scenario-kinetic-music-video

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Kinetic Music Video

动态音乐视频

A song goes in. A 1080p60 music video comes out, landscape for YouTube and X or vertical for TikTok, Reels and Shorts: a motion designer's showreel cut to the track. Kinetic type, 2D and 3D graphics and transitions land on the beats and sung syllables. Generated footage of the subject is woven through it where the concept wants a performer. The approved master is the only soundtrack. Preserve the original; an authorized audio edit produces a new, versioned master.

The work splits by what each tool does best:

Scenario supplies what code can't: a consistent subject with real physics, choreography, hair and lip-sync, in whatever art style each section needs.
The JS engine supplies what video models can't: exact typography, frame-exact timing to onsets and beats, infographics, compositing around the subject, and endless re-renders at zero generation cost.
Sub-agents give each section a dedicated designer. A shared style bible and a verification loop keep them coherent.

Connection and the core generation loop: see the

scenario

skill in this repo. For a song-to-video cut without the engine, see

scenario-seedance-music-video

. If a sibling skill named here is missing from your available skills, ask the user to install it (

npx skills add scenario-labs/skills --skill <name>

); unattended, proceed from tool schemas and flag the gap.

Read

references/lessons.md

before starting. Every item in it cost real time or money once.

references/motion_library.md

is the design menu.

输入一首歌曲，输出1080p60规格的音乐视频，支持适配YouTube和X的横屏格式，以及适配TikTok、Reels和Shorts的竖屏格式：如同动态设计师的作品集剪辑，完全贴合音乐节奏。动态文字、2D/3D图形和转场效果精准踩中节拍和演唱音节。当创意概念需要表演者时，会插入生成的角色素材。经批准的母带是唯一的音轨来源，需保留原始文件；仅在授权的音频编辑后生成新的带版本号的母带。

工作流程根据各工具的优势分工：

Scenario 提供代码无法实现的效果：风格统一的角色，具备真实物理效果、编舞、头发动效和口型同步，可适配每个片段所需的任意艺术风格。
JS引擎 提供视频模型无法实现的效果：精准的排版、与起始点和节拍帧级同步的时序、信息图表、围绕角色的合成效果，以及零生成成本的无限次重渲染。
Sub-agents（子Agent） 为每个片段分配专属设计师，通过共享的风格指南和验证循环保持整体风格一致。

连接方式和核心生成循环：参见本仓库中的

scenario

技能。如果不需要引擎，仅需歌曲转视频剪辑，参见

scenario-seedance-music-video

。如果你的可用技能中缺少这里提到的同类技能，请让用户安装（命令：

npx skills add scenario-labs/skills --skill <name>

）；在无人值守模式下，基于工具schema继续执行并标记缺失项。

开始前请阅读

references/lessons.md

，其中的每一条经验都曾耗费过实际的时间或成本。

references/motion_library.md

是设计选项菜单。

Before you start (ask once, then run)

开始前准备（询问一次后执行）

Workspace: the Scenario team and project (
```
teams_list
```
; ask if there is more than one).
Subject:
- Who or what performs: a character from reference images, a mascot, a product, the artist's own likeness (their own photos and consent only), or nobody (a pure motion-graphics or lyric video).
- Footage share: performer-led, a mix (often about 60/40), or graphics only.
Look:
- Ask for references: images, videos or links for the mood, the motion and the color. Vague color asks without references drift to safe palettes.
- Decide the 2-4 looks and which sections own which.
- Ask what must never appear.
- Defaults: generated footage shown clean (no filters on it), graphics layered on and around it, a fast and upbeat showreel energy.
Brand: any logo, wordmark or credit. Use official files in their official colors, only where the director wants them.
Budget: lean (6-10 video generations), moderate (12-20) or big. Price each paid step with
```
dry_run
```
(one run per distinct payload, then total) before spending. Relative anchors at 720p: a 6 s clip costs about twenty style frames, a 10 s clip about thirty-five, and a full mixed video about five hundred. A 1080p clip measured about 2.5 times its 720p price, so settle the footage resolution before quoting.
Format: one canvas per project, asked once: portrait 9:16 for TikTok, Reels or Shorts, landscape 16:9 for YouTube or X (unattended with no destination: landscape). It is
```
CANVAS
```
in
```
engine/timeline.js
```
, and style frames and footage are generated at that ratio, never cropped from the other. On portrait, check the destination's current length cap and plan the teaser as the platform cut when the song runs longer, or when you cannot check. Both ratios only on request: the code re-renders free with
```
--canvas
```
and most shots are windowed, but each shot that must fill the second frame (hook, lip-sync close-ups, finale) is paid again (
```
references/pipeline.md
```
§9). X caps standard accounts at 2:20, so plan a teaser.
Lyrics: use them if supplied (they beat any transcript). Otherwise transcribe and flag the uncertain words at the end.

