Write prompts that describe change, not just a scene
An image prompt describes a frozen moment; a video prompt needs a verb. 'A lighthouse in a storm' gives the model nothing to animate — 'waves slamming the lighthouse as the beam sweeps through rain' does. Structure prompts as subject, action, camera, light: 'a fisherman hauling a net over the gunwale, spray in the air, slow dolly in, low golden light.'
Keep it to one action per clip. Generations run 4-12 seconds, which is one story beat — a door opening, a turn toward camera, a wave breaking. Prompts that stack three events into one clip usually get a muddled blend of all three; three clean clips cut together always read better.
Match the model to the job
Kling 3 is the flagship for text-to-video: strong prompt adherence, end-frame control, and multi-shot support. Veo 3.1 renders the most premium realism per frame when the budget allows, with Veo 3.1 Lite covering a similar look for less. Sora 2 leans creative and stylized — the pick when you want animation energy rather than documentary realism.
For drafting, Seedance Pro is fast and cheap enough to iterate wording freely, and standard Kling covers everyday text-to-video work at a mid-range credit cost. A practical loop: block the idea on Seedance Pro, refine the prompt until the motion reads right, then rerun the final wording on Kling 3 or Veo 3.1.
Get camera movement without gambling
The fastest way to make AI video look intentional is a deliberate camera. Nidhogg ships camera-move presets — dolly-in, orbit-right, crash-zoom, drone-pullback, steadicam-glide, and dozens more — that encode the move directly, so you don't have to describe it in words and hope the model agrees.
Pair the move with the story beat: a slow dolly-in builds tension on a face, a crane-up turns a moment into an establishing shot, a crash-zoom sells a reveal. Use one preset per clip; combining contradictory moves in a single prompt is the most common reason motion comes out mushy.
From one clip to a sequence
Single generations top out around 12 seconds — that is true of every serious video model right now, and Nidhogg won't pretend otherwise. For longer stories, Seedance 2 and Kling 3 support multi-shot generation: one prompt returns several connected shots with consistent characters and lighting, like a mini scene already covered by an editor.
For full pieces, write your script as a shot list, generate each beat as its own clip with a consistent style, and assemble the cut in your editor. The script-to-video and long-video guides linked below break that workflow down step by step.