前言
上篇文章笔者分享了如何使用豆包生成AI视频，不少读者都踊跃尝试生成自己的AI视频，不过也有不少读者朋友讨论：为什么我生成的片子总有一股“塑料味儿”？有人认为是画质不够，有人觉得是人物不像。其实大多数时候，问题都出在一个关键原因上：输入提示词的质量。大家往往只告诉了模型要“拍谁”，却没有细致地告诉它“怎么拍”。“一个女孩在雨里哭”这种提示词几乎不含有效信息：她在哪儿哭、什么光线、镜头离她多远、哭到什么程度，都没有描述清楚，大模型只能靠猜。
本篇文章笔者将详细介绍大模型视频生成的提示词编写原则：它由哪些部分组成、每部分怎么写、常见坑又在哪里。还等什么？大家一起来学习吧~
一、提示词是一份导演指令包，不只是一句描述
很多人对提示词的理解还停留在“给视频生成大模型发一句话描写”。但真正能出片的提示词，是一段结构化的创作指令，用文字把主体、场景、光影、镜头、构图、风格、质感和氛围一次性交代清楚。普通提示词靠的是想象，专业提示词靠的是工业标准，用文字把影视流程复刻一遍，将选景、布光、机位、构图、调色等都压进一段话里。
看下面两组写法：
初级写法是 “古风女子站在桃花树下”，十次会生成十个样；
专业提示词则要写成 “环绕推近运镜，古风白衣女子静立十里桃林，落英纷飞，黄昏黄金时刻柔光侧逆光，丁达尔光束穿透花枝，中式留白构图，电影画幅，国风水墨写实，胶片颗粒质感，清冷孤寂氛围感”。差别不在词多词少，而在于后者把场景、光线、镜头、风格、氛围一次给全了。
二、视频生成提示词必备要素
一段合格的提示词至少需要包含八个要素：主体、场景、动作、镜头、风格、光照、情绪、氛围。
这八个要素缺一不可：缺少主体，模型就不知道要生成什么；缺少场景，人物就会像贴纸一样浮在虚空中；缺少动作，画面就会有PPT的既视感；缺少镜头，观众就没有被“带着看”的结构；缺少风格，画面就会显得调性混乱；缺少光照，画面会变得扁平灰暗；缺少情绪，表演就会像个空洞假人；缺少氛围，观众的代入感就不强。下面通过一个典型示例，来介绍具备必备要素的提示词应该怎么写。
第一步先写主体：
一位银发白袍的古代剑客
第二步再写场景，场景最好有坐标：
雨后清晨的江南古镇，青石板路湿漉漉
第三步写动作，动作一定要描述具体：
在竹林里舞剑，剑花凌厉
第四步写镜头给视角：
低角度仰拍，镜头缓慢推进
第五步再写光照赋予质感：
黄昏逆光，轮廓光勾勒发丝
剩下的风格、情绪、氛围等也依次展开。按照上述步骤，大家就可以生成一个非常完整的提示词。
三、通用提示词生成框架：SCELA+
上面的八个要素大家记起来还是比较费劲，即梦团队根据以上各个元素，总结了一套名为 SCELA+ 的提示词编写框架。它把提示词拆成场景、角色、表情动作、镜头语言、氛围情绪五个模块，末尾的加号代表可以额外补充光照、特效和参考素材。其实 SCELA+ 框架本身的字母顺序就代表了提示词的生成顺序：先搭舞台，再安排演员，指导表演、设置机位，最后注入灵魂。这正是剧组的开工顺序，也是模型最容易听懂的顺序。
具体含义如下：
- S 是场景，用三到五个具象元素搭出时空坐标：地点 + 时代 + 天气 + 核心道具 + 空间层次；
- C 是角色，用“特征锚点”锁定辨识度：一个核心特征、多个辅助特征，加服装材质；
- E 是表情动作：动作按起承转合进行拆分，表情按眼神、嘴角、微表情进行拆分；
- L 是镜头语言：遵循“景别 + 运动 + 角度 + 焦点”的公式；
- A 是氛围情绪：用通感调动多重感官；
- + 号是补充说明：用来补光照细节、特效参数，或者用 @ 引用素材。
大家在编写提示词时，按照上述顺序进行编写，拼接起来就是一份完整的视频生成提示词，例如下面的例子：
（S）废弃的 90 年代游乐场，黄昏，旋转木马生锈，积水倒映残阳。（C）银发少女，右眼下方有泪痣，褪色牛仔外套，赤脚。（E）低头咬住下唇，眼眶泛红却强忍泪水，缓步离开旋转木马。（L）远景俯拍，缓慢推近至中景，焦点锁定颤抖的手指。（A）潮湿的铁锈味，远处设施嘎吱作响。（＋）夕阳逆光勾勒轮廓，@图片1 参考泪痣位置。
生成的效果如下：
四、画面质量提升提示词关键技巧
除了上述提到的 SCELA+ 写作框架，要让你生成的画面更加完美，还有各种各样的小技巧。利用好这些技巧，会让你的画面更加生动、活灵活现。
4.1 运镜：不只写怎么动，还要写动多少
很多人只会简单描述“移动一下镜头”，这种虚的提示词模型只能瞎猜，结果要么不动，要么乱晃。
其实描写运镜可以从两个角度展开：
第一个角度是运镜方式，主要有推、拉、摇、移、跟、升降。推是向前靠近主体，拉是向后远离，摇是机身不动、镜头左右上下转，移是平行移动，跟是跟随主体，升降是垂直上下；进阶还有环绕、俯冲拉升、变焦、甩镜和轻微振动，分别对应转圈、俯冲、伪运镜、快速切换和手持的紧张感。
