Introduction

In the previous article, I shared how to use Doubao to generate AI videos. Many readers enthusiastically tried generating their own AI videos, but quite a few discussed a common issue: why does my output always have a "plastic" feel? Some attribute it to insufficient resolution, others to characters that don't look right. In most cases, the root cause comes down to one key factor: the quality of the input prompt. People often tell the model only who to film, but not how to film. A prompt like "a girl crying in the rain" carries almost no useful information: where she is crying, what the light is like, how far the camera is, the intensity of the crying—none of this is described, so the model has to guess.

This article details the principles for writing video-generation prompts: what components they consist of, how to write each part, and common pitfalls. Nothing more to wait for—let's dive in together~

1. A Prompt Is a Director's Instruction Pack, Not Just a Description

Many people still think of a prompt as simply "sending a sentence to a video-generation model." But a prompt that actually produces quality footage is a structured set of creative instructions that communicate subject, scene, lighting, camera, composition, style, texture, and atmosphere in one go. Basic prompts rely on imagination; professional prompts rely on industry standards, recreating the filmmaking pipeline in text—condensing location scouting, lighting setup, camera positioning, composition, color grading, and more into a single paragraph.

Compare the two approaches below:

A beginner writes "an ancient-style woman standing under a peach blossom tree"—ten runs produce ten different results.

A professional prompt instead reads: "Orbiting push-in, an ancient-style woman in white stands still in a ten-mile peach forest, petals flying, warm side-backlight at golden hour, Tyndall beams piercing the branches, Chinese-style negative space composition, cinematic aspect ratio, ink-wash realism, film grain texture, cold and lonely atmosphere." The difference isn't word count; the latter gives the model everything at once: scene, lighting, camera, style, atmosphere.

2. Essential Elements of a Video Prompt

A competent prompt needs at least eight elements: subject, scene, action, camera, style, lighting, emotion, atmosphere.

Each of these is indispensable: without a subject, the model doesn't know what to generate; without a scene, characters float like stickers in a void; without action, the frame feels like a PPT slide; without camera, the viewer lacks a guided structure; without style, the visual tone is chaotic; without lighting, the image becomes flat and gray; without emotion, the performance appears hollow and fake; without atmosphere, the audience's immersion is weak. Below is a typical example to show how a prompt with all essentials should be written.

First, write the subject:

An ancient swordsman with silver hair and a white robe

Second, write the scene, preferably with coordinates:

A Jiangnan water town in the early morning after rain, wet bluestone streets

Third, write the action—make it specific:

Sword dancing in a bamboo forest, crisp and sharp movements

Fourth, give the camera angle:

Low-angle upward shot, slow push-in

Fifth, add lighting to give texture:

Backlight at dusk, rim light outlining the hair strands

Unfold the remaining items—style, emotion, atmosphere—one by one. Follow these steps and you can produce a very complete prompt.

3. A General Prompt Framework: SCELA+

The eight elements above are still a bit hard to memorize, so the Jimeng team distilled them into a framework called SCELA+. It splits the prompt into five modules: Scene, Character, Expression/action, Lens language, Atmosphere/emotion; the trailing plus sign represents extra additions like lighting, special effects, and reference materials. In fact, the alphabetical order of SCELA+ itself suggests the order to follow: first build the stage, then place the actor, direct the performance, set the camera, and finally inject the soul. That's the order a film crew works in, and also the order the model finds easiest to understand.

Here's what each letter means:

1. S is Scene: use three to five concrete elements to establish a spatial-temporal anchor—location + era + weather + key props + spatial layers.
2. C is Character: lock recognizability with "feature anchors"—one core feature, several supporting features, plus clothing material.
3. E is Expression/action: break actions into beginning, development, turning point, and closure; break expressions into eyes, mouth corners, and micro-expressions.
4. L is Lens language: follow the formula "shot size + movement + angle + focus."
5. A is Atmosphere/emotion: use synesthesia to engage multiple senses.
6. The plus sign is supplementary: for light details, effect parameters, or referencing materials with @.

Write your prompt in that order; concatenate the pieces into a complete video prompt, like this example:

(S) An abandoned 90s amusement park at dusk, a rusted carousel, puddles reflecting the dying sun. (C) A silver-haired girl, a teardrop mole under her right eye, faded denim jacket, barefoot. (E) Bites her lower lip, eyes reddening yet holding back tears, slowly walks away from the carousel. (L) Wide top-down shot, slowly pushing to medium shot, focus locked on trembling fingers. (A) Damp smell of rust, distant machinery creaking. (+) Dusk backlight outlines her figure, @image1 references the position of the mole.

