A week into working with Kling 3.0 on real project prompts rather than just test cases, I've got enough of a pattern built up to share a working reference for camera movement language that actually produces consistent results. This isn't exhaustive. It's what's held up across enough repeated use that I trust it.
What consistently works
Describing camera movement as a physical action with a clear direction and pace, "the camera pushes in slowly," "a slow pan left," "handheld, following at walking pace," produces reliable results far more often than abstract mood language like "dynamic camera work," which the model tends to interpret inconsistently from generation to generation. Specificity beats evocative language every time in my testing, even though the evocative version reads better as a sentence.
Combining camera and subject motion in one clean instruction
Structuring a prompt as camera behavior first, then subject behavior, then a brief environmental note, in that order, produces more reliable results than blending everything into one dense descriptive paragraph. The model seems to weight earlier information more heavily, so leading with the camera instruction specifically has mattered more than I expected going in.
- "The camera pushes in slowly" beats "dynamic, cinematic camera work"
- Order matters: camera behavior, then subject motion, then environment
- Name a pace explicitly, slow, steady, quick, rather than assuming a default
What still requires trial and error
Complex combined moves, a push in while also panning, still produce inconsistent results often enough that I generate two or three variations rather than trusting a single pass for anything with compound camera motion. Simple, single-action camera moves are reliable enough now that I'll commit to one generation and move on. This is a living reference, not a final one, and I expect to be updating it again within a couple of months as I put more real project hours against it.