With no one around to answer, choose sensible defaults for the creative choices and write them into TREATMENT.md. Do not default the money: with no stated budget or no explicit team and project, stop before the first paid step and say what is missing.

工作空间： Scenario的团队和项目（通过
```
teams_list
```
获取；如果有多个则询问用户）。
角色主体：
- 表演者身份：参考图片中的角色、吉祥物、产品、艺人本人形象（仅可使用其本人提供的照片且需获得同意），或无表演者（纯动态图形或歌词视频）。
- 素材占比：以表演者为主、混合（通常约60/40）或纯图形。
视觉风格：
- 索要参考资料：用于氛围、动效和色彩参考的图片、视频或链接。没有参考的模糊色彩需求会倾向于使用安全调色板。
- 确定2-4种视觉风格，以及每个片段对应的风格。
- 询问绝对不能出现的内容。
- 默认设置：生成的素材保持纯净（不加滤镜），图形层叠加在素材之上或周围，整体呈现快速活泼的作品集风格。
品牌元素： 任何logo、文字商标或鸣谢信息。仅在导演需要的位置使用官方文件和官方配色。
预算： 精简（6-10次视频生成）、中等（12-20次）或充足。在产生费用前，通过
```
dry_run
```
预估每个付费步骤的成本（每个不同的请求负载运行一次，再汇总总价）。720p分辨率下的参考价格：6秒片段约等于20张风格帧的成本，10秒片段约35张，完整混合视频约500张。1080p片段的价格约为720p的2.5倍，因此报价前需先确定素材分辨率。
格式： 每个项目仅使用一种画布比例，只需询问一次：适配TikTok、Reels或Shorts的9:16竖屏，适配YouTube或X的16:9横屏（无人值守且无指定平台时默认横屏）。画布比例对应
```
engine/timeline.js
```
中的
```
CANVAS
```
参数，风格帧和素材都会按该比例生成，绝不会从另一种比例裁剪而来。如果是竖屏格式，需检查目标平台当前的时长限制，当歌曲时长超过限制或无法检查时，将预告片规划为平台适配版本。仅在用户要求时才提供两种比例：代码可以通过
```
--canvas
```
参数免费重渲染，且大多数镜头是窗口式的，但每个必须填满第二种比例画幅的镜头（钩子片段、口型同步特写、结尾）需要重新付费生成（参见
```
references/pipeline.md
```
第9节）。X平台标准账号的视频时长上限为2分20秒，因此需要规划预告片。
歌词： 如果提供了歌词就直接使用（比任何转录版本都准确）。否则自行转录，并在末尾标记不确定的歌词。

如果无人回复，为创意决策选择合理的默认值并写入TREATMENT.md文件。不要对预算做默认设置：如果没有明确说明预算，或没有明确的团队和项目，在第一个付费步骤前停止并告知缺失的信息。

Workflow

工作流程

1. Scaffold and analyze the song

1. 搭建项目框架并分析歌曲

Run

bash <skill>/scripts/scaffold.sh <song>

inside a new project folder, then follow

references/pipeline.md

§2:

```
audio_analysis.py
```
writes stems and
```
engine/data/audio.json
```
: beats, downbeats, kick/snare/vocal onsets, and 60 fps envelopes. Measure the tempo; labels lie (a "130 BPM" song was 98 with double-time drums).

lyrics_align.py transcribe

, then hand-correct

analysis/lyrics.txt

, then

lyrics_align.py align

writes

engine/data/lyrics.json

with word onsets.

