VidoGridHow Kling AI prompts are structured
Start with one plain sentence naming the subject, setting, action and mood. According to the 4.0 guide, a short idea is enough for a first draft: the model fills in staging and camera work, and when your description is already specific it follows it instead of improvising. Then add only the details that change the result, such as camera, timing, sound and anything that must stay fixed.
The older Kling text-to-video guide orders a prompt as subject, movement and scene, then camera, lighting and atmosphere. That order still works as a checklist.
Swap abstract words for things a camera can see. "Premium" or "cinematic" gives the model nothing to draw; light direction, surface texture and background objects do. Kling 4.0 accepts prompts up to 8,000 tokens, and its reference template uses four blocks: asset roles, overall settings, a timed sequence of events, and audio.
A matte black coffee grinder on a pale oak counter. Low sun from the left lights the brass crank; the wall behind is out of focus. A hand turns the crank three times and fresh grounds fall into the glass jar. Slow push-in to a close-up. Audio: beans crunching, a kettle hissing off screen. No background music. No subtitles.
Camera language, shots and timing
Camera direction is where many Kling AI prompts stay vague. Kling 4.0 plans multiple shots by default, so ask for "one continuous shot, no cuts" when you want a single take. In VIDEO 3.0, multi-shot is a switch, and Custom Multi-Shot lets you set the number and length of the shots.
Give each shot a framing (wide, medium, close-up), an angle (low, eye level, top-down) and a movement (push-in, pull-out, pan, tracking, orbit). Time ranges such as 0–4s and 4–9s are the clearest way to say when an action, a spoken line or a cut should happen.
15 seconds, realistic. A bus shelter at night in light rain. 0–5s: wide shot, a woman in her sixties in a yellow raincoat reads a paper timetable. 5–10s: medium shot, a teenage boy carrying a skateboard runs in, out of breath; she glances at him and says dryly: "Late again." 10–15s: close-up, he grins: "Then we both are." Headlights sweep across their faces. Audio: rain on the roof, no music.
Reference images, videos and keyframes
Kling 4.0 has two entry points that cannot be mixed in one task. Omni Reference, the default, takes images, videos, voice profiles and saved elements (reusable characters or objects), and also handles plain text. Keyframes takes images only and pins them to the timeline as first, last or in-between frames.
Omni Reference numbers uploads as @Image1, @Video1 and so on. Give each one a job, such as "keep her face" or "camera path only"; assets without a role leave the model guessing. To change a clip rather than imitate it, write "Edit @Video1", say what changes, and add "Keep everything else unchanged." Writing "Reference @Video1" produces a new video instead.
- Up to 15 assets per task: at most 10 images, 5 videos, 7 elements and 3 voice profiles
- Videos you reference or edit share one quota: 5 clips and 30 seconds in total
@Image1 is the character: keep her face, silver hair and green jacket. @Image2 is the greenhouse: keep its layout and hanging lamps. @Video1 sets the camera path and pacing only; copy none of its people or setting. She walks between tomato plants, stops, and lifts a leaf to check underneath. Handheld camera at shoulder height.
Dialogue, sound and on-screen text
Kling 4.0 generates speech, effects and music with the picture, so Kling AI prompts should treat audio as its own block. Put each line in quotation marks, in the language you want to hear, attached to a named speaker, and describe the voice: age, tone, pace and emotion. The 4.0 guide lists nine languages, including English, Chinese, Spanish and Hindi; VIDEO 3.0 supports five.
Music and sound effects cannot be uploaded; the only audio input is a voice profile, so describe the ambience and score in words. Signs, logos and subtitles can be requested too; for subtitles, give size, position and timing. Write "no subtitles" and "no background music" when you want neither, since the model may add both by default.
21:9, 10 seconds. Aerial shot gliding over a narrow fjord at dawn, mist on the water, one red fishing boat leaving a pale wake. At 6 seconds the camera tilts up and the words "NORTH COAST" appear in bold white capitals at the center, fading out at 9 seconds. Audio: wind, gulls, a low engine hum.
Settings that change results
These options sit outside the text of Kling AI prompts but change what the same words produce. Values come from the official guides; the product interface has the final say.
- Duration: 3–30 seconds on Kling 4.0 (default 10), 3–20 on 4.0 Flash, 3–15 on VIDEO 3.0
- Resolution: 720p, 1080p or 4K on Kling 4.0; 720p on Flash; 720p or 1080p on VIDEO 3.0
- Aspect ratio: 21:9, 16:9, 1:1, 9:16 or Auto on Kling 4.0; in Keyframes it follows the first image
- Flash suits quick drafts but does not currently take first-and-last frames or multiple keyframes
Common mistakes in Kling AI prompts
When a result misses, the 4.0 FAQ suggests checking two things first: whether each asset has a clear role, and whether the prompt leans on abstract terms.
- Mood words: replace "cozy" with a knitted throw, one low lamp and rain on the window
- Negatives: write "seen from behind" rather than "do not show her face"
- Mixed styles: for animated, clay or felt looks, add that the result should not be photorealistic
- Stacked adjectives: cut words that change neither the picture nor the sound
- Loose counts: state numbers outright, as official examples do with "exactly one" character