The generated result looks like this:

4. Key Techniques for Improving Visual Quality

Beyond the SCELA+ framework, there are various tricks to make your output even better. Using them well can make your visuals more vivid and lifelike.

4.1 Camera Movement: Don't Just Say How It Moves, Say How Much

Many people only write "move the camera a bit," a vague instruction that forces the model to guess, leading to either no movement or erratic motion.

Camera movement can be described from two angles:

The first is the type of movement: push, pull, pan, truck, follow, and crane. Push is moving in toward the subject, pull is moving away, pan is rotating the camera body while the lens moves left/right/up/down, truck is moving parallel, follow is tracking the subject, crane is vertical rise or drop. Advanced moves include orbit, dive/rise, zoom, whip pan, and slight shake—corresponding to circling, diving, pseudo-movement, fast cuts, and handheld tension, respectively.

The second angle is parameters: is the speed constant or variable, is the range "30-degree pan" or "half-circle orbit," does it start from "slowly lifting from the character's feet to a facial close-up" or "pushing evenly from wide to medium shot"?

So replace vague phrases like "slowly push" with executable instructions like "push evenly, complete in 3 seconds," which yields steadier camera moves.

4.2 Lighting: Five Dimensions to Lock Down Visual Texture

Lighting determines not just brightness, but also three-dimensionality, emotional tone, and visual texture. Professional lighting descriptions should cover five dimensions.

First, specify the position of the key light, which sets the light-dark structure. The common five are top, side, back, bottom, and front. For example, writing "key light from 45-degree left, classic Rembrandt triangle on the face" instantly establishes the overall contrast and tone.

Second, define the quality of the light. Hard light comes from direct sunlight or spotlights, with sharp shadow edges, good for emphasizing contours and strength. Soft light comes from overcast skies or softboxes, with gradual transitions, ideal for portraits. State it plainly in the prompt: "hard light with sharp edges" or "soft light diffused with natural transitions."

Then come layers and color temperature. Light can be layered into three: the key light sets the base, fill light softens shadows, rim light separates subject from background. Color temperature generally conveys emotion: warm light (3000K) feels nostalgic and cozy, cool light (6500K) feels distant, and teal-orange contrast carries cinematic quality by default.

The most striking effects come from light-shadow details: are shadows soft or hard, is there a medium in the light (Tyndall effect, beams through fog, visible dust), what is the surface response (metallic reflections, subsurface scattering on skin). An excellent prompt for light-shadow might read: "the setting sun casts striped shadows through Venetian blinds, dust dancing in the light beam." Adding that single sentence instantly gives the image three-dimensional air.

4.3 Emotion: Translate Abstract Feelings into Body Signals

Emotion is very hard to write, yet it largely determines character realism and viewer immersion, making it a decisive factor in video generation.

For video-generation models, abstract emotions like "sadness" are not understood; but phrasing like "red eyes, tears welling up, trembling mouth corners, Adam's apple rolling" is something the model can clearly generate. This is the "externalization translation method" for writing emotion: translate abstract feelings into visible bodily signals.

4.4 Don't Write "Don't"

A counterintuitive rule in prompt writing for video models: avoid using "don't" as much as possible. ByteDance's Seedance model has trouble understanding negated instructions; the more you emphasize what you don't want, the more likely it appears. So rephrase.

Change "don't have finger errors" to "fingers naturally bent, five clearly separated"; "don't have background flicker" to "scene stays stable throughout, no flicker"; "no extra people" to "only one person on screen, no additional characters."

4.5 Numeric Control: Turn Vague Estimates into Concrete Parameters

The techniques above address the "realism" of the image; numeric control addresses the "accuracy" of it.

The most useful application is the time axis. Split a 15-second clip into segments, label them with [03s], [46s], and assign one core action per segment.

Another application is specifying the intensity of detail, e.g., "motion intensity 7/10," which constrains the range while providing a concrete strength for the current shot, helping the model generate more precisely.

For complex physical effects, you can declare parameters directly, e.g., "gravity coefficient = 0.8" or "cloth flexibility = high." Numbers aren't magic—they tell the model "to what degree."