另一个角度是运镜参数：运镜速度是匀速还是变速，运镜幅度是“摇镜 30 度”还是“环绕半圈”，运镜起止点是从“人物脚部缓慢拉升至面部特写”，还是“从远景匀速推近至中景”。
所以大家以后可以将“慢慢推”这种含糊的提示词，修改为“匀速推近，3 秒完成”这类可执行指令，才能让画面生成的镜头更加稳固。
4.2 光照：五个维度定住画面质感
光照决定的不只是亮不亮，还有立体感、情绪基调和视觉质感。专业的光照描述，要补足五个维度。
先定主光源的位置，它决定明暗结构。常见就五种：顶光、侧光、逆光、底光、前光。举个例子，如果大家在提示词中写“主光源为 45 度左侧光，人物面部呈现经典伦勃朗三角光”，一句话就可以奠定画面整体的明暗关系与基调。
再定光源的性质。硬光来自直射阳光或聚光灯，阴影边缘锐利，适合强调轮廓和力量感；柔光来自阴天或柔光箱，过渡自然，适合人像。提示词里可以这样写明“硬光边缘锐利”或“柔光漫射过渡自然”。
然后是层次与色温。光的层次可以分为三层：主光定基调、补光填阴影、轮廓光把主体和背景分开。色温一般用来展示情绪，暖光（3000K）复古温馨，冷光（6500K）疏离，青橙对比自带电影感。
最出效果的是光影细节：阴影是软是硬，光线里有没有介质（丁达尔效应、雾中光束、可见的灰尘），表面是什么响应（金属反光、皮肤次表面散射）。一句很优秀的提示词对光影的描述是这样的：“夕阳穿过百叶窗投下条纹阴影，灰尘在光束中飞舞”。这一句加进去，画面立马有了立体的空气感。
4.3 情绪：把抽象感受翻译成身体信号
情绪其实非常难写，也非常决定角色的真实性和观众的代入感，是决定视频生成成败的关键因素。
对视频生成大模型来说，它并不认识具体的情绪，例如“悲伤”；但如果大家这样描述：“眼眶泛红、泪水打转、嘴角颤抖、喉结滚动”，很明显大模型可以生成出来。这就是描写情绪需要用到的“外化翻译法”。简而言之，就是把抽象情绪翻译成可视化的身体信号。
4.4 别写“不要”
在视频生成大模型的提示词编写中，有一套反直觉的原则：尽量别在提示词里写“不要”。对于字节跳动的 Seedance 大模型来说，它不太能理解否认指令，往往是你越强调不要什么，画面里越容易出现什么。所以很多时候我们需要换一个说法。
将“不要手指错误”改成“手指自然弯曲，五指分明”，“不要背景闪烁”改成“画面全程稳定，无闪烁”，“不要多余的人”改成“仅单人出镜，无额外角色”。
4.5 画面数值化：把大概变成具体参数
前面的技巧解决的是画面“像不像”的真实感问题，画面数值化主要用来解决“准不准”的问题。
画面数值化最有用的是时间轴。把 15 秒切成几段，用 [03秒]、[46秒] 标出来，每段只安排一个核心动作。
还有一种数值化的应用是细节的强度。比如“运动强度 7/10”就很实用，限定了范围的同时给出了当前画面的具体强度，模型生成得会更准。
涉及复杂物理效果时，还可以直接声明参数，如“重力系数=0.8”“布料柔韧度=high”。数字不是玄学，而是在告诉模型“到什么程度”。
4.6 多镜头
目前豆包的 Seedance 大模型支持最长生成 15s 的视频。要让这么短的时间具有叙事感，推荐使用多镜头叙事，而不是一个镜头拍到底。
对于时间轴分段的提示词来说，可以这样设计：按“起承转合”拆成三到五段，开头建环境，中间用中景带主体，再用特写抓情绪，最后拉远收尾。段与段之间的光影过渡和叙事衔接，交给模型自己计算。
还有一些进阶的做法是景别递进，让“远景到中景再到特写”形成一条视觉路径；还可以融入“术语”，比如“在恐慌时刻应用希区柯克变焦，随后环绕镜头展示空间关系”。
4.7 提示词精简
学了这么多框架和技巧，这时候大家可能会说：是不是我的提示词把每一方面都写得特别完善，生成的效果就一定好呢？
其实不然。对于 Seedance 模型来说，最佳的提示词生成区间是 30–200 个词，超出这个范围，质量会明显下滑，主要是因为模型的注意力被稀释了。有实测显示，冗余度超过 38%，模型对动作指令的响应准确率会明显降低。
当然了，精简并不是让你机械地删除文字，而是要有“保留动作骨架，剥离修饰脂肪”的思想，一定要保留主体特征和核心动作。对于一些需要大量描写面容的词句，可以使用“@图1 参考面容”上传需要的面容图片，既能保持面容的一致性，又能很好地压缩内容。
下面举一个例子：
原来冗长的提示词：
一位身着锦衣华服、腰佩温润玉佩、面如冠玉、眸光深邃的年轻男子，伫立于落英缤纷的江南庭院中，周遭有奇石、碧潭、曲廊水榭，碧潭中有锦鲤戏水，天际飘着蒙蒙烟雨，他撑着油纸伞，眉宇间流露着淡淡的惆怅，仿佛在思念远方的故人，镜头从远景徐徐拉近，最终定格在他落寞的眼眸上，背景音乐是苍凉的古琴曲，整体氛围凄美动人，电影级画质，8K 分辨率。
