Does Kling motion control follow my prompt? Only for appearance

No: on Sume's Kling 3.0 Motion Control, the optional prompt does not direct the movement. The API reference (read 2026-10-09) says that motion and framing follow the driving video and that the prompt only steers appearance details. The field accepts 1 to 2,000 characters. If you want the avatar to move differently, change the driving video, not the prompt.

What each input controls

Motion control separates three jobs. The still decides who appears. The driving video decides how they move and how long the clip runs. The prompt nudges how the result looks. Two flags handle the rest.

What a useful prompt looks like

Because the model already has the motion, spend prompt words on what the video cannot tell it: the outfit's color, the lighting you want to match, the material of a prop. Keep it short and concrete. A prompt such as 'matte red jacket, soft studio light' does the job, and a prompt like 'spin twice and wave' asks for motion that the driving video has to supply.

Prompts do not change the price. Cost is $0.1575 per output second on the catalog, billed on ceil of the motion-video second, so a 14-second reference is 14 x 0.1575 = $2.205 whatever the prompt says.

If the motion is wrong

Record or pick a new driving video. Test a short 3-second piece first: 3 x 0.1575 = $0.4725 shows whether the movement transfers to your still before you pay $4.73 for 30 seconds. Choose character_orientation: image when you want the still's framing to hold, and video when the dancer's framing should win. If the driving clip has speech or music you cannot reuse, set keep_original_sound to false.

The prompt is also not a safety valve. If the output is not what you want, rewording it cannot fix a mismatch between the still and the reference movement.

Testing a prompt cheaply

Because prompts do not change motion, you can test a prompt on a short driving clip. Cut 3 seconds from your reference ($0.4725), try two prompts, compare the look, and only then run the full 14-second clip at $2.205. That is a $0.945 test for two prompt variants against the risk of paying $4.41 for two wrong full-length renders.

If you use an avatar instead of a still, remember it resolves to the avatar's identity still, so appearance prompts may have less to steer. Use an image_url when you need to vary the look per run.