_tryout, conceptual_product, restyle, levitating, floating, splash, frozen motion, surreal, CGI, sculptural, main_image, infographic, multi_angle, detail_shot, lifestyle, whats_in_box

Fragments to drop in

• bottle of cold-brew on a sunlit kitchen counter, IG feed
• sparkling peach lemonade can for marketplace listing
• premium skincare serum, clean clinical marketplace visual system
• Bold hero shot on marble

Logo / vector mark specificationfrom production skills

One central visual idea per mark, stated in one clause; the exact phrase "no text" must appear in the constraint tail; mark_type routing is evidence-based, defaulting to abstract, never wordmark or combination mark.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

lettermark_monogram, pictorial, abstract, mascot, emblem, one dominant unified silhouette, negative-space device, motif treatment, unexpected locked color-role pairing, symmetry or intentional asymmetry, internal detail limit, small-size scalability, centered isolated presentation, minimal anchor points, SVG-friendly

Fragments to drop in

• Flat vector design, clean lines, no shadows, no texture, no text. Clean editable vector paths, SVG-friendly, minimal anchor points.
• Monoline vector design, uniform stroke weight, rounded line caps, no fills, no shadows, no texture, no text. Clean editable vector paths, SVG-friendly, minimal anchor points.
• Flat vector design with the specified locked vector gradient, no shadows, no texture, no text. Clean editable vector paths, SVG-friendly, minimal anchor points.
• Intentional hand-drawn vector strokes, approved surface variation, clear scalable silhouette, no shadows, no text.
• Describe the distinctive element concretely enough that another designer could sketch its structure without guessing.

Logo placement on a mockup surfacefrom production skills

Full-colour marks go on smooth surfaces; black monochrome on light kraft paper, natural cardboard, pale fabric, stamps, dark-ink screen printing, engraving masks and light uncoated stock; white monochrome on dark paper/boxes/fabric, reverse marks, light-ink screen printing and dark signage.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

Target surface, Position, Scale, Orientation, Color, horizontally centered, upper third, optical center aligned to panel, clear-space margin, align to <panel edge/seam/baseline>, follow surface perspective

Fragments to drop in

• [PLACEMENT LOCK] Target surface: <exact object panel/face/material>. Position: <exact alignment and location, e.g. horizontally centered, upper third, optical center aligned to panel>. Scale: logo occupies <specific proportion> of the target surface while keeping <specific clear-space margin>. Orientation: align to <panel edge/seam/baseline>; follow surface perspective without changing logo proportions. Color: use the supplied <full-color/black/white> variant exactly. State why it contrasts correctly with the material/background.
• [AUTHORITATIVE LOGO] <ImageN> is the exact approved <full-color/black/white> logo. Preserve its spelling, silhouette, geometry, proportions, internal negative space, and exact color. Do not redraw, simplify, crop, stretch, outline, or add effects.
• "Place the logo on the bag/box" is insufficient.

Seamless / tileable behaviourfrom production skills

{MATERIAL} in the texture templates is 3-6 concrete words from the reference, e.g. "hand-painted cracked dry earth with small grey stones".

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

seamless tileable game texture tile, uniform pattern density, perfectly seamless edges that wrap horizontally and vertically, no border, no vignette, no single focal object, photographed-flat appearance, seamless repeating pattern

Fragments to drop in

• seamless tileable game texture tile of <description>, uniform pattern density,
• , perfectly seamless edges that wrap horizontally and vertically, no border, no vignette, flat even lighting, no single focal object
• seamless tileable surface texture of <description>,
• , perfectly seamless edges, flat even lighting, no perspective, no objects, photographed-flat appearance
• The repeating units must be aligned so rows continue across the tile borders: unit boundaries at the left edge continue into the right edge and across top/bottom.
• Reduce the visible periodic repetition of distinctive features so the pattern feels organic.

Key colour for transparencyfrom production skills

Models cannot output alpha, so asking for "transparent background" produces a fake checkerboard. Default magenta; switch to green if the subject's own colours are near pink/magenta/purple; blue if both are taken. The chosen colour must appear in both the prompt suffix and the keying script.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

bright magenta #FF00FF, bright green #00FF00, bright blue #0000FF

Fragments to drop in

• on a solid uniform bright <KEY COLOR> background, crisp edges, no drop shadow outside the element
• Clean uniform <key color> background.

Reference-image role and what it controlsfrom production skills

For image-to-image the prompt should describe what changes, not redescribe the input: bad "a man with brown hair in a leather jacket holding coffee, made into anime"; good "transform into anime style, vibrant colors, soft cel shading".

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

style donors only, Image0, Image1, Image2, official logo, approved base scene, style-reference, representative-visual, grade/energy only

Fragments to drop in

• Make an Animated Explainer. Take only the visual render style and color grading of the input image(s); mix the styles if there is more than one image. Never use the characters, inscriptions, etc. from the input image(s) unless the instructions below ask you to. Use only the render style, and follow the user's instructions below:
• Use references as style donors only. Never copy their people, text, logos, or objects.
• Label references by role in prompts: `Image 1: official logo`, `Image 2: approved base scene`.
• State what each reference controls and what it must not control.