精简后的提示词：
古装男子撑油纸伞立于雨中江南园林，眼神忧伤。[0~~4 秒] 远景推近，[5~~8 秒] 特写眼眸。细雨绵绵，古琴 BGM（background music，背景音乐），电影质感。@图 1 参考面容服饰。
生成效果如下：
五、视频提示词生成skill: 上述流程的自动化实现
这时候又有读者朋友要说了：“我也不是专业的呀，上面这些知识也太多啦，我怎么能记得住呢？”作为最宠粉的大模型博主，笔者的目标是让每个人都能具备驾驭 AI 大模型的能力，因此已经把上述流程整理成了一整套 Skill。（如果对 Skill 不熟悉，可以看笔者的文章 Agent Skills完全指南：核心概念丨设计模式丨实战代码）。大家日常使用 workbuddy、codex、claude code 还是 deepseek harness，都可以安装笔者的 Skill。只需简单输入想要描述的场景，Skill 即可辅助你生成一整套专业提示词。下面笔者以 DeepSeek Harness 来演示。如果对 DeepSeek Harness 不熟悉，可以看笔者的专栏：DeepSeek Harness从入门到进阶
- 设定工作区，将Skill放在工作区中，让DeepSeek Harness安装该 Skill。笔者这里已经安装过了。
将 video-prompt-director skill 添加到你的skill技能库中
- 输入想要生成的相关场景，得到提示词加持的智能体会给出需要输入的建议和必需参数
- 按照交互提示输入必要参数，直到生成满意的提示词
- 将提示词输入豆包生成视频
- 最终生成的效果如下，是不是非常简单
以上就是本篇文章的全部内容，video-prompt-director skill大家感兴趣可添加笔者同名微信公众号大模型真好玩并私信视频生成skill 自取。
六、总结
一轮梳理下来，笔者的感受是：视频模型对你“言听计从”的前提，是大家先学会像导演一样说话。从SCELA+框架的五个模块，到运镜、光照、情绪外化、避免“不要”指令、数值化控制、多镜头叙事和提示词精简，这套方法本质上是在用结构化的语言复刻影视工业的决策流程，让模型不再靠猜，而是按指令执行。
工具层面，豆包的Seedance、即梦、可灵各有所长；真正拉开差距的，始终是提示词的结构化程度。结构化提示词不只是一段描述，它本身就是AI视频创作最核心的技术能力。笔者也会继续打磨video-prompt-director skill，让“输入一个场景，得到一套专业提示词”的体验再顺滑一些，后续的更新与分享也会第一时间同步给大家~ 当然笔者的主要文章分享方向还是大模型前沿技术的研究与深入探索，欢迎各位热爱大模型、智能体技术的读者朋友们关注笔者的掘金账号与专栏，也可关注笔者的同名微信公众号大模型真好玩，一起把大模型玩明白~
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Foreword
The author of the article shared how to use Doubao to generate an AI video, with a lot of readers trying to generate their own AI video, but a lot of readers' friends discussed: Why do I always have a "plastic smell"? Some people think it's not enough. Some people think it's a character. In fact, most of the time, the problem arises for one key reason: input quality of prompt. You tend to tell the model who to shoot, but not how to shoot. prompt, “A girl crying in the rain”, has little information: where she cries, what light, how far away she is from shot, and how much she cries, is not clearly described and LLM can only guess.
