01 The Director's Brief
Kling 4.0 behaves less like a clip generator and more like a small shot crew. It plans multiple shots by default, expands a thin prompt into blocking and camera moves, and accepts up to 8,000 tokens of direction. Stacking adjectives no longer helps. What helps is a brief: what is in frame, what happens in what order, how the camera behaves, what the sound does, and what must not change. This chapter builds that brief from one sentence to a full shot plan.
1.1 The Five-Part Brief
Every strong Kling 4.0 prompt answers five questions. Scene sets the world; action moves it; references cast it; camera and audio direct it; constraints protect it. Only scene and action are mandatory — a reference-free text-to-video prompt is still a complete brief. Write in this order and you can stop at any line.
[Scene] + [Action] + optional [References] [Camera] [Audio] [Constraints]
Scene
Where and when the shot happens, plus the visual tone.
Action
What the subject does, in order, with body parts and pace.
References
The assets that define identity, product or setting: @Image1, @Video1, @Audio1.
Camera
Shot size, lens, camera path and pace. One move per shot.
Audio
Dialogue, ambience, SFX and music. Stereo by default.
Constraints
What must stay fixed, and the rare negative instruction worth writing.
A Sparse Brief the Model Completes
Prompt
A street violinist packs up her case as the first heavy drops of rain hit the pavement. The camera holds on her hands, then tilts up to the awning as the street empties. Quiet, wet, end-of-day mood.
A short prompt is a valid starting point: Kling 4.0 expands it with blocking, shot sizes and pacing. Treat the result as a first draft. When a specific look matters, fill in the camera and audio lines yourself. Suggested: 10 seconds, 1080p, 16:9.
1.2 The Fifteen-Block Director's Brief
When the shot has to be repeatable, write the long brief. Kling 4.0's sample prompts often use fifteen blocks: scene context, active references, location map, first frame, shot list, optics, camera behavior, action timeline, physics, lighting, audio, character acting, style, quality and positive constraints. You will rarely need all fifteen. Load the blocks that carry the shot — a multi-shot action sequence needs the timeline and shot list; a product close-up needs optics and lighting.
A Six-Block Chase Brief
Prompt
SCENE CONTEXT: A 10-second rooftop chase at blue hour. Exactly two people: a courier in a yellow rain shell and a pursuer in a dark coat. The roof is wet, gravel-covered and cluttered with vents. ACTIVE REFERENCES: @Image1 defines the courier's face, build and yellow shell. @Image2 defines the rooftop layout: water tank left, fire escape right, skylight center. SHOT LIST: Shot 1 — 0 to 3.5s, low tracking shot behind the courier. Shot 2 — 3.5 to 7s, hard cut to the pursuer's boots splashing through puddles. Shot 3 — 7 to 10s, the courier vaults the skylight, lands, and looks back. CAMERA: Shot 1, 24mm, chest height, handheld pursuit. Shot 2, 50mm, ground level, shallow focus. Shot 3, 28mm, orbit left as she lands. One camera move per shot. ACTION TIMELINE: 0–3.5s the courier sprints, dodges a vent, glances back. 3.5–7s the pursuer closes in, coat snapping. 7–10s she plants one hand on the skylight frame and clears it; the pursuer stops at the edge. PHYSICS AND LIGHTING: Rain streaks, wet gravel kicks, fabric and hair lag behind each turn. Cool blue ambient light; warm stairwell light spills from the skylight. AUDIO: Stereo. Near: boots on gravel, breath, rain on metal. Far: city traffic, a distant siren. A low percussive score enters at 3.5s and cuts at the landing. CONSTRAINTS: The courier's yellow shell and face stay identical across all three shots. Keep the rooftop layout consistent. No subtitles.
A compact version of the fifteen-block template: six blocks that control a multi-shot action scene. On Kling 4.0 you can stretch the same timeline to 20 or 30 seconds; Try this loads the 10-second version. Requires two uploaded images.
1.3 Multi-Shot Is the Default
Kling 4.0 splits a scene into shots and organizes pacing without being asked. That is useful, but it is not the same as directing the cuts. When you need control, write a shot list: number each shot, give it a time range, one camera move and one action. Hard cuts are the default; write "hard cut" when the transition matters. Keep screen direction consistent so the audience never has to re-orient.
