lengyi-seedance2.5-prompt
设计 Seedance 2.5 视频提示词。输入一段文本(小说片段、一句剧情或几个主题词),扩写成 20-30s(最长
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/woyin2024--lengyi-seedance2-5-prompt)HTML
<a href="/listings/woyin2024--lengyi-seedance2-5-prompt"><img src="/api/listings/woyin2024--lengyi-seedance2-5-prompt/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Who Is This Skill For
Any AI Agent that supports the `Agent Skills` spec (e.g. Codex, Claude Code, WorkBuddy, TRAE Work, Qoder Work, Kimi Work, etc.). As long as your Agent can read and obey a Skill, it can use this to design Seedance 2.5 video prompts.
In principle, models with stronger reasoning are recommended, e.g. Claude Opus 5, K3, DeepSeek-v4-pro, Qwen3.8, GPT-5.6, etc.
Installation
# Clone the repo git clone https://github.com/woyin2024/lengyi-seedance2.5-prompt.git
Drop the entire `lengyi-seedance2.5-prompt` folder into your Agent's Skills directory:
- **Claude Code project-level**: `<project>/.claude/skills/lengyi-seedance2.5-prompt/` - **WorkBuddy user-level**: `~/.workbuddy/skills/lengyi-seedance2.5-prompt/` - **Other Agents**: follow your tool's Skill installation guide
Usage
Once installed, trigger it in conversation any of these ways:
- "Turn this novel excerpt into a Seedance video storyboard" - "Use the lengyi-seedance2.5-prompt skill to design a 30-second multi-shot prompt, the theme is…" - "Make a shotlist, image-to-video"
Trigger keywords (Chinese): `seedance 视频提示词`, `多镜头提示词`, `视频分镜脚本`, `镜头脚本`, `shotlist`, `文生视频 / 图生视频`, `AI 视频镜头脚本`.
Recommended Workflow
1. Use the Skill to generate the "Asset Image T2I Prompts" first 2. Feed the T2I prompts into a text-to-image model (e.g. Image2) to generate subject / scene asset images 3. Feed the asset images + the Skill's "Video Prompt" output into Seedance 2.5
---
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 154 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: lengyi-seedance2.5-prompt
- description: 设计 Seedance 2.5 视频提示词。输入一段文本(小说片段、一句剧情或几个主题词),扩写成 20-30s(最长
- 30s)的连续镜头脚本,输出一条统一格式的多镜头提示词(主体 / 场景 / 音乐 +
- 镜头1、镜头2…),镜头数根据故事和节奏合理安排。先锚定主体与场景资产图,给出对应的文生图提示词,以锁定一致性。当用户要做 seedance