The author of this article will describe in detail the prompt writing principles of the LLM video: what part it consists of, how it is written in each part, and where the common pit is located. So what? Let's study together.
One, prompt is a director's bag, not just a description.
Many people's understanding of prompt still lingers on “show a word to the video generation LLM”. However, the real prompt is a structured creative instruction to make the subject, scene, photo, shot, drawing, style, quality and atmosphere clear in words. Normal prompt is based on the imagination that professional prompt is based on industrial standards, re-graving the video process with words, and crushing the scenery, light, slot, diagram, color etc.
Look at the following two sets:
The first term is "the woman of the ancient wind stands under the peach tree" and 10 times it produces ten.
Specialized prompt is going to say, "Swirling mirrors, white-haired women standing at a distance of ten miles, flying at a time of twirling gold, light-side reverse light through a flower branch, medium-sized white drawings, film drawings, national wind ink writing, film granular sense, cold and lonely atmosphere." The difference is not that there are fewer words, but that the latter gives the scenes, light, shot, style, atmosphere at once.
II. EQUIPMENT FOR Video Generation prompt
A qualified prompt requires at least eight elements: subject, scene, action, shot, style, light, emotion, atmosphere.
The eight elements are missing: without the subject matter, the model does not know what to generate; without the scene, the person will float like a sticker; without action, the image will have a perception of the PPT; without shot, the audience will not have a structure “with it”; without style, the image will appear to be confused; without light, the picture will be flat and gray; without emotion, the performance will be like an empty false; without atmosphere, the audience will not feel strong. A typical example of how prompt, with the requisite elements, should be written is presented below.
The first step is to write about the subject:
An ancient swordsman in silver and white.
The second step is to write the scene.
In the morning after the rain, Bluestone Road was wet.
Step three, action must be specific:
Swords danced in the bamboo forest. Swords are strong.
Step four, shot for perspective:
Low angle. shot slow advance.
The fifth step is to write about the quality:
Twilight against light, contours twirling.
The rest of the styles, emotions, atmospheres, etc. are in sequence. Following the above steps, you can generate a very complete prompt.
General prompt Generation Framework: SCELA+
The eight elements above are still hard to remember, and the Jimeng team summarized a framework prepared by a prompt called SCELA+ based on the above elements. It dissects prompt into five modules of scenes, role, expression action, shot language, atmosphere, and at the end the plus sign represents additional light, effects and reference asset. In fact, the alphabetical order of the SCELA+ framework itself represents the order in which prompt is generated: first set the stage, then arrange the actors, guide the show, set the slot, and finally inject the soul. That's the order in which the show starts and the order in which the models are most easily understood.
The meaning is as follows:
- S is the scene, with three to five pictograms of the time and space coordinates: location + age + weather + core props + spatial level;
- C is the role, with a "characteristic anchor" for identification: a core feature, multiple auxiliary features, plus clothing material;
- E is an expression: the action is transposed to split, with an eye, a mouth, a micro-expression;
- L is the shot language: following the formula "Specific + Motion + Angle + Focus";
- A is an atmosphere: the mobilization of multiple senses with a sense;
- + is a supplementary description: used to fill in light details, special effects parameters, or to quote asset.