Check timing by eye with
```
plot_lyrics.py
```
. Audition when playback is available; a spectrogram locates onsets but cannot prove pronunciation or musical continuity.

在新项目文件夹中运行

bash <skill>/scripts/scaffold.sh <song>

，然后按照

references/pipeline.md

第2节执行：

```
audio_analysis.py
```
会生成音轨分轨和
```
engine/data/audio.json
```
文件：包含节拍、重拍、底鼓/军鼓/人声起始点，以及60fps的包络线。务必实测速度；标签信息可能不准（比如一首标注“130 BPM”的歌曲实际是98 BPM，只是用了双倍速鼓点）。
运行
```
lyrics_align.py transcribe
```
转录歌词，然后手动修正
```
analysis/lyrics.txt
```
，再运行
```
lyrics_align.py align
```
生成包含单词起始点的
```
engine/data/lyrics.json
```
文件。
使用
```
plot_lyrics.py
```
肉眼检查时间轴。如果有播放功能就试听；频谱图可以定位起始点，但无法验证发音或音乐连贯性。

1a. Edit the soundtrack when requested

1a. 按需编辑音轨

For pronunciation repairs, lyric changes, section replacements or structural cuts, read references/audio_editing.md. Use

scenario-ace-step

for its Cover/Repaint contracts. Keep the current cut and master until a candidate is selected; preserve audio outside the agreed window and rebuild affected timing data after applying it. A picture-only revision reuses the approved soundtrack.

如果需要修正发音、修改歌词、替换片段或结构剪辑，请阅读references/audio_editing.md。使用

scenario-ace-step

的Cover/Repaint合约功能。在候选版本被选中前，保留当前剪辑和母带；保留约定编辑窗口之外的音频内容，应用修改后重建受影响的时间轴数据。仅修改画面的修订版本可复用已批准的音轨。

2. Research (parallel, background)

2. 调研（并行后台执行）

Launch two sub-agents with web access while you work on other things:

Culture: for the song's subject, find recognizable events, memes and visual tropes, each with a verified fact, its iconic visual form, an insert idea and a short on-screen string.
Craft: study the user's motion references (download them, contact-sheet them, read the frames), plus genre direction techniques. Extract timing numbers, color hex values and implementable JS moves.

Save the condensed results as

analysis/research_*.md

. Scene agents read them alongside

references/motion_library.md

在你处理其他工作时，启动两个具备网络访问权限的子Agent：

文化调研： 围绕歌曲主题，查找知名事件、迷因和视觉梗，每个都要包含已验证的事实、标志性视觉形式、插入创意和简短的屏幕文字。
技术调研： 研究用户提供的动态参考资料（下载、制作联系表、逐帧分析），以及该类型的导演技巧。提取时间参数、颜色十六进制值和可实现的JS动效。

将精简后的结果保存为

analysis/research_*.md

。场景Agent会结合这些文件和

references/motion_library.md

一起阅读。

3. Concept, style bible and treatment

3. 创意概念、风格指南和制作方案

Give the video a spine. The strongest concepts make the edit itself carry the song's meaning: a device that progresses through the whole video, for example a counter, clock or countdown that moves with the song while the cut rate or the looks escalate with it. The spine usually lives in the HUD counters.
STYLE.md (the law):
- a one-line concept and the spine
- the footage rules (clean plate and the back/mid/front layers, or another policy)
- the looks, each with ground, ink, accent, rendering rule and matching footage style
- the palette with roles
- 4-5 type roles with named fonts
- lyric display rules
- a motion grammar mapping beats to effects, with the cut density per section
- the subject (identity, design details usable as graphic motifs)
- brand rules, references and bans
TREATMENT.md:
- song facts and the energy map
- the arc
- the hook for the first 3 seconds (the strongest idea in the video, and it must work as a muted autoplay)
- one row per section: time range on downbeats, lyrics, look, must-have ideas
- a footage map (filled in as clips land)
PLAN.md: a one-page version for the director to approve before you spend on footage (unattended, this is the stop point if the budget is unstated).

