Generating Video with Kling

I'll say this for AI video generation — it's the one area where I think most people genuinely underestimate how much the prompt matters. More so than images, even.

That is a generous framing of what is, in practice, people typing 'cool cinematic video of a car' and then being surprised when they get something that looks like a fever dream rendered in margarine.

Which is exactly the point. The tool isn't broken. The instruction was.

Precisely. Kling is doing exactly what it was asked to do. It's doing it faithfully, enthusiastically, and in the complete absence of the information it actually needed. You have to describe the scene, the objects, the style, the mood — if you wanted a director, you had to behave like a screenwriter.

And here's the thing people find uncomfortable: your first output probably won't be good. That's not a flaw in the process. That is the process.

I realise that sounds like I'm asking people to tolerate failure. I am. Specifically, I'm asking them to tolerate it, study it, and use it. The ones who find that discouraging are, statistically, not the ones who end up very good at anything.

Though I'd push back slightly on the idea that iteration is some grand discipline. Sometimes the first prompt is bad because it was vague, you tighten it, and the second one is fine. It doesn't have to be a spiritual journey.

...That is a fair point, actually.

I know. I made it on purpose.

The broader issue is that people treat Kling as a finished-video machine. It is not. It generates clips — short ones, with their own inconsistencies around things like hands and physics and objects that occasionally decide to become other objects mid-shot. Treating the output as raw material rather than a final product is not a workaround. It's the correct mental model.

Which is where integrating it into an editing workflow actually matters. A five-second Kling clip that you grade, trim, and cut against other footage is genuinely useful. A five-second Kling clip that you post directly because you couldn't be bothered to finish the job is a different proposition entirely.

You had one job in this transaction — to make the thing watchable. The AI gave you a starting point. What you do with it is still your problem.

I think the honest version of the lesson is: this is a skill. It takes a few attempts to develop an intuition for what the model responds to. But it's a learnable skill, and it compounds — the more you use it, the faster the iteration gets.

Yes. And the people who grasp that — that 'generating video' and 'prompting well' are not the same thing — are the ones who will find this genuinely useful rather than merely intermittently impressive.