Three Shots, One Coffee Ritual
Prompt
Three shots, 10 seconds, one continuous morning ritual. Shot 1 — 0 to 3s: close-up of a hand grinding coffee beans; the burr turns, grounds fall into a glass jar. Static camera, shallow focus, warm window light. Shot 2 — 3 to 6.5s: hard cut to a kettle pouring a thin stream into a pour-over cone; steam curls toward the lens. Camera slowly pushes in. Shot 3 — 6.5 to 10s: hard cut to the finished cup on a wooden counter, a hand lifts it toward the window; the city outside is out of focus. Camera holds, then racks focus to the window. AUDIO: Stereo. Grinder crunch, water pour, ceramic clink, faint street ambience. No music. CONSTRAINTS: Same kitchen, same warm morning light in all three shots. Keep the cup and jar consistent.
A shot list is the most reliable way to get a multi-shot sequence. Number the shots, give each one a time range, and write one camera move per shot. Suggested: 10 seconds, 1080p, 16:9.
02 Omni Reference
Omni Reference is the 4.0 entry point for anything that must look like something: up to ten images, five videos, seven elements, three voice profiles and fifteen assets in total. Uploads are numbered in order — @Image1, @Image2, @Video1 — and the prompt is where you assign each one a job. A reference without a role is a suggestion; a reference with a role is a constraint. This chapter covers the four patterns that matter: casting, multi-character scenes, camera borrowing and storyboards.
2.1 Give Every Asset a Job
Write the role before the scene: "Use @Image1 as the actor, @Image2 as the product, @Image3 as the set." Then state what must not carry over. When two assets overlap — a product photo and a set photo both contain a table — say which table wins. This one sentence prevents most identity drift.
[@Image1 role] + [@Image2 role] + [@Video1 role] + [New scene and action] + [What stays consistent]
Assets
Each asset and its job: identity, product, set, costume, motion or voice.
New scene
The scene, action and camera the assets are dropped into.
Consistency
What must stay identical across the shot, and what must not carry over.
Camera
A camera move that suits the new scene, not the reference.
Audio
Dialogue, ambience or a voice profile, if the scene needs sound.
An Actor, a Product and a Set
Prompt
Use @Image1 as the actor: a woman in her early thirties with short curly hair, a linen shirt and a calm, direct expression. Use @Image2 as the product: a matte black electric kettle, including its handle shape and small amber indicator light. Use @Image3 as the set: a bright Scandinavian kitchen with white oak counters and a large window on frame-left. The woman stands at the counter, fills the kettle from the tap, sets it on its base and switches it on. The camera starts at a medium shot, pushes in to the indicator light as it turns amber, then pulls back as steam rises. Keep her face, hair and shirt identical to @Image1, the kettle identical to @Image2, and the kitchen layout identical to @Image3. Do not carry the table from @Image2 into the new kitchen. No subtitles.
Three roles, one conflict rule. Naming what should not carry over is what keeps a product reference from dragging its original background into the new scene. Requires three uploaded images.
2.2 Cast Multiple Characters
For a two-person scene, give each character their own image and their own line of description. Then lock the relationship: who stands where, who looks at whom, and what must not swap. A group photo is a poor reference for two people; separate, clear images work better. If a character appears in more than one shot, repeat their identity description in the shot list.
A Kitchen Argument, Two References
Prompt
Use @Image1 as Maya: mid-forties, dark hair in a loose bun, green apron, expressive eyebrows. Use @Image2 as Daniel: late forties, grey stubble, blue shirt with rolled sleeves, tired eyes. Maya and Daniel stand on opposite sides of a kitchen island. Maya sets down a mug too hard; Daniel looks up from a laptop. They argue in low voices, then Daniel closes the laptop and leaves frame-right. The camera holds a two-shot, then slowly pushes toward Maya as she exhales. Keep Maya's face, hair and apron identical to @Image1 and Daniel's face, stubble and shirt identical to @Image2 in every shot. Do not swap their clothing or facial features. Maya stays frame-left, Daniel stays frame-right. No subtitles.
Separate references per character, plus a spatial rule — "Maya stays frame-left, Daniel stays frame-right" — do most of the consistency work. Requires two uploaded images.
2.3 Borrow Camera Moves from a Video
A reference video can lend its camera path, pacing, choreography or attack timing without lending its subject. Write what the video is for, then explicitly replace everything else. If the reference contains a person, say that the person's face and clothing do not carry over. This is the fastest way to reproduce a move you have seen but cannot describe.