- 视频提示词、多镜头提示词、视频分镜脚本、镜头脚本、shotlist、文生视频/图生视频、AI 视频镜头脚本时使用。
- ---
- # lengyi-seedance2.5-prompt skill:Seedance 2.5 视频提示词设计
- 把用户输入的文本设计成 **20-30s 的连续镜头脚本**,并输出为**一条**可直接复制进 Seedance 2.5 模型的多镜头视频提示词:头部锚定主体、场景、音乐,正文按 `镜头N | …` 顺序排列。
- ## 默认前提(硬性)
- - **必须先锚定资产**:主体与场景都是视频的"资产图",每次输出都要先设计出它们的**文生图(T2I)提示词**,再写视频提示词(见第 4 步)。头部 `主体:`/`场景:` 里填的是这些资产的名称。
- - **音乐行固定不变**:`音乐:不要生成任何背景音乐,只生成对应音效`。这一行是全片唯一的声音指令——**各镜头里不再描述任何音效**,音效交由模型根据画面自动匹配。
- - **镜头正文只描述画面**:不写音效,不写剧情作用/叙事目的。叙事在设计阶段想清楚,但不落进提示词。
- - 默认用中文写提示词。用户要求或目标为海外工具时给英文/中英双语。
- ## 能力边界
- - 输入三类,分别处理:①小说/剧情/脚本片段→提炼最具画面感的故事节奏;②一句简短剧情→补全前因后果与转折;③几个主题词→创意扩写出"起承转合"短片创意,再拆镜头。
- - 输出必须明确:**总时长(默认 20-30s,上限 30s)**、**镜头数(按内容定,无固定档位)**、每镜时间区间。
- - 用户显式指定时长/镜头数则优先;超过 30s 时按 30s 处理并一句话说明。
- ## 工作流程(严格按序执行)
- ### 第 1 步 · 解析与扩写输入
- 在心里列出并补全:
- 1. **故事节奏:**拆成有先后、连续性、逻辑紧密的故事画面。故事节奏是你的设计依据,写进【脚本总览】即可,不写进视频提示词。
- 2. **主体**:有几个角色/物体/道具主体?各自外观线索(缺失则合理设定)。
- 3. **场景**:地点、时间、天气、空间。跨场景时列清场景数量与切换点。
- 4. **基调/类型**:从故事气质推断(科幻/悬疑/温情/热血/广告/武侠…)。
- ### 第 2 步 · 定总时长与镜头数(由故事和节奏决定)
- **镜头数完全取决于故事和节奏**,没有固定档位,也没有数量上限。流程是:定总时长 → 排叙事节拍 → 给每个节拍选镜头类型 → 按类型给时长 → 加总校准。**不要用"总时长 ÷ 平均值"倒推镜头数。,**
- - 总时长:默认 **20-30s**;内容简单取 20-24s,叙事完整取 25-30s。用户另有要求时,则听用户。,
- - **每镜时长按镜头类型定**(核心依据):
- | 镜头类型 | 时长 |
- | - | - |
- | 特写 / 快节奏镜头 | 2-3s |
- | 中景 / 叙事镜头 | 3-6s |
- | 全景 / 固定镜头 | 6-7s |
- - **镜头数是上面选择的结果**。参考量级:30s 视频通常落在 **8-14 镜**(均值约 2-4s);20s 约 7-10 镜。全景、固定镜头多则镜头数偏少,特写快切多则偏多。镜头的数量主要取决于**故事**和**节奏**。
- **反狂切镜头规则**(约束的是"无理由地切",不是"短镜头"):
- 1. **每一次切镜都要有叙事理由**——换空间、换时间、换视角、换信息层级、情绪断点。为炫技而切、为"看起来热闹"而切,一律合并。这是唯一真正的反狂切标准。
- 2. **能一镜说完的不切两镜**:优先用「运镜 + 主体动作 + 环境变化」在一镜内完成推进(推近代替切特写、跟拍代替切两个机位)。
- 3. **分镜节奏要有起伏,不能通篇同一时长**:全篇都是 2-3s 就是狂切,全篇都是 6s 则拖沓。长短交替,让快剪段落成为对比出来的高潮。
- 4. **快剪成组使用**:连续 2-3s 的短镜适合表现打斗、蒙太奇、情绪爆点,成组出现后要用一个中景或全景把节奏收住。
- 5. 长镜头是优点:建立镜可用 6-7s,靠运镜与动作分层填满,不必急着切。
- **时长算术(唯一硬约束)**:各镜时长之和 = 总时长,且总时长 ≤ 30s。每镜标明时间区间(如 `0.0–6.5s`),自检末镜结束时间 = 总时长。
- **节奏规则**:开场用全景/建立镜(6-7s)交代环境与主体 → 中段用中景叙事镜稳步推进 → 高潮切入 2-3s 特写快剪组 → 结尾回到较长镜头收束(≥3s)。
- ### 第 3 步 · 建立主体与场景档案(一致性核心)
- 1. 为**每个主体**写**锚点描述**——固定不变的视觉特征:年龄、发型发色、瞳色、脸型、服装/材质、标志性道具。措辞一旦确定,全程不改。
- 2. 为**核心场景**写场景锚点(地点、时间、光线色调、天气、空间布局);同场景全程沿用同一措辞。
- 3. 给每个主体/场景取一个**短名(资产名)**,如「武松」「猛虎」「景阳冈山道」。头部 `主体:`/`场景:` 只填这些短名,镜头正文也只用短名。