为视频设定核心主线。 最优秀的创意会让剪辑本身承载歌曲的含义：一个贯穿整个视频的核心装置，比如随着歌曲推进的计数器、时钟或倒计时，同时剪辑节奏或视觉风格也随之升级。主线通常体现在HUD计数器中。
STYLE.md（最高准则）：
- 一句话概念和核心主线
- 素材规则（纯净底版和后/中/前层，或其他规则）
- 视觉风格：每种风格包含底色、线条色、强调色、渲染规则和匹配的素材风格
- 调色板及各颜色用途
- 4-5种文字角色及对应指定字体
- 歌词显示规则
- 动态语法：节拍与效果的映射，以及每个片段的剪辑密度
- 角色主体（身份、可作为图形母题的设计细节）
- 品牌规则、参考资料和禁用内容
TREATMENT.md（制作方案）：
- 歌曲基本信息和能量分布图
- 故事弧光
- 前3秒钩子（视频中最有冲击力的创意，必须在静音自动播放时也能吸引人）
- 每个片段一行：重拍对应的时间范围、歌词、视觉风格、必备创意
- 素材分布图（随着素材到位逐步填充）
PLAN.md： 单页版本的方案，供导演在你投入素材生成费用前批准（无人值守模式下，如果预算未明确，这一步是停止点）。

4. Subject, style frames and footage on Scenario

4. 在Scenario上生成角色、风格帧和素材

Skip this for a graphics-only video, apart from any stills or 3D props. Otherwise follow

references/pipeline.md

§3-5:

Subject sheet: a turnaround, expressions and callouts from the user's references. Look at it, then crop identity references from it.
Style frames: one still per shot at the canvas ratio (an image model that takes reference images), in its section's look. Compose with negative space for type (on portrait, the face and the type space inside the safe band), on a set that composites (black void with rim light, flat color cyc, or white high-key). No text. In a lip-sync frame nothing covers the mouth (no mic, hand or prop): a covered mouth fails the sync check and is paid again. Contact-sheet them and get the look approved cheaply (unattended: proceed only inside a stated budget).
Footage routes:
- Image-to-video from the frame for actions.
- A video model with reference images and reference audio, given a synthesized beat track for dances that lock to the song, then beat-warp.
- Audio-to-video from a frame plus a vocal-stem slice for lip-sync. The alternative is reference-audio video plus a lip-sync correction model.
- Code instead of generation for networks, charts, UI and particles.
Never send the song to a video model. Turn the model's native audio off when its schema offers it (the render takes sound only from the master either way), and launch everything in parallel (
```
jobs_wait
```
re-called with
```
pending_job_ids
```
, never
```
job_get
```
, never a relaunch).
Prep every clip with
```
tools/prep_clip.sh
```
: frames, a subject matte (Apple Vision, with Scenario background removal as the fallback; see
```
references/pipeline.md
```
§5) and tracking (MediaPipe pose plus the matte bbox). Inspect the mattes. Write each clip's content, key moments, slot and matte quality into the footage map.

如果是纯图形视频，除了静态图或3D道具外，跳过此步骤。否则按照

references/pipeline.md

第3-5节执行：

角色设定表： 根据用户提供的参考图，生成角色的多角度视图、表情和标注说明。查看后从中裁剪出身份参考图。
风格帧： 每个镜头一张对应画布比例的静态图（使用接受参考图的图像模型生成），采用对应片段的视觉风格。构图时要为文字留出负空间（竖屏格式下，人脸和文字区域都要在安全区内），背景采用便于合成的场景（带轮廓光的黑色虚空、纯色背景板或白色高调背景）。不要加文字。口型同步的帧中不能有任何东西遮挡嘴巴（不能有麦克风、手或道具）：嘴巴被遮挡的帧会通不过同步检查，需要重新付费生成。将风格帧拼成联系表，低成本地获得风格确认（无人值守模式下：仅在明确预算内继续执行）。
素材生成路径：
- 基于风格帧的图生视频，用于生成动作。
- 使用带参考图和参考音频的视频模型，提供合成节拍轨来生成贴合歌曲的舞蹈动作，然后进行节拍扭曲对齐。
- 音频生视频： 基于一帧画面加人声音轨切片生成口型同步效果。替代方案是用参考音频生成视频，再搭配口型同步修正模型。
- 网络、图表、UI和粒子效果用代码实现，不用生成模型。
绝对不要把整首歌发送给视频模型。如果模型的schema支持关闭原生音频，就关掉（反正渲染时只会从母带获取音频），并且所有任务并行启动（通过
```
pending_job_ids
```
重复调用
```
jobs_wait
```
，绝不要用
```
job_get
```
，绝不要重新启动任务）。
使用
```
tools/prep_clip.sh
```
预处理每个片段：生成帧、角色遮罩（使用Apple Vision，备用方案是Scenario的背景移除功能；参见
```
references/pipeline.md
```
第5节）和跟踪数据（MediaPipe姿态检测加遮罩边界框）。检查遮罩质量。将每个片段的内容、关键时刻、对应位置和遮罩质量写入素材分布图。