A Camera Move, a New Car
Prompt
Use @Video1 only for camera movement and pacing: a low tracking shot that accelerates alongside a moving subject, then rises into a three-quarter orbit. Use @Image1 as the car: a silver 1970s coupe with chrome bumpers. Use @Image2 as the setting: a rain-slicked coastal road at dusk, cliffs on frame-right. Replace the subject in @Video1 with the car from @Image1 and the setting with the road from @Image2. Match the camera's speed and rhythm to @Video1, but do not copy its subject, background or color grade. The car keeps the shape, chrome and paint of @Image1 throughout. No subtitles.
Three clauses do the work: what to borrow, what to replace, and what must not carry over. This example needs one reference video and two images, so Try this stays unavailable until Kling 4.0's video-reference slot lands in the generator.
2.4 Use a Storyboard as a Shot Map
A storyboard image is not a style reference. It is a map: shot order, framing, camera path, screen direction, emotional progression. Tell the model to read the board for structure and explicitly not to render the board itself. This is especially useful for action and comedy, where timing matters more than any single frame.
A Four-Panel Storyboard, Rendered as a Scene
Prompt
Use @Image1 as a four-panel storyboard. Read it only as a shot map: panel 1 is a wide shot of a bakery at dawn; panel 2 is a close-up of a baker sliding a tray into the oven; panel 3 is an over-the-shoulder shot of a customer pointing at a pastry; panel 4 is a wide shot of the customer leaving with a paper bag. Do not render the storyboard, its panel borders, arrows or annotations. Render the four shots as live action in that order, with hard cuts between them. The bakery interior stays consistent: pale green tiles, brass fixtures, warm pendant lights. The baker wears a white apron and a dark cap; the customer wears a rust-colored coat. No subtitles.
The negative line is essential here: without "do not render the storyboard" the model may treat the panel borders as part of the art direction. Storyboard references are display-only for now; Kling 4.0's reference slot accepts them when it launches.
03 Keyframes
Keyframes are the other way to control time: instead of describing motion, you pin stills to moments. Kling 4.0 accepts up to ten keyframes in one task, including first frame, last frame, first-and-last, and middle frames anywhere on the timeline. Use them when the destination matters as much as the journey — a character's entrance, a product's final position, a scene change that must land exactly. Keyframes and Omni Reference are separate entries: you cannot combine them in one generation, and keyframes do not accept reference assets.
3.1 First and Last Frame
The first-and-last formula is one sentence of setup plus one sentence of transformation. Name the frames, state what happens between them, and say what must stay fixed. Do not describe the stills in detail — the images already do that. Describe the motion, the pace and the one thing that changes.
Use @Image1 as the first frame and @Image2 as the last frame. [What happens between them] + [What stays fixed]
First frame
The opening still: composition, subject, lighting.
Last frame
The closing still: the state the shot must land on.
Transformation
The visible change between the two frames, in order.
Fixed elements
What must not change: identity, set, product shape.
Audio
Sound that supports the transformation.
The Same Street, Day to Night
Prompt
Use @Image1 as the first frame: a narrow Lisbon street in daylight, laundry lines overhead, a tram stop on frame-right. Use @Image2 as the last frame: the same street after dark, lit by warm shop windows and a single streetlamp. In between, the sun drops, the sky turns from pale blue to deep indigo, shop signs flicker on one by one, and a tram passes from frame-left to frame-right. The camera holds its position; only the light and the tram move. Keep the buildings, laundry lines and tram stop identical to the two frames. No subtitles.
The camera stays still so the transformation reads. The prompt describes only what changes between the two stills: light, signs, tram. Requires two uploaded keyframes.
3.2 Up to Ten Keyframes
With more than two keyframes, you are writing a state progression, not a transformation. Give each keyframe a moment and a visible state: frame 1 assembled, frame 3 half-disassembled, frame 5 packed in a box. The model interpolates the steps; your job is to make each step unambiguous. Keep the camera and setting stable unless the change itself is the camera move.
A Product Assembling in Five States
Prompt
Use @Image1 through @Image5 as five keyframes on the timeline. @Image1 is the product fully assembled on a workbench. @Image2 is the outer shell lifted two centimeters above the base. @Image3 is the shell removed, revealing the battery and circuit board. @Image4 is the battery sliding out toward frame-right. @Image5 is the empty shell and battery laid side by side. Interpolate continuously from @Image1 to @Image5 over 10 seconds. The camera holds a fixed three-quarter view above the workbench; the product stays centered. Lighting stays constant: a soft overhead key with a cool rim light. No hands enter the frame. No subtitles.
