TutorialVideo Generation

How to Write AI Video Prompts (With Examples)

July 5, 2026 · Updated September 30, 2026 · 9 min read

A good AI video prompt describes one shot — the subject, what moves, how the camera moves and what the light looks like — in one to three concrete sentences. The rest of the result comes from settings: which model, how many seconds, which aspect ratio, and how you stitch clips together afterward. This guide covers both, with prompts you can paste straight into a generator.

The four-part prompt structure

Every strong video prompt covers four things:

  1. Subject — what is in the shot.
  2. Action / motion — what moves, and how.
  3. Camera — the camera move and framing.
  4. Look — lighting, time of day, style, mood.

Miss any one and the model fills the gap with a guess — usually a boring one.

ByteDance's prompt guide for Seedance 2.0 (updated September 2026) stretches the same idea into a longer formula — subject, action details, scene, lighting and color, camera movement, style, image quality, constraints — and adds rules worth trying on any model:

Camera language that works

Models respond well to explicit camera terms. Useful ones:

Lighting language that works

Lighting sets the entire mood. “Golden hour,” “overcast soft light,” “dramatic low-key lighting,” “neon-lit night,” “backlit silhouette” — these steer the look more than almost anything else.

Example prompts to copy

Each rewrites a weak prompt into a strong one and suggests settings to start from. They're starting points to adapt, not guaranteed results.

Nature:

❌ “a forest”
✅ “Slow aerial drone shot gliding over a misty pine forest at dawn, golden light breaking through the trees, cinematic, calm.”
Start with: 16:9, 5 seconds.

Product:

❌ “a watch”
✅ “Slow 360-degree rotating shot of a luxury watch on a black reflective surface, dramatic side lighting, shallow depth of field, premium and minimal.”
Start with: 1:1, 5 seconds.

Urban:

❌ “a city”
✅ “Tracking shot moving down a rain-soaked city street at night, neon signs reflecting in puddles, shallow depth of field, cinematic, moody.”
Start with: 16:9, or 9:16 for Reels, 5 seconds.

Abstract / background:

❌ “colorful background”
✅ “Slow-moving abstract gradient of deep purple and pink, soft flowing shapes, static camera, smooth and hypnotic, ideal as a background plate, no text.”
Start with: your layout's ratio, 10 seconds for a longer plate.

Food B-roll:

❌ “coffee”
✅ “Macro close-up of espresso pouring into a white ceramic cup, crema swirling, steam rising, fixed camera, warm morning window light, shallow depth of field, no text.”
Start with: 9:16 or 1:1, 5 seconds.

Vertical social:

❌ “a rainy day”
✅ “Low tracking shot alongside a paper boat drifting down a rain-filled gutter, raindrops rippling the water, overcast soft light, playful and calm, no text.”
Start with: 9:16, 5 seconds.

Establishing shot:

❌ “a harbor town”
✅ “Wide establishing shot of a small harbor town at blue hour, lights switching on along the waterfront, slow push-in from over the water, cinematic and quiet, no text.”
Start with: 16:9, 8 seconds — a length Seedance 2.0, Kling 3.0, Veo 3.1 and Happy Horse 1.0 all support.

Short sequence (multi-shot):

✅ “Shot 1: Wide shot of a lighthouse on a cliff at dusk, fixed camera, waves breaking below. Shot 2: Cut to a close-up of the lamp as it switches on, the warm beam sweeping past the lens. Shot 3: The camera slowly pulls back to a wide shot as the beam crosses the dark sea. Cinematic, moody, no text.”
Start with: 16:9, 12–15 seconds, Seedance 2.0 (its guide recommends this format).

From prompt to finished video: the text-to-video workflow

The prompt is step one. Four decisions shape the rest.

Step 1 — Pick the model

Our Video Generator offers five models, and all of them cost the same 10 credits per second, so choose on fit rather than price. What the makers document, as of September 2026:

The quickest way to choose: run one prompt on two models at a short length. Two 5-second clips cost 100 credits.

Step 2 — Pick the length

The duration slider only offers lengths the chosen model renders — 4–15 seconds on Seedance 2.0, 3–15 on Kling 3.0 and Happy Horse 1.0, 4, 6 or 8 on Veo 3.1, 6 or 10 on Grok Video — and you're charged for the seconds you select. Draft at 5 seconds (4 or 6 on Veo, 6 on Grok): enough to judge the motion, at a third of the price of 15. If a model refuses a setting at submission, you'll see an error and no credits are taken. For longer scenes, generate several clips or extend one.

Step 3 — Pick the aspect ratio and resolution

Decide where the video will run before you generate: 16:9 for YouTube and landscape embeds, 9:16 for TikTok, Reels and Shorts, 1:1, 4:3 or 3:4 for feeds. Cropping later costs resolution — a 9:16 crop from a 1080p 16:9 frame is only about 608 pixels wide. ByteDance also notes that Seedance adds stray subtitles less often in landscape than in portrait, so if captions keep appearing in 9:16, add “no subtitles” to the prompt.

720p and 1080p cost the same per second here, so there's no price reason to draft at 720p.

Step 4 — Generate, compare, iterate

The same prompt gives a different clip every run. Generate two or three variations, keep the best, and change one element of the prompt at a time so you learn what each change does. Three 5-second variations cost 150 credits, or $6.75 at Starter Pack prices. Failed jobs are refunded automatically; clips you don't like aren't.

Step 5 — Stitch clips into a finished video

A finished video is almost never a single generation. A workflow that holds up:

Ten 5-second shots come to 500 credits: $22.50 in Starter Pack credits, $16.39 with the Power Pack, or $12.25 of a Pro plan's monthly 2,000 (prices from our pricing page, September 30, 2026). For social, you can often skip the heavy edit: one strong 9:16 clip plus a caption is a complete post.

Common mistakes

Put it into practice

Open our Video Generator, paste one of the example prompts above, and tweak from there. For a closer look at how the models compare, see our best AI video generators roundup.

Try a prompt now →