5. Engine and scene agents

5. 引擎和场景Agent

Smoke-test first:
- Set
```
PAL
```
  in
```
core.js
```
  , and in
```
engine/timeline.js
```
  set
```
CANVAS
```
  , the sections on downbeats,
```
HUD
```
  (title, mark, section codes, spine counters) and
```
POST
```
  (neutral by default).
- Write a tiny
```
_smoke.js
```
  : plate → type behind the subject → matte → tracked label.
- Render one still and a 5 s clip, then confirm audio and flashes line up with
```
av_sync_check.py
```
  .
Brief the agents. Fill in
```
AGENTS_BRIEF.md
```
from the template, then launch one background agent per section in a single message. Each gets its file, time range, lyrics, must-have ideas and clips, and reads AGENTS_BRIEF, STYLE, TREATMENT, ENGINE_API and the research first. Log the agent ids. For a teaser or anything under about 30 s, the lead may write the scenes itself and skip the research agents, PLAN.md and the per-section agents; keep the smoke test.
Deliver clips as they land. Message the agent that needs each one ("clip ready: content, key moments, slot").
Review every report. Contact-sheet the best stills and Read them. Fix engine-level problems yourself (they affect everyone) and tell the agents what changed.

先做冒烟测试：
- 在
```
core.js
```
  中设置
```
PAL
```
  参数，在
```
engine/timeline.js
```
  中设置
```
CANVAS
```
  、基于重拍的片段划分、
```
HUD
```
  （标题、标记、片段代码、主线计数器）和
```
POST
```
  （默认中性效果）。
- 编写一个简短的
```
_smoke.js
```
  文件：底版 → 角色后方文字 → 遮罩 → 跟踪标签。
- 渲染一张静态图和一个5秒的片段，然后用
```
av_sync_check.py
```
  确认音频和闪光效果同步。
给Agent下达指令。 从模板填写
```
AGENTS_BRIEF.md
```
，然后在一条消息中为每个片段启动一个后台Agent。每个Agent会获取对应的文件、时间范围、歌词、必备创意和素材，并且会先阅读AGENTS_BRIEF、STYLE、TREATMENT、ENGINE_API和调研资料。记录Agent的ID。对于预告片或时长约30秒以下的内容，主导Agent可以自行编写场景，跳过调研Agent、PLAN.md和分片段Agent；但保留冒烟测试。
素材就绪后立即交付。 向需要该素材的Agent发送消息（“素材就绪：内容、关键时刻、对应位置”）。
审核所有报告。 将最佳静态帧拼成联系表并查看。引擎层面的问题自行修复（因为会影响所有Agent），并告知Agent修改内容。

6. Assemble, watch, iterate

6. 组装、观看、迭代

Render the full cut (3 workers, about 5 min for 2:00 on an M-series Mac). Wait for it in the foreground (or poll until it prints
```
wrote
```
) and carry on to verification and delivery: never end the run while a render or encode is still going, because a headless session that stops kills it.
Verify:
- ```
av_sync_check.py
```
  : expect 0.0 ms audio and 0 to +1 frame visual. Below a correlation of about 0.15 the visual lag means nothing (type leading by
```
E.LEAD
```
  reads as up to -3 frames): check onset stills instead.
- ```
freezedetect
```
  : no frozen footage.
- Contact sheets at 0.5 s steps, 20 s per sheet.
- Every section boundary, frame by frame.
- On portrait,
```
still.mjs --safe
```
  sheets, 20 s each: no word, logo or face outside the outlined band.
Global look pass: a still sheet across all sections, then tune
```
POST
```
, the palette and the HUD theme.
Revisions: map each director note (they arrive as timestamps) to its section and send it back to the owning agent via SendMessage. Resuming keeps its context. Fix a weak shot by replacing it with a new generation in a different look, or with code, never by adding effects. Offer alternatives as timeline
```
variant
```
s (render with
```
--variant b
```
) instead of guessing.