- 4. **多主体**:列出人物关系与出场顺序;每镜标注在场主体。
- 5. **风格**:全片一句统一的风格关键词,写在资产锚定区块,视频提示词里只在镜头1点明一次,后续不重复。
- 6. 锚点档案是**资产图生成依据**,**不在各镜头里逐字复述**(原因见第 5 步)。
- ### 第 4 步 · 写资产图文生图(T2I)提示词(每次必做)
- 主体和场景都要给 T2I 提示词,先生成资产图,再连同视频提示词一起使用:
- - **主体资产(人物/物体/道具)**:背景必须简洁(纯色/白底/浅景深虚化),**只突出主体本身**,不带叙事场景,便于模型提取特征。内容 = 逐字复述锚点描述 + 姿态/角度(正面视角、全身或半身清晰)+ 统一风格 + 画质词(8K、精致细节、studio 柔光)。
- - **场景资产**:展示核心场景的**代表性全貌**,**排除人物与临时道具**,光线/色调/布局与场景锚点一致 + 统一风格 + 画质词。
- - **数量**:主体按实际数量给(通常 1-3 个),场景按实际场景数给(通常 1-2 个)。保持精简,不要为凑数编造资产。
- ### 第 5 步 · 写多镜头视频提示词(只写画面,且是会动的画面)
- **两条核心原则:、**
- - **只描述画面**:镜头正文写的是"摄影机拍到了什么、怎么动的",重点是介绍视频的画面内容。不写音效,不写"剧情:本镜推进了什么",不写情绪解释与叙事意图。
- - **是视频不是静帧**:每镜是一段会动的视频片段,禁止写成静态画面 caption。
- 严格按统一模板(见"输出格式")。每个 `镜头N` 一行,必须含以下要素:
- - **时间区间**(写在镜号后)
- - **场景短名**(多场景时每镜标明当前场景;单场景可在镜头1标明后省略)
- - **主体**:**仅用短名(如"武松")+ 本镜变化点**(如"毡笠帽背系脊梁、发束被风吹乱")。**不复述固定外观**——外观由资产图锁定,重复描写会抢占模型注意力、弱化动作与运镜。
- - **景别 + 镜头运动**:明示方向与速度(缓慢推近 / 拉远 / 上摇 / 横移 / 跟拍 / 环绕 / 手持轻晃 / 慢动作 / POV)。景别要与第 2 步定的时长匹配。6-7s 的建立镜必须写出运镜的**阶段变化**(如"先低角度缓推,至中段上摇至全景");2-3s 的特写只给一个明确、干净的运镜,别塞复杂运动。
- - **动作过程**:强动词写出本镜内的连续动作变化("步伐由快转慢→缓缓抬头→眉头微皱")。3s 以上的镜头至少写 2 个递进动作;2-3s 的短镜只写 1 个动作到位,写多了模型做不完。
- - **时间内变化**:光线、衣物飘动、落叶飞舞、表情转换、色温过渡…
- - **构图**:主体位置、前后景。
- **每镜写完自检(视频化三要素,缺一重写)**:① 至少 1 个强动词;② 至少 1 个镜头运动词;③ 至少 1 个环境/光线动态。
- ❌ 静帧式(错):「武松站在山路上,夕阳余晖洒在身后,目光锐利」
- ❌ 夹叙事/音效(错):「…夕阳从山脊缓缓下沉;音效:碎石脚步、山风。剧情:醉态登场,为猛虎出场铺垫」
- ✅ 纯画面视频式(对):「武松大步从画面深处走来,长袍下摆被风向后吹起,镜头从地面缓慢上摇至全景,夕阳从山脊缓缓下沉」
- 镜头之间要有**景别/机位/情绪的变化与剧情递进**,不要每镜雷同——但递进靠画面本身呈现,不靠文字解说;变化靠设计,不靠多切。
- ### 第 6 步 · 控制总字数(硬约束 ≤4000 字)
- - 全部输出(资产 T2I 提示词 + 多镜头视频提示词,不含表头说明)**总字数 ≤ 4000 字**。
- - 预算:每镜约 `3000 ÷ 镜头数` 字,资产部分 ≤1000 字。
- - 若超字数:优先压缩重复的环境/风格描述,合并次要镜头(顺便改善节奏),不要靠删运镜或动作过程来省字。
- ### 第 7 步 · 交付前质检(逐项核对)
- - [ ] 已明确总时长(默认 20-30s,≤30s)与镜头数,各镜时长之和 = 总时长。
- - [ ] 每镜时长与其镜头类型匹配(特写/快切 1-3s,中景/叙事 3-5s,全景/固定镜头/长镜头 5-7s)。
- - [ ] 镜头数由故事节奏推出,不是拿平均值除出来的。
- - [ ] 时长有起伏,不是通篇同一节奏;快剪组后有较长镜头收住;每次切镜都有叙事理由。
- - [ ] 已给出所有主体与场景的资产图 T2I 提示词;主体用干净背景只展示主体,场景排除人物;措辞与档案锚点逐字一致。
- - [ ] 视频提示词严格按统一模板:`主体:` / `场景:` / `音乐:不要生成任何背景音乐,只生成对应音效` / `镜头N | …`。
- - [ ] 头部主体、场景填的是资产短名,镜头正文也只用短名 + 变化点,不逐字复述固定外观。
- - [ ] 镜头正文**只有画面**:无音效描述、无 `剧情:…`。
