Kling v3 prompt format
How to write a prompt Kling v3 reads the way you meant it, with each character and
product held to its element. Kling opens exactly on the first frame and animates
forward: the prompt says how the scene evolves from it, not a fresh description of it.
The request is in kling-v3-image-to-video. Worked examples are in
references/examples.md.
How to handle @ names, entrances, subjects not in the frame and reference style is
shared by every video model: guide-animate-a-shot. Never put a person's name or a
handle in the prompt outside a quoted line: Kling renders it as on-screen text.
Layout
Two parts, the prompt then the sound line:
<prompt>
Background audio: <sound>
Elements
Each subject with reference pictures goes in the request as an element, and the prompt
cites it as @Element1, @Element2, … in the order of the elements list. Up to 3
elements. The first frame has no label.
- Cite a subject by its label at every mention. At first mention follow it with a few plain words: "@Element1, the woman,".
- A subject in the first frame is tied to where the frame shows them, at first mention: "@Element1, the small figure in the yellow raincoat at the end of the pier,". Take the where and how from the frame only.
- When another person or animal is in the shot, start each sentence about a subject with its label again, not a pronoun; within the sentence a pronoun is fine.
- Never a possessive label (
@Element1's): write "his", "her", "its". - A product is an element like a person: "@Element2, the black tumbler in her hand,".
- Never cite a label that isn't in the list, and never
@Image.
The prompt
Two to five short sentences of plain prose, under 130 words. No Subject: or Camera:
labels, no lists of moods or adjectives, never one long chain of clauses.
- One sentence anchoring the first frame: the setting and where each subject already in frame is. Name a non-live-action style here (2D cartoon, 3D CG, claymation).
- The action, in the order given. Subject motion explicit and physical; camera motion explicit (holds static, pushes in, pulls back, pans, tilts, tracks alongside, orbits, rises, dives, racks focus), with how fast when it matters ("slowly").
- Where the shot lands.
One continuous take: never a cut, a second shot or a timestamp. Big, clear, realistic motion. On-screen text in double quotes, verbatim.
Dialogue
Only for words the person actually wrote.
- Exact words: the speaker's action first, then
[<speaker>, <voice and tone>]: "<the exact words>". The speaker is the subject's label ([@Element2, low gravelly voice]) or, for anyone without one, the noun phrase the prompt uses for them ([The ferry captain, bright carrying voice]). Keep each speaker's tag identical every time, and describe each speaker's voice in different words. One tag per line, in spoken order; when a different speaker answers, start their action with "Immediately," so the voices don't merge. Keep every word; a whisper stays a whisper. - No exact words: no tag, no quoted speech, and no speaking verb for anyone (
says,talks,explains,asks,mouths); write the visible gesture. End the prompt with every person in frame keeping their lips closed (animals and objects need no such line). Smiles and laughs closed-mouthed.
The sound line
One or two short sentences naming actual sounds: ambience, the action's sounds,
footsteps, breathing, fabric. No music unless asked. Never a label, a mood or something
only visible. No human voice, and these words are banned even as metaphors: voice,
vocal, speech, words, talking, chatter, murmur, conversation, whisper,
dialogue. Fabric rustles, a river babbles, a crowd shuffles.
Check it
- Every element's label in the prompt, none in the sound line, none unknown.
- No
@handle or name outside a quoted line. - Lips-closed sentence when there are no words; tags when there are.
- Under 130 words, two to five sentences.