When you were preparing prompt, you did so in the order described above, so you put it together as a complete video generating prompt, for example:
(S) Abandoned '90s playgrounds, twilight, rotting horses rusted and water sprung to the ground. (c) Silver-haired girls, tear moles, faded cowboy coats, barefoot under the right eye. (E) Head down and bite down the lower lip, with red eyes and tears, slowly leaving the carousel. (L) Perspectives push over, slowly push to the midpoint, and focus locks the shivering fingers. (A) The smell of moist rust, which squeaks from far away. (+) Sunset counterlight sketches the contours, @photo1 refer to the position of the tear mole.
The resulting effects are as follows:
IV. prompt KEY TECHNOLOGY FOR INVESTIGATION
In addition to the SCELA+ writing framework mentioned above, there are a variety of small techniques to make your picture better. Use these techniques to make your picture more lively and flexible.
4.1 Mirrors: Not just how to move, but how much to write
Many simply describe “move shot”, a false model of prompt that can only be guessed, with the result that it is either not moving or wandering.
In fact, the mirror can be developed from two angles:
The first is the mirror mode, which is mainly push, pull, shake, move, follow, rise and fall. The push is forward close to the subject, pulls back and forth, shakes back and forth, turns around shot, moves back and forth in parallel, follows the subject, rises and down vertically; the steps are also revolving, dived up, zooming, shatters and slight vibrations, corresponding to the tension of the circle, dives, pseudo-species, rapid switching and handheld, respectively.
Another angle is the mirror parameter: whether the mirror speed is flat or variable, whether the mirror range is "cross 30 degrees" or "circle" and the mirror starting from "slowly pulls the person's foot to face feature" or "spreads from vision to medium."
So you can then change the vague prompt by “slowly pushing” to “spacing up, 3 seconds to complete” so that the shot generated by the image becomes more robust.
4.2 Lights: 5 dimensions to fix the image Sensor.
It's not just the brightness, but the stereo, the emotional tone and the visual quality. Professional light descriptions are supplemented by five dimensions.
First determine the location of the main light source, which determines the bright and dark structure. Five types of light are common: top light, side light, reverse light, bottom light, front light. For example, if you write in prompt, "The main light is 45 degrees left, and the character has the classic Rembrandt triangle on his face," a sentence could lay out the dark relationship and tone of the picture as a whole.
The nature of the re-light source. Hard light comes from direct sunlight or spotlights, the edge of the shadow is sharp, and it is appropriate to emphasize the contours and sense of power; soft light comes from the dark or the soft box, and transitions to nature and is appropriate for human beings. In prompt, it can be stated that “hard light edges are sharp” or “soft light flashing over nature”.
Then the layer with the color temperature. The light can be divided into three layers: the main light sets the tone, the shading of the light, and the contours separate the subject from the background. Colour temperature is usually used to display emotions, warm light (3,000 K) retrospect, cold light (6,500 K) alienation, and orange contrasts with a sense of cinema.
The most effective is light detail: the shadow is soft and hard, there is no medium in the light (butdar effect, beams in the fog, visible dust) and what the surface response is (metallic reflection, dispersion of the subsurface skin). A very good prompt describes the shadow of light as follows: “The sunset casts a striped shadow through the blinds and the dust dances in the beam”. This sentence is added to the image with a 3D air feeling.
4.3 Emotions: translation of abstract feelings into physical signals
Emotions are very difficult to write, and they determine the true nature of the role and the intergenerational perception of the audience, which is a key factor in the success or failure of video generation.
For the video to generate LLM, it does not know specific emotions, such as “sorture”; but if you describe it as “the eyes are red, the tears are turned, the mouth is shaking, the throat is rolling”, it is clear that LLM can be generated. This is the "out-of-the-ground translation" that is required to describe emotions. In short, it is the translation of abstract emotions into visualized body signals.
4.4 Do not write "No"
There is a set of anti-intuitive principles in the prompt writing of the video generating LLM: try not to write “No” in prompt. For byte's beat Seedance LLM, it's hard to understand the denial of instructions, and it's often the more you emphasize not to do anything, the easier it gets in the picture. So many times we need a different narrative.
Replace the words “without errors in the fingers” with the words “naturally curved fingers, with clear fingers”, “without background blinking” with the words “in a stable and non-scrambling picture” and “without redundant people” with the words “only human beings, without additional roles”.
4.5 Numericalization of the image: translating it into specific parameters
The techniques that lie ahead solve the problem of the “like” perception of reality of the picture, which is essentially used to solve the problem of “failure”.
The most useful thing about digitalization is the timeline. Cut 15 seconds into segments, mark it with [03 seconds], [46 seconds] and schedule only one core action for each segment.