- - [ ] 每镜含强动词 + 镜头运动 + 环境/光线动态,无任何一镜读起来像静帧。
- - [ ] 多主体已注明关系与出场顺序,每镜标了在场主体;多场景每镜标了场景。
- - [ ] 风格关键词全片统一,只在镜头1点明一次。
- - [ ] 输出总字数 ≤ 4000 字。
- ## 输出格式
- 按 `references/templates.md` 输出,含完整范例与风格/运镜/光线词库速查。结构:
- 1. **【脚本总览】**:总时长 / 镜头数 / 镜头数理由(节拍如何对应镜头类型)/ 节奏骨架(各镜类型与时长序列)/ 风格 / 主体档案 / 场景锚点。叙事与镜头规划的说明**只写在这里**;
- 2. **【资产图文生图提示词】**:主体资产 + 场景资产;
- 3. **【视频提示词】**:严格套用统一模板的单条提示词(可直接整块复制),**纯画面描述**;
- 4. 文末标注当前总字数,确保 ≤4000。
- 统一模板(不可改动结构与措辞):
- ```
- 主体:主体1、主体2
- 场景:场景名
- 音乐:不要生成任何背景音乐,只生成对应音效
- 镜头1 | xxxx
- 镜头2 | xxxx
- ```
- ## 硬约束(必须遵守)
- 1. 每次必给**主体 + 场景**的资产图文生图提示词(主体干净背景只展示主体;场景展示代表性全貌、排除人物)。
- 2. 默认总时长 20-30s,上限 30s;各镜时长之和 = 总时长。
- 3. 镜头数完全由故事和节奏决定,无固定档位、无数量上限;每镜时长按镜头类型给(特写/快切 1-3s,中景/叙事 3-5s,全景/建立 5-7s),30s 通常落在 8-14 镜。
- 4. 视频提示词严格套用统一模板,音乐行原文不变:`音乐:不要生成任何背景音乐,只生成对应音效`。
- 5. 一致性:头部与各镜只用资产短名 + 本镜变化点,不复述固定外观(外观交给资产图)。
- 6. 镜头正文**只描述画面**:不写音效、不写剧情作用、不写意图解释。
- 7. 每镜是视频提示词不是图片提示词:含强动词、镜头运动、时间内变化三要素。
- 8. 输出总字数 ≤ 4000 字。
README.md · README · 282 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- <div align="center">
- # Seedance 2.5 Prompt.skill
- A video prompt Skill built specifically for **Seedance 2.5**.<br>
- **Turn an idea / a fragment of a theme / a novel excerpt into a polished 30-second film.**
- [](LICENSE)[](./SKILL.md)[](https://www.seedance.ai)[](./references/templates.md)
- <p align="center">
- <a href="./README.md" style="display:inline-block;padding:6px 18px;background:#4f46e5;color:#fff;border-radius:6px 0 0 6px;text-decoration:none;font-weight:600;">English</a> <a href="./README.zh.md" style="display:inline-block;padding:6px 18px;background:#f3f4f6;color:#4b5563;border-radius:0 6px 6px 0;text-decoration:none;font-weight:600;">简体中文</a>
- </p>
- </div>
- <p align="center">
- <img src="public/海报.jpg" width="100%" alt="Poster" />
- </p>
- ---
- ## 🌟 Why This Skill?
- Writing AI video prompts is a path paved with traps.
- It took a long time of stubborn refining before these things became clear:
- **1.** Seedance isn't picky about prompts, but it still helps to spell out the **subject, scene, music, and shots** clearly.
- - **Subject**: specify exactly which reference image; the name must stay consistent throughout.
