Kling AI Prompts Cheatsheet - Kling Video Camera Moves & Structure
Kling clips usually fail on camera motion rather than image quality: a single prompt that demands a push-in, a pan and an orbit at once leaves the model picking one and smearing the rest into jitter. This sheet breaks camera grammar into eight enumerable moves - push, pull, pan, tilt, tracking, orbit, roll, zoom and static - each annotated with its start and end frame behaviour, so you can decide which one actually fits the shot instead of stacking verbs. A fixed six-section structure then separates subject, action, environment, lighting, camera and style, which is what stops the model from silently dropping half your intent. For reproducible results, the camera_control block with its six axes in [-10, 10] is the ground truth; when a text description drifts, reconcile against those numeric values. The rules of thumb are short: split any composite move into separate single-move shots, and lock duration (5-15s) plus aspect ratio (9:16 or 16:9) before writing a single word of prompt, because an undecided timeline is where late-shot collapse comes from.
Camera Moves 5
push-in / pull-outpan (left-right) / tilt (up-down)dolly (fore-aft) / orbit (around subject)crane (up-down) / tracking (follow)static / handheldCamera Rules 5
one move per shot ruleKling 3.0 multi-shot allows different moves per shotPut camera move in its own segmentModerate move amplitude, avoid vertigoStatic subject + slow push-in feels premiumSix-Segment Structure 5
[Subject][action/motion][camera movement][environment][lighting] + [style]API Params 5
camera_control: { horizontal/vertical/pan/tilt/roll/zoom ∈ [-10,10] }aspect_ratio: 16:9 / 9:16 / 1:1duration: 5-15smodel: kling-v2-6 / kling-v3negative_prompt to exclude unwanted elementsSubject & Motion 5
Be specific about the subject (age/outfit/material)Write continuous verbs, avoid ambiguityAdd mood/state (pensive/cheerful)Label relations for multiple subjects (A looks at B)Speed words (slowly/rapidly)Templates 5
Product: {item} rotating showcase, slow orbit, studio light, clean bgPerson: {who} walks slowly to the window, push-in, soft light, cinematicLandscape: sea of clouds over mountains, pan, golden dawn, wide shotduration 10s, 16:9, model kling-v3negative: distorted face, flicker, watermarkCommand Examples
Product showcase: slow orbit under studio light
[Subject] a matte-finish stainless steel thermos flask
[action/motion] rotates slowly to show its form
[camera movement] slow orbit
[environment] clean bg
[lighting] + [style] studio lighting, clean background, minimalismA slow orbit on a static subject is the cinematic pairing; naming the material (matte metal) controls texture more reliably than naming a brand.
Character beat: walking to the window with a push-in
[Subject] a woman in her thirties wearing a beige knit sweater
[action/motion] walks slowly toward the window
[camera movement] push-in
[environment] an old apartment in the early morning
[lighting] + [style] soft light, cinematicAge and clothing make the subject concrete, the camera move stands alone as a single move per the one-move rule, and the speed word sets the pace.
Landscape: lateral pan under golden morning light
[Subject] a continuous mountain range with a sea of clouds
[action/motion] the cloud layer drifts slowly
[camera movement] pan
[environment] wide shot across an open ridge
[lighting] + [style] golden morning light, crisp, deep focusA horizontal pan with a wide shot is the standard landscape pairing, and keeping the action on the clouds stops in-scene motion from competing with the camera.
Settle duration, aspect ratio, model, and negative prompt in one block
duration: 10s
aspect_ratio: 16:9
model: kling-v3
negative_prompt: distorted face, flicker, watermarkFix duration and aspect ratio before writing prose, otherwise the back half of the clip drifts; 9:16 for vertical shorts, 16:9 for product and landscape, with kling-v3 favoring quality and kling-v2-6 favoring speed.
Common Pitfalls
- Cramming more than one camera move into a single shot (push-in and orbit together), which is the single most common source of broken footage; split multi-move sequences into numbered multi-shots.
- Asking for a locked-off camera without writing static or zeroing camera_control, or relying on a phrase like stable camera, which the model often reads as gentle handheld and turns into persistent micro-shake.
- Describing the subject as just a person with no age, clothing, or material: the vaguer it is, the less controllable the result.
- Leaving relationships unstated when there are multiple subjects, so the model decides who is who and frequently reverses the intended interaction.
- Filling negative_prompt with only static-quality words and omitting the three motion defects, distorted face, flicker, and watermark, which then show up at the usual rate.
Tips
- Kling prompts center on "six-segment structure + one-move-per-shot": camera move in its own segment, one motion per shot; use Kling 3.0 multi-shot to set different moves per shot.
- The API controls lenses precisely via camera_control horizontal/vertical/pan/tilt/roll/zoom (∈[-10,10]) — more reproducible than prose like "camera slowly pushes in".
- Portrait shorts use 9:16, product/landscape use 16:9; duration 5-15s; prefer kling-v3 (quality) or kling-v2-6 (speed).
- Video negatives target motion defects: distorted face / flicker / watermark are frequent fixes.
FAQ
What is the one-move-per-shot rule and why is it non-negotiable?
One camera move per shot. The grammar is enumerable: push-in / pull-out, pan (horizontal) / tilt (vertical), dolly (depth travel) / orbit (around the subject), crane (rise) / tracking (follow), and static / handheld. Ask for push-in plus pan plus orbit in one sentence and the model can only pick one; the rest degrades into jitter. When you genuinely need several moves, use Kling 3.0 multi-shot and write "Shot 1: ... Shot 2: ...", keeping each numbered shot to a single move.
In what order should the six prompt sections be written?
Follow [Subject], then [action/motion], then [camera movement], then [environment], then [lighting] plus [style]. Camera movement gets its own section precisely so it is separated from subject and action and the model reads it as the camera moving rather than the scene shifting. Make the subject concrete (age, clothing, material), express action as a continuous verb phrase (walks slowly toward the window), state relationships when there are multiple subjects (A is looking at B), and add a speed word (slowly, rapidly) to set the pace.
How do I pick among the eight camera moves without guessing?
Decide from the camera-to-subject relationship rather than by stacking verbs. Subject moving means tracking or orbit. Camera changing its distance to the subject means push-in / pull-out or dolly. Camera turning away from the subject means pan or tilt. A change in height means crane. If nothing should move, say static explicitly and reserve handheld for deliberate shake. When the text description and the API disagree, the numeric camera_control values win.
How should camera_control be filled in, and is there a reference range for the values?
It exposes six axes — horizontal, vertical, pan, tilt, roll, zoom — each bounded to [-10, 10], and it is the reproducible way to control move magnitude. Choose the magnitude tier before choosing the move type, because excessive speed reads as motion sickness, which is why keeping the amplitude moderate is a rule in its own right. A slow push-in on a static subject is the familiar cinematic pairing and gives better results than stacking several moves.
What should negative_prompt actually exclude in video?
Motion artifacts rather than static image quality. The standard set here is distorted face, flicker, and watermark: faces warping through movement, frame-to-frame instability, and burned-in branding. Those three account for most failed generations. Negative prompting only defuses defects; if the five positive sections (subject, action, camera, environment, lighting and style) are incomplete, no amount of exclusion rescues the clip.
Official References
Each command links to its official documentation below, so you can verify the latest usage and read deeper.
Maintained by LaoHand
Publicly updated on Aug 21, 2026, continuously proofread against official docs.
Contact Us
Wrong command or description? Send us corrections, business inquiries or product feedback by email.
Contact Us