Intermediate10 min
Directing Runway and Kling Like a DP
A beginner writes what is in the shot. A practitioner writes how the camera sees it. Once you treat the prompt like instructions to a director of photography, your hit rate jumps and you stop relying on luck.
The five-part prompt structure
Strong prompts answer five questions in order: subject, action, camera, lighting, and style. Generators weight the front of the prompt more heavily, so lead with what must be right.
prompt-template.txt
[subject] [doing action],
[camera angle] + [camera movement],
[lighting], [time of day],
[lens / depth of field],
[overall style and grade]
Example:
a lone hiker cresting a ridge,
low-angle tracking shot following from behind,
golden hour backlight, long shadows,
35mm shallow depth of field,
muted cinematic teal-and-orange gradeNegative space and what to leave out
Over-stuffed prompts confuse the model. If you name six adjectives for lighting, you get mush. Pick the two that matter. The model fills sensible defaults for the rest, and defaults are usually fine.
Lock the seed when you are close
When a generation is almost right, reuse its seed and change one phrase. This gives you a controlled edit instead of a brand-new roll of the dice.
Camera moves the models actually understand
| Move | Phrase that works | Use it for |
|---|---|---|
| Push in | slow dolly push-in | Building focus on a subject |
| Orbit | camera orbits around subject | Revealing a product in 3D |
| Tracking | tracking shot following the subject | Motion and energy |
| Crane | crane up revealing the scene | Establishing a location |
Runway - same scene, two grades
seed 48213 grade: teal-and-orange -> keeper
seed 48213 grade: warm vintage 16mm -> alt
diff: only the final style clause changed
Result: two clips of the same scene that differ only in grade, generated on purpose rather than by accident. That is direction, not luck.