How to Write AI Video Prompts That Actually Work (2026 Guide)

Most disappointing AI video output is a prompting problem, not a model problem. The same model that produces a muddy, drifting clip from one prompt produces a clean cinematic shot from another. This guide covers what actually changes the result.

Describe a shot, not a scene

The single most common mistake is writing a story instead of a shot.

Weak:

A woman walks into a coffee shop, orders a latte, sits down, and starts reading a book.

That is four shots. A model asked to produce all of it in 6 seconds will either rush every beat or drift between them incoherently.

Strong:

Medium shot of a woman in a wool coat sitting by a rain-streaked coffee shop window, steam rising from a latte, warm interior light, slow push in.

One subject, one action, one camera move. Video models generate a continuous take — give them a take.

Use real camera language

Models are trained on footage described in production vocabulary, so that vocabulary works:

• Shot size: wide, medium, close-up, extreme close-up
• Camera move: static, slow push in, pull back, pan left, tracking shot, handheld, aerial
• Lens feel: shallow depth of field, anamorphic, macro
• Light: golden hour, overcast, hard studio light, neon practicals, backlit

Compare:

A car in a city at night.

against:

Low tracking shot alongside a black sedan on wet asphalt, neon signage reflecting in the paint, shallow depth of field, night.

Same subject. The second one specifies what the camera does and what the light does, which is most of what makes a shot look intentional.

Put the important thing first

Prompt weight tends to fall off toward the end. Lead with your subject and action; leave style descriptors for the tail.

Subject and action → camera → lighting → style

A grey cat stretching on a windowsill, static close-up, soft morning light, shallow depth of field, natural film grain.

Motion needs to be plausible

Ask for physically coherent motion and you will get it more often. Ask for something the model has to invent and it will produce artifacts.

Reliable: walking, drifting smoke, falling rain, rippling water, cloth in wind, slow camera moves.

Unreliable: complex hand manipulation, precise text appearing on screen, fast multi-limb action, exact object counts.

If you need on-screen text, add it in post. No current model renders arbitrary text reliably.

Choose the workflow before the prompt

Three workflows exist and each takes a different prompt:

Text to video — the prompt carries everything. Be specific about subject, setting, and camera.

Image to video — the image already establishes subject, composition, and palette. Describe motion only. Re-describing what is visible in the image fights the reference:

Slow push in, subject turns head slightly toward the light, subtle fabric movement.

First/last frame — you supply both endpoints and the model interpolates. Describe the transition, not the frames:

Smooth camera drift from the wide establishing view to the close detail, continuous motion, no cuts.

Not every model supports every workflow. Kling O3 has no text-to-video path; Happy Horse 1.1 is text-only; Wan 3 has no first/last-frame mode. Check before you write.

Negative space beats negative prompts

Rather than listing what you do not want, describe what should be there. "No blur" is weaker than "sharp focus on the subject's eyes." Positive specification steers generation; negation often does not.

Iterate cheaply

Credit costs differ substantially — 20 credits for Seedance 2.0 or Wan 2.7 against 35 for Veo 3.1. Iterating on an expensive model is the fastest way to waste a balance.

A workable loop:

1. Draft the prompt on a cheap model until composition and motion are right.
2. Change one variable at a time. Changing three things at once tells you nothing about which one helped.
3. Once the prompt is stable, re-render on your target model.

Prompts transfer between models well enough that this works. The composition you dial in on Seedance 2.0 will usually hold on Veo 3.1.

Keep the ones that work

Prompts that produce good results are reusable assets. Save them with a note about which model and workflow produced the result — the same text can behave differently across models, and next month you will not remember which combination worked.

A template

[Shot size] of [subject] [specific action],
[setting detail], [lighting], [camera move],
[style / finish]

Filled in:

Medium close-up of a ceramicist shaping a bowl on a spinning wheel, clay-dusted hands, workshop window light from the left, static camera with shallow depth of field, documentary finish.

Start there, change one thing per attempt, and keep what works.