4.6 Multi-Shot

Currently, Doubao's Seedance model supports up to 15 seconds of video. To give such a short duration narrative coherence, recommend multi-shot storytelling rather than one continuous shot.

For time-segmented prompts, design three to five segments following "beginning, development, turning point, conclusion": open with establishing environment, middle uses medium shots to carry the subject, then close-ups for emotion, and finally pull back to close. Let the model handle the light transitions and narrative joins between segments.

Advanced approaches include gradual shot-size progression, letting "wide to medium to close-up" form a visual path; or incorporating terminology, e.g., "apply a Hitchcock zoom during the panic moment, then an orbiting shot to show spatial relationships."

4.7 Prompt Compression

After all these frameworks and techniques, you might ask: if I write every detail flawlessly, will the result necessarily be good?

Not necessarily. For Seedance, the optimal prompt length is 30–200 words; beyond that, quality noticeably drops because the model's attention gets diluted. In testing, when redundancy exceeds 38%, the model's accuracy in following action instructions drops significantly.

Compression doesn't mean mechanically deleting characters; it means "keep the action skeleton, trim the decorative fat." Preserve the subject's key features and core actions. For parts that require extensive face description, you can use "@image1 reference face" to upload a reference image, maintaining face consistency while compressing content.

An example:

Original verbose prompt:

A young man in ornate brocade robes, a warm jade pendant at his waist, face like jade, eyes deep and bright, standing in a Jiangnan courtyard amid falling petals, surrounded by odd rocks, a turquoise pool, meandering corridors and waterside pavilions, koi fish playing in the pool, a drizzly mist in the sky, holding an oil-paper umbrella, a faint melancholy in his brows as if missing a distant friend, the camera slowly pulls from wide to a final close-up on his desolate eyes, background music is an ancient guqin piece, overall atmosphere is mournful and beautiful, movie-grade, 8K resolution.

Compressed prompt:

Ancient-style man holding an oil-paper umbrella standing in a rainy Jiangnan garden, sad eyes. [0~~4s] wide shot push-in, [5~~8s] close-up of the eyes. Fine drizzle, guqin BGM, cinematic texture. @image1 reference face and clothing.

Generated output:

5. Video Prompt Skill: Automating the Process Above

At this point some readers will say: "I'm not a professional—there's too much knowledge here, how can I remember it all?" As a model blogger who cares about fans, my goal is to give everyone the ability to harness AI models, so I've packaged the entire workflow into a Skill. (If you're unfamiliar with Skills, check out my article Agent Skills Complete Guide: Core Concepts | Design Patterns | Practical Code.) Whether you use workbuddy, codex, claude code, or deepseek harness, you can install my Skill. Simply input the scene you want, and the Skill helps generate a complete set of professional prompts. Below I demonstrate using DeepSeek Harness. If you're unfamiliar with DeepSeek Harness, see my column: DeepSeek Harness from Beginner to Advanced

1. Set up the workspace, put the Skill inside, and let DeepSeek Harness install it. I've already installed it here.

Add the video-prompt-director skill to your skill library

1. Input the scene you want to generate; the prompt-assisted agent will ask for suggestions and required parameters.

1. Follow the interactive prompts to input necessary parameters until you get a satisfying prompt.

1. Input the prompt into Doubao to generate the video.

1. The final result looks like this—simple, right?

That's all for this article. If you're interested in the video-prompt-director skill, you can add my WeChat public account 大模型真好玩 and send the message 视频生成skill to get it.

6. Summary

After this full tour, my takeaway: the prerequisite for a video model to "obey" you is that you first learn to speak like a director. From the five modules of SCELA+ to camera movement, lighting, externalizing emotion, avoiding "don't," numeric control, multi-shot narratives, and prompt compression, this method essentially uses structured language to replicate the decision-making process of film production, making the model execute instructions instead of guess.

On the tool side, Doubao's Seedance, Jimeng, and Kling each have strengths; the real gap is always the degree of prompt structuring. Structured prompts are not just a description—they are the most core technical ability in AI video creation. I will continue refining the video-prompt-director skill to make the experience from "input a scene" to "get a full set of professional prompts" even smoother, and I'll share updates with everyone as they come~ Of course, my main writing focus remains in-depth research and exploration of frontier LLM technologies. Welcome interested readers of LLMs and agent tech to follow my Juejin account and column, or my same-named WeChat public account 大模型真好玩, to have fun understanding LLMs together~