Design-board direction vocabulary (web/layout imagery)from production skills

Commit to one option per category before prompting and hold it across all boards; one board per section, landscape 16:9 or 3:2, never one tall full-page image.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

Pristine Light (paper/cream/off-white, dark ink), Deep Dark (charcoal/graphite), Bold Studio Solid (oxblood, royal blue, forest, vermilion, emerald fields), Quiet Premium Neutral (bone, sand, taupe, stone, smoke), technical grid/dot field, solid with soft ambient depth, full-bleed cinematic imagery, tactile paper/material texture, cinematic centered minimalist, asymmetric split, floating polaroid scatter, inline typography behemoth, editorial offset, massive image-first with restrained text, modular bento rhythm, alternating editorial blocks, poster-stacked storytelling, gallery-led cadence, Swiss grid discipline, asymmetric premium flow, artifact/collectible, journey/waypoints, tool/precision instrument, living system/garden, stage/spotlight, archive/dossier

Fragments to drop in

• "website design mockup, desktop landing page section, [SECTION ROLE], [theme paradigm + exact palette words], [typography character] typography, [hero architecture / composition anchor], [background mode], [narrative spine motif], professional layout, clear hierarchy and spacing, award-winning web design" , plus: "no watermark, no browser chrome".
• Name real content in the prompt (the actual headline wording you plan) so type sits believably.

Reference photos for identity trainingfrom production skills

5-20 photos, 8-12 the sweet spot, clear face with eyes visible, single person per photo, sharp, ≥1024×1024 ideal.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

front, 3/4 left, 3/4 right, slight up/down, indoor, outdoor, soft, harsh, neutral, smiling, talking, head shot, head-and-shoulders, full body

Fragments to drop in

• Higher variety = better identity capture.

Image-to-3D input framingfrom production skills

One figure per input image , a 4-view character sheet reconstructs as "a statuette of 4 fused figures".

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

t-pose, a-pose, Full body in frame, nothing cut off, Clean/plain background, Limbs visually separated from the torso, One character, no props overlapping the silhouette, lowpoly, three-quarter isometric view

Fragments to drop in

• Full body in frame, nothing cut off.
• Clean/plain background; no shadows, text, watermarks, logos.
• Limbs visually separated from the torso (arms not pressed to the body) , fused silhouettes produce fused geometry and break auto-rig.

Duration38.0% of audio prompts

Length of the piece.

A generator will otherwise give you a loop of arbitrary length that has to be cut, which is where the musical ending gets lost.

Values

15 seconds, 30 seconds, 60 seconds, 90 seconds, 120 seconds

Also used in production skills, not counted

--duration 30, --duration 12, --duration 4, --duration 2

Fragments to drop in

• 90-second bed.
• 30 seconds with a clean ending.
• 15-second sting.

From production skills

• mirelo_text_to_audio --prompt "glass breaking in a large hall" --duration 4
• sonilo_music --prompt "cinematic synthwave track" --duration 12

Sonilo Music and Mirelo Text to Audio both require an explicit --duration; Seed Audio does not.

Role

What the audio is for.

A bed under narration and a standalone track need opposite decisions about melody and midrange, so state the job before the style.

Also used in production skills, not counted

no music, no voice, no ambience, dry studio recording, close microphone, short isolated, single, sound effects, ambience, foley, impacts, transitions, environmental sounds

Fragments to drop in

• Instrumental music bed to sit under narration.
• Standalone track, no voice over it.
• Short sting to close a scene.

From production skills

• short isolated sword impact, dry studio recording, no music
• single heavy wooden door slam, close microphone, no ambience
• glass breaking in a large hall
• cinematic rain ambience with distant thunder
• Describe one sound per SFX prompt. Do not request a complete mixed scene.
• State `no music`, `no voice`, or `no ambience` when isolation matters.

This is the single richest audio prompt guidance in the repo. One sound per prompt is the hard rule.

Tempo38.7% of audio prompts

Speed, ideally as a number.

39% of audio prompts in the corpus give a BPM. A number is obeyed far more reliably than a word like upbeat.

Values

70 BPM, 90 BPM, 100 BPM, 116 BPM, 128 BPM

Fragments to drop in

• Buoyant mid-tempo around 116 BPM.
• Slow, roughly 70 BPM.
• Driving, 128 BPM, steady four on the floor.

Key and tonality15.3% of audio prompts

Major or minor, and the key if it matters.

The cheapest emotional lever in audio. Major reads as safe and bright, minor as unresolved, and models respond to both.

Values

major key, minor key, modal, pentatonic

Fragments to drop in

• Bright major key.
• Minor key, unresolved.
• Warm, clean major-key resolution in the final eight seconds.