- - **Scene**: uploading one scene image is enough.
- - **Music**: strictly prompt "do not generate any BGM," only sound effects — otherwise post-production editing becomes a nightmare.
- - **Shots**: timestamps depend on the content, but the shot numbers must be clearly written, e.g. Shot 01, 02.
- **2.** Don't blindly chase resolution. **720P follows prompts best**, 1080P is next, 4K is mediocre.
- **3. The core is still picture control.** You need concrete storyboards, framing, transitions, action, camera moves, and plot in your head — good pictures make good stories.
- **4.** If you write a prompt you wouldn't even read yourself, don't expect Seedance to produce a high-quality video from it.
- These insights hardened into rules, templates, and checklists, and so this Skill was born. **Give it any text — a novel excerpt, a one-line plot, a few theme words — and it returns a multi-shot prompt you can paste straight into Seedance 2.5, along with all the text-to-image prompts for the asset images you need.**
- > If you've been tormented by AI video pacing, consistency, and camera work — this Skill is for you.
- ---
- ## ✨ What It Does
- ### 🎬 Direct-to-Seedance 2.5 Prompts
- | Input Type | How It's Handled |
- |------------|------------------|
- | Novel / script / story excerpt | Distills the most visual story beats into consecutive shots |
- | A single line of plot | Fills in cause, effect, and turning points, then breaks into shots |
- | A few theme words | Creatively expands into a short film with a full arc, then breaks into shots |
- Whatever the input, the output is always the same structure: **【Script Overview】+【Asset Image T2I Prompts】+【Video Prompt】**, ready to copy as a single block.
- ### 🎯 Default Deliverables
- - **Total duration 20–30s** (30s hard cap, auto-collapsed if exceeded)
- - **Shot count decided by story rhythm** — no fixed tiers (30s typically lands at 8–14 shots)
- - **Every shot carries a time range**, and the sum of all shot durations strictly equals the total
- - **All required asset image T2I prompts included**: subjects on clean backgrounds, scenes with people excluded — to lock consistency *before* you generate the video
- ### 📐 Skill Workflow
- 1. **Parse & expand**: break down story rhythm, inventory subjects & scenes, set the tone
- 2. **Set duration & shot count**: lay out shots by story beats, assign duration per shot type, sum and calibrate
- 3. **Build subject & scene profiles**: write anchor descriptions, name assets, lock consistency for the whole film
- 4. **Write asset image T2I prompts**: one for each subject and scene — generate images first, then video
- 5. **Write the multi-shot video prompt**: strictly follow the unified template, describe only moving pictures
- 6. **Control word count**: hard constraint ≤ 4000 characters
- 7. **Pre-delivery QA**: an 11-point checklist — nothing ships until it passes
- ---
- ## 🛠️ Design Philosophy
- This is the most valuable part of the Skill. Every rule below was earned the hard way.
- ### 1. Asset Anchoring: The Real Fix for Consistency
- The instinctive way to hold consistency is to "re-describe the protagonist's appearance in every shot" — that's exactly wrong: the more you restate, the more attention the model spends on restating, and the weaker the motion and camera work become.
- **The fix**: a subject's and scene's fixed traits are written exactly once in the profile, as the basis for generating asset images (T2I); the video prompt uses only the **short name + this shot's delta** (e.g. "Wu Song": straw hat strapped to his back, hair blown loose by the wind), never re-stating "~30yo brawny man, thick brows…". Consistency is locked by the asset images; the prompt is responsible only for "motion."
- ### 2. Configure Shots by Story Type
- Many people first decide "I want 12 shots," then divide total duration by 12 to get 2.5s per shot — the result is a flat wall of fast cuts with no rhythm at all. That's chaos cutting.
- **Hard rule**: shot count is the *result* of laying out story beats and assigning shot types — never a target you set first. Duration is set by shot type: **close-up / fast cut 2-3s, medium / narrative 3-6s, wide / establishing 6-7s**. Every cut needs a narrative reason (change of space / time / viewpoint / information layer / emotional breakpoint). If one shot can tell it, don't use two. Alternate long and short. After a fast-cut group, settle the rhythm with a longer shot.
