How to Write AI Video Prompts: From Timeline Storyboards to Model Dialects
Why timeline storyboarding is the most valuable skill right now
Three months ago people were still arguing with me that writing AI video prompts is just piling up natural-language descriptions of what you want, and the model's smart enough to figure the rest out on its own. Anyone still thinking that today is watching their clips fall apart by the 5-second mark, with the subject drifting the moment there's more than one shot involved.
The real turning point was when timeline-based storyboarding showed up. Instead of one big blob of a prompt, you cut it into time segments — 00:00-00:05, 00:05-00:10 — and each segment only describes the action, camera movement, and composition happening in that window. The jump in execution accuracy is visible to the naked eye. The reason is straightforward: video generation models are essentially conditioning a continuous temporal signal, and the finer-grained and more time-anchored your conditioning is, the less likely the model is to wander off-topic.
This isn't something anyone made up out of thin air — it's been verified repeatedly against Seedance 2.5's own official examples and real-world testing from a bunch of director-minded creators. Seedance 2.5's grasp of multi-shot scripts is currently top-tier, largely because it parses timeline-structured prompts exceptionally well. Flip that around: if you're still describing 30 seconds of footage in one continuous paragraph, it doesn't matter which model you're using — you're basically gambling.
How to actually format the timeline
A usable timeline prompt looks like this:
- [00:00-00:03] Core description of shot / subject / action / environment
- [00:03-00:06] Change in camera movement, or the next beat of the subject's action
- [00:06-00:08] Closing shot, usually a close-up emphasizing one detail
The point isn't making it look pretty — it's that each segment only carries one action intent. A lot of people try to cram a slow push-in, a character turning around, and a lighting change all into one segment, and the model ends up picking one or two to execute while dropping the rest. Better to split into more segments than let a single one carry more than one motion instruction.
Segment length matters too: for a clip around 5 seconds, cut a new time block every 2-3 seconds; for anything over 10 seconds, you can cut every 3-4 seconds. Cutting too finely (a new segment every single second) actually makes the transitions between segments feel stiff — you can visibly see the seams.
Shot-language vocabulary cheat sheet
Chinese cinematography terms you can drop straight into your prompts.
Shot size: wide shot / full shot / medium shot / close-up / extreme close-up — tightening in that order from environment down to facial detail. Spelling this out cuts down massively on subject-proportion issues going haywire.
Camera movement: push in (dolly in) / pull out (dolly out) / pan / track / crane / handheld follow / orbit. Domestic (Chinese) models generally respond more accurately to Chinese camera-movement terms than to stiffly translated English ones.
Lighting: backlit silhouette, side-light rim light, oppressive top light, diffused soft light, Tyndall effect (god rays) — lighting terms are usually ten times more effective than vague adjectives like "cinematic."
Pacing: slow motion / normal speed / quick cuts — it helps to directly note a speed multiplier or frame-rate leaning, like "0.5x slow-motion close-up."
Combine these four categories — shot size + camera movement + lighting + pacing — and you'll cover about 90% of standard shot requirements.
Negative prompts aren't a magic fix, but you can't skip them
Negative prompts get both overrated and underrated by different crowds. The overrated camp expects them to fix compositional mistakes — negative prompts are basically useless against composition drift, that's a job for your positive description to solve. The underrated camp skips them entirely and ends up with the model freely improvising a pile of elements nobody asked for.
Where negative prompts genuinely earn their keep is excluding known model quirks, like:
- Excluding extra limbs, extra people (the classic headache in multi-subject scenes)
- Excluding built-in watermarks/captions (some models add these even when you didn't ask)
- Excluding overly smooth, plastic-looking skin
- Excluding excessive camera shake (if what you want is a stable shot)
Different models support negative prompts to different degrees — Sora 2, for instance, doesn't have a public API right now, so how well negative prompts work depends more on the in-app interface settings than on stacking plain text. Best to check the specific model's input spec before writing, rather than slapping one negative-prompt template onto every platform.
When structured JSON/YAML prompts are actually worth using
Structured prompts — writing shot, subject, action, and camera parameters into separate JSON or YAML fields — earn their keep in two scenarios: one, workflows that need bulk production with prompts auto-assembled by a script, where structured fields make programmatic parsing and substitution easy; two, when you need precise control over several independent parameters at once (say, aspect ratio, shot duration, and camera-movement type) and the target model's API explicitly supports structured input.
But for a single one-off creation — manually writing a prompt for one 10-second clip — wrapping it in JSON is a pure waste of time. The model still has to convert those structured fields back into natural language internally anyway, so just writing natural language directly is more efficient and leaves room for fine-tuning the phrasing by feel. What actually determines the outcome was never whether the format is JSON or plain text — it's whether you broke the shot down properly.
Simple rule of thumb: if the prompt needs to be programmatically generated or reused (like batch-testing the same storyboard with different subjects), go structured. If it's a one-time polish job, plain language is enough.
Model dialects: Seedance eats Chinese storyboards, Veo eats dialogue, Sora eats physics
This is the part that's easiest to overlook but actually caps how good your final output can be: the exact same prompt approach can completely derail on a different model, because each model's training data leans differently, forming its own dialect.
Seedance 2.5 / Seedance 2.0: strongest at understanding Chinese timeline storyboards. For wuxia, guofeng (Chinese-aesthetic), and multi-shot narrative material, writing detailed storyboards directly in Chinese gets more accurate execution than translating to English first.
Veo 3 / Veo 3.1: its core strength is native audio, so the prompt needs to spell out the dialogue content, tone, and ambient sound clearly — just stacking camera-movement descriptions wastes its strongest feature. The writing style here is closer to a script with dialogue than a shot list.
Sora 2: physical simulation is its strength — describing an object's weight, material, collision behavior, and fluid dynamics works better than describing camera language. It's far more responsive to details like "how will this ball bounce" than to cinematic-shot phrasing.
Kling 3.0: highly parameterized camera control, good for directly specifying the exact magnitude and speed of push/pull/pan/track moves — action and fight-scene content especially benefits from this.
Take the same storyboard script, write it as a Chinese shot list for Seedance, rewrite it as a dialogue-driven scene script for Veo, and swap it for physical-action descriptions for Sora — the quality gap in the output will be very obvious.
Three prompt examples you can copy straight away
Example 1 (Chinese storyboard, for Seedance-series):
[00:00-00:03] 中景,一名身着靛蓝长衫的女子在竹林中缓步前行,侧光透过竹叶形成斑驳光斑,风吹动衣袂与竹叶同步摆动
[00:03-00:06] 镜头缓慢推进至近景,女子转身回眸,发丝随风扬起,背景虚化加深
[00:06-00:08] 特写定格于女子眼神,光线由侧光转为逆光剪影,收尾
负面提示词:多余人物、画面水印、过度平滑皮肤
Example 2 (dialogue script, for Veo-series):
场景:清晨厨房,窗外阳光斜射进来。一位中年男子一边煎蛋一边回头对镜头外的孩子说:今天不许迟到,校车七点半就走。语气带点无奈的笑意,背景有煎蛋滋滋声与远处电视新闻的模糊人声。镜头固定机位,轻微手持晃动感。
Example 3 (physical action, for Sora-series):
一颗玻璃弹珠从一米高的木质斜坡顶端滚落,途中撞击三个多米诺骨牌依次倒下,最后一块骨牌推动一个装满水的塑料杯翻倒,水花四溅到桌面反光,全程一镜到底,固定俯视机位,正常速。
Copy these three as-is and just swap in your own subject and setting — the skeleton doesn't need to change.