steal

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Steal

Someone's edit works and you want it on your own videos. This skill measures their edit, writes it down as a style, and proves it on one of your takes before you use it.

Output:

video-agent/styles/<name>.md

: the style, read by

/short-editing <take> style <name>

```
video-agent/styles/<name>/
```
: its stills, sound effects and stickers.

video-agent/edits/<slug>-<name>/final.mp4

: the test video.

Read references/learnings.md before you start. Each rule in it comes from a copy that got rejected.

某人的剪辑效果很棒，你想把这种风格用到自己的视频中。这个 skill 会量化对方的剪辑效果，将其记录为一种风格，并先在你的一段拍摄素材上进行验证，然后再供你使用。

输出：

video-agent/styles/<name>.md

：风格文件，由

/short-editing <take> style <name>

读取。

```
video-agent/styles/<name>/
```
：其中的静帧、音效和贴纸。

video-agent/edits/<slug>-<name>/final.mp4

：测试视频。

开始前请阅读 references/learnings.md。其中每条规则都源于一次被否决的复刻尝试。

How to run it

运行方法

Preflight:

ffmpeg -version && yt-dlp --version

. Missing:

/studio-setup tools

Talk in the creator's language. Say what you do in one line before each long step.
```
<name>
```
: the creator's handle in lowercase with dashes, unless the user names the style.
Scratch files go in
```
video-agent/.tmp/steal-<name>/
```
(
```
W
```
below).

预检：

ffmpeg -version && yt-dlp --version

。如有缺失，执行

/studio-setup tools

。

使用创作者的语言交流。每个耗时步骤开始前，用一行说明你将做什么。
```
<name>
```
：默认使用创作者账号名的小写形式并以连字符分隔，除非用户为该风格命名。
临时文件放在
```
video-agent/.tmp/steal-<name>/
```
（下文简称
```
W
```
）。

1. Get the reel

1. 获取 Reel

A local file: use it.
A YouTube link:
```
yt-dlp -o "W/reel.%(ext)s" "<url>"
```
.
Instagram or TikTok: try
```
yt-dlp
```
first. If it is blocked (login wall), and Treg is set up (
```
treg balance
```
answers),
```
treg catalog search "download an Instagram reel video file"
```
, then
```
treg catalog get <id>
```
for the price, then
```
treg call <id> --query url=<url>
```
and download the video URL it returns. Otherwise ask the user to save the video and give you the file.
When you can, study 2 or 3 reels of the same format. One reel can be the exception.

本地文件：直接使用。
YouTube 链接：
```
yt-dlp -o "W/reel.%(ext)s" "<url>"
```
。
Instagram 或 TikTok：先尝试
```
yt-dlp
```
。如果被拦截（登录墙），且已配置 Treg（
```
treg balance
```
有响应），则执行
```
treg catalog search "download an Instagram reel video file"
```
，再通过
```
treg catalog get <id>
```
查看价格，然后执行
```
treg call <id> --query url=<url>
```
并下载其返回的视频 URL。否则，请用户保存视频并提供文件。
条件允许时，研究同一格式的 2 或 3 条 Reel。只有一条时可以例外处理。

2. Measure

2. 量化

Do all three passes. A style copied from the frames alone misses half of it: the sound and the rhythm.

Faster with sub-agents. If your agent can run sub-agents (Claude Code, Codex), make the sheets and

cuts.txt

yourself (first block below), then hand Frames, Sound and Zooms to three sub-agents at the same time. Give each one the absolute paths, its section copied whole, and one file to write:

W/notes-frames.md

W/notes-sound.md

W/notes-zooms.md

, numbers and timestamps only, never playing audio aloud. Look at the sheets yourself while they work: you judge the look, they bring the numbers. Then write the style from their notes. Without sub-agents, do the passes one after the other.

完成全部三轮分析。仅从画面帧复制会漏掉一半风格：声音和节奏。

使用 sub-agent 更快。 如果你的 Agent 可以运行 sub-agent（Claude Code、Codex），则自行生成接触表和

cuts.txt

（下方第一个代码块），然后把 Frames、Sound 和 Zooms 三部分同时交给三个 sub-agent。向每个 sub-agent 提供绝对路径、完整复制的对应章节，以及一个要写入的文件：

W/notes-frames.md

、

W/notes-sound.md

、

W/notes-zooms.md

。只记录数字和时间戳，绝不外放音频。它们工作时，你自行查看接触表：你负责判断视觉风格，它们负责提供数字。然后根据这些笔记编写风格文件。如果没有 sub-agent，则依次完成各轮分析。

Frames and cuts

Frames 和 cuts

bash

ffmpeg -v error -y -i W/reel.mp4 -vf "fps=2,scale=270:480,tile=8x3:padding=4:color=black" W/sheet-%02d.jpg
ffmpeg -hide_banner -i W/reel.mp4 -vf "scdet=threshold=12" -an -f null - 2>&1 | grep -o "scd.time: [0-9.]*" | cut -d' ' -f2 > W/cuts.txt
ffprobe -v error -show_entries format=duration -of csv=p=0 W/reel.mp4

Each sheet holds 24 tiles at 2 per second. Sheet N, tile i (from 0) shows second 12(N-1) + i/2.
Seconds per shot = duration / (number of lines in
```
cuts.txt
```
+ 1).