Five keyframes, one fixed camera, one continuous action. The model needs each state to be visually distinct; ambiguity between frames produces morphing. This example is display-only until Kling 4.0's multi-keyframe slot is available.
04 Sound, Voice and On-Screen Text
Kling 4.0 generates two-channel stereo with the picture and improves lip sync. Dialogue, ambience and music are written like the rest of the brief: name what you want, say where it sits in the mix, and keep the line in the language you want performed. On-screen text is a separate skill: specify the exact words, the placement and the behavior, then check it in motion.
4.1 Build a Stereo Soundscape
Write sound in layers. The nearest sound gets the most detail; midground sounds establish the place; a distant bed keeps the scene from feeling empty. Assign left and right when direction matters. If you want a quiet scene, say so explicitly — an empty audio prompt invites invented music.
A Market Waking Up
Prompt
A covered market at 6 a.m., shutters rolling up one by one. A vendor arranges oranges into a pyramid; another drags a crate across wet concrete. The camera walks slowly down the aisle at eye level, past hanging bulbs and crates of greens. AUDIO: Stereo. Near, frame-left: a metal shutter rattling up. Near, frame-right: oranges thumping into a crate. Midground: a radio playing a low, muffled bolero. Far: a street sweeper and pigeons. No music beyond the radio. Keep dialogue out of the scene.
Near, midground and far are the three layers; left and right place the listener. "No music beyond the radio" is a constraint, not a negative prompt. Suggested: 10 seconds with native audio.
4.2 Dialogue in Nine Languages
Kling 4.0 speaks Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, with regional accents and dialects including Cantonese, Sichuanese, Beijing, Taiwanese Mandarin, Northeastern Mandarin, American, British and Indian English. Write the line in the target language inside quotation marks, name the speaker, and add a tone cue. Keep lines short: one exchange per beat is easier to lip-sync than a monologue.
A Three-Language Exchange at a Counter
Prompt
A bright bakery counter. A Japanese tourist in a beige coat, a Spanish barista and a Chinese baker stand in that order from frame-left to frame-right. The camera holds a medium shot. Tourist (in Japanese, slightly hesitant): 「クロワッサンはまだありますか？」 Barista (in Spanish, warm and quick): «Sí, quedan tres. ¿Quiere el último?» Baker (in Chinese, calling from the back): 「刚出炉的，给你留了一个。」 The tourist smiles and nods. The barista wraps a croissant in paper and slides it across the counter. Stereo audio, bakery ambience, no background music. Keep each speaker's mouth in sync with the language they speak.
Three languages, one shot, three short lines. Short exchanges keep lip sync readable; naming the language before each line is what makes the model switch voices. Quoted lines stay in their original languages in every locale of this guide.
4.3 Direct a Voice Profile
A voice profile is a reference asset, not a text description. Upload five to thirty seconds of clean speech and cast it like an actor: "Use @Audio1 as the narrator's voice." You can stack up to three voice profiles per task. Describe delivery — pace, pitch, energy — in the prompt; the profile supplies identity, not performance.
A Documentary Narration
Prompt
Use @Audio1 as the narrator's voice: a low, unhurried baritone with a slight rasp. The scene is a slow aerial shot over a salt flat at sunrise, the surface cracking into pale polygons. The camera drifts forward at walking pace. Narrator (measured, intimate, in English): "By the time the sun clears the ridge, the water has already gone. What's left is a map of where it stood." AUDIO: Stereo. Wind across the flat, distant birds, the narrator centered and dry, no music. Keep the voice identical to @Audio1 throughout.
The profile carries identity; the prompt carries performance. Upload a clean five-to-thirty-second clip, then describe pace and tone. This example needs an audio reference, so it is display-only until Kling 4.0's voice-profile slot is live.
4.4 Render Text That Holds Still
Kling 4.0 renders text in the same nine languages, plus emoji and logos, and is designed to keep it stable through camera and subject movement. Write the exact characters in quotes, say where they sit, and specify the material. For subtitles, give the position, size, line count and timing — "subtitles" alone is not a direction. When you do not want burned-in text or music, "no subtitles, no background music" is the rare negative line worth writing.
A Bookshop Sign at Closing Time
Prompt
A small independent bookshop at closing time. The camera slowly pushes toward the window from the street. Behind the glass, a shopkeeper stacks books; warm pendant lights reflect on the pane. The hand-painted sign above the door reads "MOLYIN BOOKS" in cream serif lettering on dark green wood. The sign stays sharp and legible as the camera moves. A small paper notice on the door reads "CLOSED" in black caps. AUDIO: Quiet street ambience and a latch turning. No subtitles, no background music.