Instrumentation71.2% of audio prompts

Which instruments play.

Named in 71% of audio prompts, the most used aspect of any type. Listing instruments works better than naming a genre, which models interpret loosely.

Values

percussion, ukulele, piano, bass, handclaps, marimba, bells, strings, guitar, xylophone, drums, harp, synthesizer, pads, 808, synth, flute, choir, synthesized

Also used in production skills, not counted

instrumental, no vocals, cozy forest loop, warm marimba and soft strings, cinematic synthwave track, backing tracks, instrumental beds, jingles, musical moods

Fragments to drop in

• Marimba, ukulele, xylophone, light percussion and soft handclaps.
• Solo piano with a distant string pad.
• Deep electronic ambience with soft mechanical clicks.

From production skills

• instrumental cozy forest loop, warm marimba and soft strings, no vocals
• cinematic synthwave track
• Music prompts specify mood, tempo, instrumentation, and `instrumental/no vocals` when lyrics are unwanted.

Tempo is named as a required aspect but the repo gives no BPM values or key/scale vocabulary anywhere , this is genuinely thin.

Mood20.2% of audio prompts

The feeling.

Pairs with tempo and key. Together those three settle most of the composition.

Values

warm, playful, cheerful, buoyant, wholesome, ambient, uplifting, driving, triumphant, minimal, dreamy, suspenseful

Also used in production skills, not counted

material, era, energy, tonal language

Fragments to drop in

• Cheerful, wholesome and playful.
• Calm and spacious.
• Suspenseful and restrained.

From production skills

• Preserve the game's STYLE FORMULA conceptually through material, era, energy, and tonal language even though audio is not visual.

The visual style contract is carried into audio prompts by translation, not by pasting the formula.

Arrangement and dynamics10.4% of audio prompts

How the piece develops.

Without this you get four bars looped flat. Naming a build and an ending is what makes it usable against picture.

Values

airy, midrange, headroom, no abrupt hits, narration-friendly, smooth transitions, gentle dynamics

Also used in production skills, not counted

-6 dBFS, -10 to -12 dBFS, -18 to -20 dBFS, -3 dBFS true peak, short fades at loop boundaries

Fragments to drop in

• Airy and narration-friendly, gentle dynamics, no abrupt hits.
• Build subtle energy throughout, then a celebratory lift at the end.
• Clear midrange space left for a voice.

From production skills

• Voice around -6 dBFS.
• SFX around -10 to -12 dBFS.
• Music around -18 to -20 dBFS.
• Final true peak at or below -3 dBFS.
• Voice stays above SFX; SFX stays above music.
• Never stack raw model outputs at full gain.

Post-generation mixing, not prompt text, but it is the only mix guidance in the corpus.

Vocals62.6% of audio prompts

Whether anything sings or speaks.

Named in 63% of audio prompts, almost always as a prohibition. Generators add humming and wordless vocals unprompted.

Fragments to drop in

• No vocals, no lyrics, no spoken words.
• Wordless humming only, no intelligible words.
• No prominent lead melody competing with the narration.

Ending

How it stops.

The most common practical complaint about generated audio is that it fades or stops mid-phrase. Ask for a resolution.

Fragments to drop in

• End on a clean major-key resolution, no fade.
• Ring out and decay naturally.
• Hard stop on the final beat.

Voice and spoken linefrom production skills

The prompt IS the line to be spoken; the voice itself is selected by id, never described in prose. Adjust --speech_rate modestly rather than rewriting a take.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

voice_type preset|element, voice_id, speech_rate, one voice per speaking entity, exact line, performance direction

Fragments to drop in

• The gate is open. Move!
• same voice, calmer delivery
• Voice prompts contain the exact line; performance direction belongs in the model-supported prompt, not undocumented flags.
• Voice: lock one voice per speaking entity and reuse it. Keep lines short enough not to block gameplay.

Narration writing (spoken script)from production skills

Narration is the one thing written in the user's chosen language; every image and video prompt stays English.

Documented in production prompt skills rather than measured in the corpus, so there is no frequency for it. Treat it as practice worth copying, not as a count.

Values

20-24 words, about 8-9 seconds, under roughly 9.5 seconds, plain spoken text only, no timecodes, emotion cues, parentheticals, or stage directions, Spell numbers out, hook through payoff, ~150 words per minute

Fragments to drop in

• Write one plain line per clip, sized for about 8-9 seconds and normally 20-24 words. Keep every take under roughly 9.5 seconds.
• Use no timecodes, emotion cues, parentheticals, or stage directions.
• Spell numbers out.
• Set tone through word choice and concrete detail.
• Never say "in this video."
• For four and a half thousand years, the pyramids of Egypt have stood against the desert, silent and immense.
• Write scripts for speech, not text: short sentences, natural pauses, ~150 words per minute target. A 60-second video = ~150 words.
• Don't pad scripts to fit duration. The model paces itself; over-stuffed scripts get rushed delivery.