DEV Community

Maya Chen
Maya Chen

Posted on

Treat AI video prompts like a spec, not a wish: a template I use for multi-shot clips

Most bad AI video I have made came from a prompt that read like a mood board: "cinematic, moody, beautiful woman walking in rain, 4K". The model fills every gap with its own defaults, and the defaults change from shot to shot. The fix that worked for me was boring: write the prompt the way you would write a ticket for another engineer.

The five fields

I keep every shot in a tiny block with five fields, always in the same order:

SUBJECT:   who/what, with 2-3 fixed identifiers (hair, jacket color, prop)
ACTION:    one verb phrase, one motion only
CAMERA:    lens feel + one move (slow push-in, static, orbit left)
LIGHT:     source + time of day
CONSTRAINT: what must NOT change (face, logo, outfit)
Enter fullscreen mode Exit fullscreen mode

The CONSTRAINT line is the one people skip, and it is the one that saves the most regenerations. If the jacket is red in shot 1, say "jacket stays red, no pattern" in every shot.

One motion per shot

Asking for "she turns, smiles, picks up the cup and walks to the window" in five seconds is asking for rubber limbs. Split it. Two short clips with one clean motion each cut together better than one clip with four half-finished motions.

Lock the ends, let the middle move

If your generator supports first-frame and last-frame input, use it for anything with a product or a logo. You control the two states that matter (closed box, open box) and the model only has to invent the transition. It is the closest thing to a keyframe you get.

Test harness, not vibes

I run each spec three times before judging it. If one out of three is good, the prompt is not good yet; it got lucky. I only change one field between runs, like a normal A/B test.

Where I run this

Lately I have been testing these specs on Kling4Pro, an independent web front end (not the official Kuaishou site) that exposes Kling 4.0 Pro alongside older Kling versions, Veo 3.1 and Seedance. The useful part for this workflow is that it takes first/last frames and up to nine reference images, and new accounts get some free credits, so you can run the three-run test before spending anything. Whatever tool you use, the spec format is the part that transfers.

If you try the template, the one habit I would keep is the constraint line. It turns "why did the face change" from a mystery into a bug report.

Top comments (1)

Some comments may only be visible to logged-in visitors. Sign in to view all comments.