There is also a numerical application that is the strength of detail. For example, "Sport strength 7/10" is practical, limiting the scope while giving the specific strength of the current picture and making models more accurate.
Where complex physical effects are involved, parameters such as "gravity factor = 0.8" "combination of fabric = high" can also be directly stated. The numbers are not grammatics, but are telling the model “to what extent”.
4.6 Multishot
Currently Doubao's Seedance LLM supports a maximum of 15s video. In order to give such a short period of time a sense of narrative, it is recommended to use multi-shot narratives instead of a shot to the end.
For prompt in the time-axis segment, it can be designed as follows: three to five segments, to be broken down by “start-and-turn” and to begin with an environment with a middle zone, with a medium area of the subject, with a close-up of emotions, and eventually with a long end. The light transition and narrative connection between the paragraphs is left to the model for its own calculations.
There are also steps forward that allow for a visual path of “looking to the center” to be developed; they can also incorporate “terminologies”, such as “the application of the Hitchcock focal point in times of panic, followed by the display of space relations around shot”.
4.7 prompt Streamlining
Having learned so many frameworks and techniques, one might say, "Did my prompt write everything so well that it would produce good results?
Actually not. For the Seedance model, the best prompt generation range is 30-200 words, beyond which the quality will decline significantly, mainly because the model's attention is diluted. It has been measured that redundancy exceeds 38 per cent and that the model ' s response to action instructions is significantly less accurate.
Of course, streamlining does not allow you to mechanically delete words, but rather to have the idea of “remaining action skeletons and decorating fat” and must retain the main features and core actions. For some words that require extensive presentation, the image of face that is required can be uploaded using “@figure 1 reference face” to keep the face consistent and to compress the content well.
The following is an example:
The long prompt:
A young man dressed in tiara, with a warm waist, with a long face, with a bright look, standing in the courtyard of Jiangnan, surrounded by a cold old song, a cold vibrator, an all-breathing atmosphere, a smoky sky, an oil-paper umbrella, a light twirling twirl between his eyebrows, as if shot had approached him from a long distance, and had finally fallen on his eyes, background music was a pale old song, a darkly moving person, a film-class painting, 8K resolution.
Streamlined prompt:
The old man's oil-paper umbrella was in the middle of the rain and the eyes were sad. [0~4 seconds] Vision closes, [5~8 seconds] closes eyes. Tiny rain, ancient BGM (background music), film quality. @Chart 1 Reference Faces.
The results are as follows:
V. Video prompt generation skill: Automation of the above processes
And it's time for a friend of the reader to say, "I'm not a professional, I'm too knowledgeable to remember." As the most popular producer of LLM, whose aim is to equip everyone with the ability to manage AI LLM, the process has been consolidated into a single Skill. (If not familiar with Skill, the author's full guide to Agent Skills: The Core Concept Design Model is based on the field code). You can always use workbooks, codex, claude code, or deepseek harness. Just input the scene you want to describe, and Skill helps you generate a set of professional prompts. Here's a demonstration by DeepSeek Harness. If you're not familiar with DeepSeek Harness, you can read the column of DeepSeek Harness from the beginning to the next step.
- Set up a workspace, place Skill in the workspace, and let DeepSeek Harness install the Skill. The pen has been installed here.
Add video-prompt-directorkill to your skill skills library Medium
- Enter the relevant scenes that you want to generate, and AI agent, with prompt, gives the suggestions and parameters that need to be entered
- Enter the parameters necessary for interactive alerts until satisfactory prompt is generated
- Enter prompt into Doubao to generate video
- The effect is as follows, isn't it very simple?
This is the full text of this article, and the video of WeChatWeChat Official Account43, also known as WeChat LLM, is very interesting for everyone interested in adding it.
Summary
And the way it's done, the way it's done, the way it's done, the way it's done, the way it's done. From the five modules of the SCELA+ framework, from mirrors, light, emotional externalization, avoidance of “no” commands, numerical control, multi-shot narratives and prompt streamlining, the approach essentially mirrors the industry's decision-making process in structured language, so that models are no longer guessed, but implemented as ordered.
At the tool level, Doubao ' s Feedance, Jimeng, and Kling have their own strengths; what really opens the gap is the extent to which prompt has been structured. Structured prompt is not just a description, but is itself the core technical capability of AI video creation. The writer will also continue to sharpen the video-prompt-director skill, making it easier to “enter a scene and get a set of professional prompt” experiences, and the following updates and sharings will be synchronized for the first time.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.