- Typical 30s skeleton: `Wide establishing (6-7s) → Medium narrative ×2-3 → Close-up fast-cut group (2-3s) ×3-4 → Medium transition → Wide closing`
- ### 3. Write Only the Picture, Don't Distract the Model
- ```
- ❌ Still-frame: Wu Song stands on the mountain path, sunset glow behind him, sharp gaze
- ✅ Video: Wu Song strides out from deep in the frame, his long robe's hem blown back by
- the wind, the camera slowly tilts up from the ground to a wide shot, the sun slowly
- sinks behind the ridge
- ```
- The shot body describes only "what the camera captures and how it moves" — no sound effects (handled by the fixed Music line in the header), no `Plot: …`, no emotional explanation — without distracting the model.
- **The three video-essentials** (self-check after every shot; rewrite if any is missing): ① strong verb ② camera-movement word ③ environment/light dynamic.
- ### 4. Unified Prompt Template
- ```
- Subject: subject1, subject2
- Scene: scene name
- Music: Do not generate any background music, only generate corresponding sound effects
- Shot 1 | 0.0–6.0s | …
- Shot 2 | 6.0–9.5s | …
- ```
- - The Music line is fixed (the single audio instruction for the whole film; no shot describes sound); style keywords are unified film-wide and declared once in Shot 1.
- - **Hard constraints**: the sum of all shot durations = total duration ≤ 30s; total output ≤ 4000 characters. When over budget, compress redundant environment/style descriptions and merge minor shots first — **never save words by cutting camera movement**.
- ---
- ## 📁 Project Structure
- ```
- .
- ├── SKILL.md # Main Skill file (core logic, workflow, hard constraints)
- ├── references/
- │ └── templates.md # Output template + full example + shot-count decision
- │ # cheat sheet + style/camera/light lexicon
- ├── public/
- │ ├── 关于作者.md
- │ ├── 公众号二维码.png
- │ └── 海报.jpg
- ├── README.md # English (this file)
- └── README.zh.md # 简体中文
- ```
- ---
- ## 🚀 Quick Start
- ### Who Is This Skill For
- Any AI Agent that supports the `Agent Skills` spec (e.g. Codex, Claude Code, WorkBuddy, TRAE Work, Qoder Work, Kimi Work, etc.). As long as your Agent can read and obey a Skill, it can use this to design Seedance 2.5 video prompts.
- In principle, models with stronger reasoning are recommended, e.g. Claude Opus 5, K3, DeepSeek-v4-pro, Qwen3.8, GPT-5.6, etc.
- ### Installation
- ```bash
- # Clone the repo
- git clone https://github.com/woyin2024/lengyi-seedance2.5-prompt.git
- ```
- Drop the entire `lengyi-seedance2.5-prompt` folder into your Agent's Skills directory:
- - **Claude Code project-level**: `<project>/.claude/skills/lengyi-seedance2.5-prompt/`
- - **WorkBuddy user-level**: `~/.workbuddy/skills/lengyi-seedance2.5-prompt/`
- - **Other Agents**: follow your tool's Skill installation guide
- ### Usage
- Once installed, trigger it in conversation any of these ways:
- - "Turn this novel excerpt into a Seedance video storyboard"
- - "Use the lengyi-seedance2.5-prompt skill to design a 30-second multi-shot prompt, the theme is…"
- - "Make a shotlist, image-to-video"
- Trigger keywords (Chinese): `seedance 视频提示词`, `多镜头提示词`, `视频分镜脚本`, `镜头脚本`, `shotlist`, `文生视频 / 图生视频`, `AI 视频镜头脚本`.
- ### Recommended Workflow
- ```
- 1. Use the Skill to generate the "Asset Image T2I Prompts" first
- 2. Feed the T2I prompts into a text-to-image model (e.g. Image2) to generate
- subject / scene asset images
- 3. Feed the asset images + the Skill's "Video Prompt" output into Seedance 2.5
- ```
- ---
- ## 🖼️ Output Example
- Input — a single line of plot: **"Wu Song, drunk on Jingyang Ridge, encounters a fierce tiger and fights it bare-handed."**
- Output (excerpt; full version in [`references/templates.md`](./references/templates.md)):
- ```
- 【Script Overview】
- · Total: 30s · Shots: 9 · Avg: 3.3s
- · Shot-count rationale: beats are "entrance — awareness — tiger appears —
- dodge — staff breaks — pin down — exhaustion — ending." The opening needs
- one 6s wide shot to establish the barren mountain and the drunk at once;
- the midsection uses two medium narrative shots; the fight is the emotional
- peak, handled with a group of three 2.5s close-up fast cuts; then it
- returns to medium-close and wide to settle the rhythm.