Exact characters, exact material, exact position. Test text in motion: if it wobbles, shorten the line or lock the camera. The two allowed negatives are the only ones this prompt needs.
05 Editing, Recreation and Long Takes
The last chapter is about working with existing footage and with time. Kling 4.0 can edit up to five clips totaling thirty seconds, borrow the pacing of one video for a new cast, and generate a single take up to thirty seconds. Each has its own prompt grammar; using the wrong verb is the most common failure.
5.1 Edit, Don't Reference
For an edit, the verb matters. Write "Edit @Video1" and then state the change. If you write "Reference @Video1", the model may generate a new shot in the style of the video instead of editing it. Add "Keep everything else unchanged" whenever the rest of the frame must survive. Supported edits include adding an element, replacing a character or scene, replacing multiple characters, removing an element and changing the style.
Edit @Video1: [specific change] + Keep everything else unchanged
Source clip
The clip being edited. Use @Video1 in the prompt.
Change
One specific edit: add, replace, remove or restyle.
Preservation
What must not change: framing, performance, background, grade.
Reference
An image that supplies the replacement character, product or texture.
Replace the Drink, Keep the Shot
Prompt
Edit @Video1: replace the glass of orange juice on the table with the ceramic cup from @Image1. Match the cup's size, handle and glaze to @Image1. Keep everything else unchanged: the same table, the same hand entering frame, the same window light, the same camera movement and timing. The cup should cast a plausible shadow and reflect the window. No subtitles.
One change, one preservation clause. "Keep everything else unchanged" is not filler — without it, the model may re-stage the shot. This example needs a source clip and a reference image, so Try this is display-only.
5.2 Recreate the Rhythm, Replace the Cast
Creative recreation borrows narrative pacing, shot composition or presentation style from an existing video and applies it to new characters, products or scenes. It works best when the reference and the target are close in pacing and shot structure. Write what to borrow, what to replace, and what must not carry over — especially the original faces and setting.
Recast a Two-Person Scene
Prompt
Use @Video1 for narrative pacing and shot composition only: a two-person conversation across a desk, opening on a wide shot, cutting to a close-up on each speaker, ending on a shared laugh. Use @Image1 as the new lead: a woman in her sixties with silver hair and round glasses. Use @Image2 as the new scene: a sunlit pottery studio with shelves of unglazed bowls. Replace both people in @Video1 with the woman from @Image1 and a colleague of similar age and build. Keep the shot order, camera positions and timing of @Video1. Do not carry over the original faces, clothing or office set. The woman's identity stays identical to @Image1 in every shot. No subtitles.
Borrow structure, replace identity. The closer the new scene's blocking is to the reference, the more reliably the recreation lands. Display-only until Kling 4.0's video-reference slot is live.
5.3 Direct a 30-Second Take
A single Kling 4.0 generation can run from 3 to 30 seconds in one-second steps, with a default of 10 seconds. Longer is not automatically better: five seconds is enough for a single action, ten to fifteen holds a short scene, and twenty to thirty is for a continuous take with a beginning, a turn and an ending. Write the beats on a timeline, not as a list of adjectives. If the shot needs more than thirty seconds, that is a multi-shot sequence or a future extension, not a longer prompt.
A 15-Second One-Take in a Kitchen
Prompt
A 15-second continuous take, no cuts. A cook in a white T-shirt works at a kitchen island, making pasta by hand. The camera starts at the far end of the counter and moves slowly toward her. 0–3s: she dusts the board with flour and cracks two eggs into a well. 3–6s: she pulls the flour into the eggs with a fork, then switches to her hands. 6–10s: the dough comes together; she kneads it with the heel of her hand, turns it, repeats. The camera is now at a medium close-up. 10–13s: she wraps the dough and sets it aside; a small pot of water starts to boil behind her. 13–15s: she looks toward the pot, wipes her hands on a towel and reaches for the lid. The camera holds as steam rises. AUDIO: Stereo. Fork on ceramic, flour on wood, dough pressing against the counter, water coming to a boil. No music. No subtitles. CONSTRAINTS: One continuous shot. Same cook, same kitchen, same light throughout.
A timeline turns a long prompt into a performance. Kling 4.0 can run this at 20 or 30 seconds; Try this loads a 15-second version, the longest Kling 3.0 can generate today.