Most people prompt an AI video model the way they'd type a search query — a few nouns and a hope. Then they blame the model when the result is stiff, generic, or ignores half of what they asked. The models are better than that. The gap is almost always in the prompt. Here are the techniques that actually move the output, drawn from how these models genuinely behave.

“Write a shot, not a search query.

"woman walking city" is a search query. "A woman in a red coat walks briskly across a rain-slicked crosswalk at night, neon signs reflecting in the puddles" is a shot. The model has far more to work with in the second, and almost none of it is decoration — every clause is a decision it would otherwise make randomly.

It is tempting to stack comma-separated tags: "cinematic, 8k, dramatic, moody, film grain." Current models read natural language better than tag soup. A plain descriptive sentence outperforms a pile of adjectives, because the model can parse relationships between words rather than guessing which of twelve tags matters most.

The clearest prompts describe three things in order: what is in the shot, how it moves, and how the camera moves. Blur these together and the model has to disentangle them. Keep them distinct — subject, then subject motion, then camera language — and adherence jumps.

This is the single biggest fix for image-to-video. When you start from an image, the model can already see the picture — describing what is in the frame is wasted words. Spend the prompt on what happens: "the camera slowly pushes in as she turns her head toward the window." Describe the motion, not the photograph.

"The camera pans left, then cranes up, then zooms in" asks for three moves in a few seconds and usually produces a confused blur. Keep it to a single, clear camera move per generation. If you need a sequence, generate the beats separately and cut them together — you will get cleaner motion and more control.

Models are trained on film, and they respond to film vocabulary: *dolly in*, *tracking shot*, *low angle*, *shallow depth of field*, *handheld*. These are not magic words, but they are precise instructions the model actually understands, where "make it look cinematic" is not.

Lighting is where "fine" becomes "striking." Say where the light comes from and what it does: *golden hour backlight*, *hard overhead light*, *soft window light from the left*. Add the atmosphere — *fog*, *dust in the air*, *volumetric light* — and the frame gains depth the model will not add on its own.

Models have real, distinct strengths, and the same prompt does not behave identically across them. This is worth learning per model rather than assuming:

• For Kling, see the complete guide to prompting Kling.
• For LTX-2, the LTX-2 prompting guide covers its start/end-frame control.
• For Seedance 2.0, the Seedance 2.0 guide covers its native-audio and reference behaviour.
• For voice and audio, the ElevenLabs voice & audio guide is the equivalent.

Several models let you set the opening and closing frame of a clip. This is the most underused control there is. Fixing where a shot begins and ends is the practical difference between clips that cut together and clips that drift — it turns generation from a slot machine into direction.

When a result is close but wrong, resist rewriting the whole prompt. Change one thing — the camera move, or the light, or the subject's action — and regenerate. You learn what the model is actually responding to, and you converge on the shot instead of thrashing between unrelated attempts.

If there is a single principle underneath all ten, it is this:

✅ Specific and singular — one clear subject, one motion, one camera move, described in a plain sentence, wins almost every time.

⚠️ Vague but hopeful — "cinematic, epic, beautiful" gives the model nothing to act on, so it falls back on the average of its training data. That average is what "generic AI video" looks like.

❌ Overloaded — three camera moves, five adjectives, and two subjects in one prompt asks the model to resolve conflicts you could have resolved yourself.

Prompting is a skill you build by watching what changes when you change the prompt — which is fast and cheap when you can pick the right model per shot and see the result in seconds. Describe one specific shot, generate it, then change exactly one thing and generate again. Two iterations in, you will prompt better than most people ever bother to.