渲染完整剪辑（3个工作进程，M系列Mac上2分钟的视频大约需要5分钟）。在前台等待（或轮询直到输出
```
wrote
```
），然后继续验证和交付：绝对不要在渲染或编码还在进行时结束运行，因为无头会话停止会杀掉进程。
验证：
- ```
av_sync_check.py
```
  ：预期音频偏差0.0毫秒，视觉偏差0到+1帧。如果相关度低于约0.15，视觉延迟数据没有意义（文字提前
```
E.LEAD
```
  值可能显示为最多-3帧）：改为检查起始点静态帧。
- ```
freezedetect
```
  ：确保没有冻结的素材。
- 每0.5秒一帧的联系表，每张覆盖20秒内容。
- 逐帧检查每个片段的边界。
- 竖屏格式下，生成
```
still.mjs --safe
```
  联系表，每张20秒：确保没有文字、logo或人脸超出安全区轮廓。
全局视觉调整： 生成覆盖所有片段的静态帧联系表，然后调整
```
POST
```
效果、调色板和HUD主题。
修订： 将每条导演意见（通常以时间戳形式给出）对应到所属片段，通过SendMessage发回给负责的Agent。恢复会话会保留上下文。修复不佳的镜头时，用不同风格的新生成素材替换，或用代码实现，绝不要靠加特效。提供时间轴
```
variant
```
（变体）作为替代方案（用
```
--variant b
```
渲染），而不是自行猜测。

7. Deliver

7. 交付

Videos: a master, an upload file (landscape: the master for YouTube, an X-ready 2-pass 21 Mb/s file under 512 MB for X; portrait: the vertical upload in
```
references/pipeline.md
```
§8), a teaser (verse plus first chorus), plus one file per variant, and one set per canvas, suffixed
```
_16x9
```
and
```
_9x16
```
, when both were asked for. Use distinct filenames, because macOS is case-insensitive.
Stills: full-res PNGs rendered by the engine, never grabbed from the compressed video.
Style sheet / making-of: publish it as an artifact or doc: palette, type, subject sheet, source-vs-final footage pairs, a frame index, the pipeline and iteration counts.
Report: tell the user what was verified, which lyrics you guessed, and the CU spent, summed from this run's successful
```
jobs_wait
```
rows (
```
cuCost
```
, logged in
```
analysis/jobs.md
```
; a failed job is refunded but its row keeps a
```
cuCost
```
);
```
usage
```
is project-wide and includes other work.

视频文件： 母带、上传用文件（横屏：YouTube用母带，X用符合平台要求的2次编码21Mb/s、大小低于512MB的文件；竖屏：按照
```
references/pipeline.md
```
第8节的竖屏上传规范）、预告片（主歌加第一段副歌），以及每个变体对应一个文件；如果要求两种画布比例，则每种比例一套文件，分别加上
```
_16x9
```
和
```
_9x16
```
后缀。使用不同的文件名，因为macOS不区分大小写。
静态图： 引擎渲染的全分辨率PNG文件，绝不要从压缩视频中截图。
风格表/制作记录： 作为制品或文档发布：包含调色板、字体、角色设定表、原始素材与最终素材对比、帧索引、工作流程和迭代次数。
报告： 告知用户验证了哪些内容、哪些歌词是猜测的，以及消耗的CU总量，从本次运行成功的
```
jobs_wait
```
记录中求和（
```
cuCost
```
字段，记录在
```
analysis/jobs.md
```
中；失败的任务会退款，但记录中仍保留
```
cuCost
```
）；
```
usage
```
是项目级别的用量，包含其他工作的消耗。