- · Rhythm skeleton: Wide 6.0 → Medium 3 → Close-up 2 → Medium follow 4.0
- → Close-up×3 → Medium-close 3.0 → Wide 3.5
- · Style: cinematic wuxia, film grain, realistic texture, warm sunset to cold moon
- 【Asset Image T2I Prompts】
- - Subject "Wu Song": ~30yo brawny man, thick brows, square face, black beard,
- hair tied high, earth-yellow coarse-cloth short martial robe, dark-brown
- cloth sash at waist, straw hat on back, holding a wooden staff, full-body
- front view, cinematic wuxia realism, film grain, clean white background
- showing only the subject, 8K, fine detail, studio soft light
- - Subject "Fierce Tiger": adult South China tiger, orange-yellow with black
- stripes, sturdy build, eyes glowing green, side-body coiled-strike pose…
- - Scene "Jingyang Ridge mountain path": barren mountain dirt road, scattered
- rocks and dry grass, dead trees leaning in on both sides, blood-red sunset…
- 【Video Prompt】
- Subject: Wu Song, Fierce Tiger
- Scene: Jingyang Ridge mountain path
- Music: Do not generate any background music, only generate corresponding sound effects
- Shot 1 | 0.0–6.0s | Jingyang Ridge mountain path, cinematic wuxia, film grain,
- realistic texture. Extreme wide shot, low-angle slow push-in, tilting up over
- the scattered rocks at the midpoint, settling on a medium shot of the figure
- by the end. Subject "Wu Song" stumbles in his steps, flushed with drink,
- striding out from deep in the frame, staff dragging on the ground carving a
- dirt trail, slowing as he nears and looking around. Long robe hem blown back
- by the mountain wind, sun keeps sinking behind the ridge, shadow stretching
- ever longer.
- Shot 2 | 6.0–9.5s | Medium shot, camera steadily tracking Wu Song's right side.
- Wu Song lifts a gourd, tilts his head back to drink, wipes his mouth…
- (remaining shots in the full example)
- ```
- ---
- ## 🤝 Contributing
- This Skill was born from repeatedly refining something that didn't feel right when I used it — but it can certainly be better.
- If you find, while using it, that:
- - A certain input type (ads, anime, product showcases…) has room for better shot rhythm
- - The lexicon needs additions (new camera, light, or environment-dynamic terms)
- - The template could fit Seedance 2.5's latest features more tightly in specific scenarios
- - You'd like more output modes (English-only / bilingual)
- contributions are very welcome:
- 1. Fork this repo
- 2. Create a feature branch: `git checkout -b feature/your-feature`
- 3. Commit your changes: `git commit -m 'Add some feature'`
- 4. Push the branch: `git push origin feature/your-feature`
- 5. Open a Pull Request
- Extra credit for PRs that include **a real, run-through case** (input → output → model-generated result). That's worth more than any abstract discussion.
- ---
- ## 📜 License
- This project is open-sourced under the [MIT License](LICENSE).
- For personal learning use, you're free to use, modify, and distribute it; for commercial use, please obtain authorization in advance — contact via WeChat: lengcp2013.
- ---
- ## 👤 About the Author
- **Leng Yi (冷逸)**, blogger at **沃垠AI (WoyinAI)**, one of China's top AI tech media outlets. A Vibe Coding developer who doesn't write code, obsessed with Prompt, Skills, and Agent engineering.
- - **Unified handle across platforms**: 沃垠AI (WoyinAI)
- - **Tags**: Product & operations background, in-depth AI reviews, super OPC
- - **Content matrix**: WeChat Official Account, Xiaohongshu, Zhihu, GitHub, Bilibili, X, and more
- Follow the **沃垠AI** WeChat Official Account for more AI insights, hands-on Prompt engineering, and Skill design methodology:
- <p align="center">
- <img src="public/公众号二维码.png" width="90%" alt="沃垠AI WeChat Official Account QR Code" />
- </p>
- ---
- <div align="center">
- ### If it helped you, leave a ⭐
- **Thank you 🙏**
- </div>
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.