Quality bar

质量标准

Every word shows on time. Each sung word is visible at its onset (as a hero or in the subtitle track), and hero words peak on the onset frame.
Integration, not decoration. Every footage moment has a graphic that reacts to it (behind via the matte, tracked to it, timed to its move, or continuing its shapes). The footage pixels are untouched.
No static frames. Footage always plays live, and something always moves.
One focal point per frame. Type frames the subject and never covers the face. Behind-the-subject type still reads. On portrait, every word, logo and face sits inside the safe band.
Section variance. Mix the looks, 2D and 3D, poster slams and quiet subtitles. Never use the same type treatment on two consecutive lines. Cut density climbs with the song's energy, and showcase moments (logo, end card) get time.
Photosensitivity. At most 3 full-frame luminance flips per second.
Watch it yourself before calling it done, then take a second look at the first 3 seconds.

每个字都准时出现。 每个演唱的单词都在其起始点可见（作为主视觉文字或在字幕轨中），主视觉文字在起始帧达到峰值效果。
融合而非装饰。 每个素材画面都有对应的图形与之互动（通过遮罩在后方、跟踪素材、与动作同步，或延续素材的形状）。素材的原始像素不做修改。
无静态帧。 素材始终动态播放，总有元素在运动。
每帧一个焦点。 文字围绕角色排版，绝不遮挡脸部。角色后方的文字也要清晰可读。竖屏格式下，所有文字、logo和人脸都要在安全区内。
片段多样性。 混合使用不同视觉风格、2D和3D效果、海报式跳切和安静的字幕。连续两行歌词绝不使用相同的文字处理方式。剪辑密度随歌曲能量提升，展示性时刻（logo、结尾卡）要留足时间。
光敏性保护。 每秒最多3次全画幅亮度翻转。
完成前务必亲自观看，然后再重点检查前3秒。

Files in this skill

本技能包含的文件

Setup and analysis: scaffold.sh, audio_analysis.py, lyrics_align.py, plot_lyrics.py
Approved-master stem lookup for transcription, plots and vocal slices: audio_paths.py
Footage generation helpers: synth_beat.py, beatwarp.py
Footage prep: prep_clip.sh, prep_video.py, matte.swift, fixmatte.py, matte_keyed.py, matte_fallback.py, track.py
Sync checks: lipsync_check.py, dance_sync_check.py, av_sync_check.py
Rendering and review: render.mjs, still.mjs, serve.mjs, sheet.py
Engine, copied into the project by scaffold.sh: index.html, main.js, canvas.js, core.js, plate.js, roto.js, typekit.js, hud.js, plus timeline.template.js and scene.template.js
references/pipeline.md: every command and Scenario call, step by step
references/engine_api.md: the scene contract and the full
```
E
```
API (canvas and safe band, plates, mattes, tracking, post, HUD, variants)
references/motion_library.md: looks, type system, a 40-move technique library, integration ideas, transitions, timing, easing and springs, sync and cut density
references/agent_brief_template.md: the brief each scene agent reads
references/lessons.md: production gotchas and direction defaults that held

搭建与分析：scaffold.sh、audio_analysis.py、lyrics_align.py、plot_lyrics.py
已批准母带的音轨查找（用于转录、绘图和人声切片）：audio_paths.py
素材生成辅助工具：synth_beat.py、beatwarp.py
素材预处理：prep_clip.sh、prep_video.py、matte.swift、fixmatte.py、matte_keyed.py、matte_fallback.py、track.py
同步检查：lipsync_check.py、dance_sync_check.py、av_sync_check.py
渲染与审核：render.mjs、still.mjs、serve.mjs、sheet.py
引擎文件（由scaffold.sh复制到项目中）：index.html、main.js、canvas.js、core.js、plate.js、roto.js、typekit.js、hud.js，以及timeline.template.js和scene.template.js
references/pipeline.md：逐步说明所有命令和Scenario调用
references/engine_api.md：场景合约和完整的
```
E
```
API（画布与安全区、底版、遮罩、跟踪、后期、HUD、变体）
references/motion_library.md：视觉风格、文字系统、40种动效技巧库、融合创意、转场、时序、缓动和弹簧效果、同步与剪辑密度
references/agent_brief_template.md：每个场景Agent阅读的指令模板
references/lessons.md：制作中的坑点和经过验证的导演默认规则