Read EVERY sheet and write down:

Layouts: full camera, split screen (where the seam sits), flat color with text, cut-out person on a color, B-roll, screenshots.
Captions, per layout: font feel (condensed, rounded, serif), case, weight, size, where they sit on the body, how a word appears, which word gets emphasised and how.
Stickers and images: type, border, size, when they land (on the word that names them?), whether they stack.
Hook (first 2 s), CTA (last 3 s), color grade (bright and flat, or contrasty and dark).

Save 6 to 10 key stills to

video-agent/styles/<name>/

bash

ffmpeg -v error -y -ss <seconds> -i W/reel.mp4 -frames:v 1 video-agent/styles/<name>/t<seconds>.jpg

bash

ffmpeg -v error -y -i W/reel.mp4 -vf "fps=2,scale=270:480,tile=8x3:padding=4:color=black" W/sheet-%02d.jpg
ffmpeg -hide_banner -i W/reel.mp4 -vf "scdet=threshold=12" -an -f null - 2>&1 | grep -o "scd.time: [0-9.]*" | cut -d' ' -f2 > W/cuts.txt
ffprobe -v error -show_entries format=duration -of csv=p=0 W/reel.mp4

每张接触表包含 24 个图块，采样频率为每秒 2 帧。第 N 张表中从 0 开始计数的第 i 个图块对应第 12(N-1) + i/2 秒。
每个镜头的平均时长 = 总时长 /（
```
cuts.txt
```
中的行数 + 1）。

务必查看每一张接触表，并记录：

布局：全景镜头、分屏（接缝位置）、纯色背景配文字、抠像人物叠在色块上、B-roll、截图。
字幕（按布局分别记录）：字体风格（窄体、圆体、衬线体）、大小写、字重、字号、位于身体的什么位置、单词如何出现、哪个词被强调以及如何强调。
贴纸和图像：类型、边框、尺寸、出现时机（是否落在点明其名称的单词上）、是否叠加。
Hook（前 2 秒）、CTA（后 3 秒）、调色（明亮平调，或高对比暗调）。

将 6 到 10 张关键静帧保存到

video-agent/styles/<name>/

：

bash

ffmpeg -v error -y -ss <seconds> -i W/reel.mp4 -frames:v 1 video-agent/styles/<name>/t<seconds>.jpg

Sound

bash

ffmpeg -v error -y -i W/reel.mp4 -lavfi "showspectrumpic=s=3000x500:legend=0:scale=log:stop=11000" W/spectrum.png

The image is 3000 px wide, so a cut at t seconds sits at x = 3000 × t / duration. Read it next to
```
cuts.txt
```
.
A bright thin vertical stroke high up = a click. Wide and low = a boom. A rising smear = a whoosh. A dark gap across the steady music band = the beat drops out.

Zoom on one cut to see its hit:

ffmpeg -v error -y -ss <cut - 1> -t 2 -i W/reel.mp4 -lavfi "showspectrumpic=s=1000x500:legend=0:scale=log:stop=11000" W/cut-<n>.png

Beat: count the regular low strokes over 10 s on the spectrogram (or tap along while it plays): strokes per 10 s × 6 = BPM.

Music level under the voice (optional, needs

ELEVENLABS_API_KEY

video-agent/.env

, uses a few credits, say so first):

bash

set -a; . video-agent/.env; set +a
ffmpeg -v error -y -i W/reel.mp4 -vn -ac 1 -ar 44100 W/full.wav
curl -s -X POST https://api.elevenlabs.io/v1/audio-isolation -H "xi-api-key: $ELEVENLABS_API_KEY" -F audio=@W/full.wav -o W/voice.mp3
ffmpeg -v error -y -i W/voice.mp3 -ac 1 -ar 44100 W/voice.wav
for f in full voice; do ffmpeg -v error -i W/$f.wav -af "asetnsamples=882,astats=metadata=1:reset=1,ametadata=print:key=lavfi.astats.Overall.RMS_level:file=-" -f null - | grep -o "RMS_level=.*" | cut -d= -f2 > W/rms-$f.txt; done
paste W/rms-voice.txt W/rms-full.txt | awk '$1<-40{print $2}' | sort -n | awk '{a[NR]=$1}END{print "music while the voice is silent:", a[int(NR/2)+1], "dB over", NR/50, "s"}'
awk '$1>-30' W/rms-voice.txt | sort -n | awk '{a[NR]=$1}END{print "voice while talking:", a[int(NR/2)+1], "dB"}'

These commands measure the level every 20 ms. The music is the full mix in the moments the isolated voice is silent.
Music under the voice = the voice level minus the music level.
Under 1 s of silence (an edit with every pause cut) is too little to trust: judge it on the spectrogram.
Never subtract the isolated voice from the mix: it is not in phase with the original, so the result comes out louder than the voice.
Without a key, judge the level on the spectrogram.

Write down where the music starts and which event gets which sound: cut, sticker, zoom, impact.

bash

ffmpeg -v error -y -i W/reel.mp4 -lavfi "showspectrumpic=s=3000x500:legend=0:scale=log:stop=11000" W/spectrum.png

图像宽度为 3000 px，因此位于第 t 秒的剪切点对应 x = 3000 × t / total duration。将它与
```
cuts.txt
```
对照查看。
高处的一条明亮细竖线代表 click（点击声）；宽而低的竖线代表 boom（低频冲击声）；向上延伸的拖影代表 whoosh（呼啸声）；稳定音乐频带上的一道暗色间隙表示节拍在此中断。

放大某个剪切点以查看其命中音效：

ffmpeg -v error -y -ss <cut - 1> -t 2 -i W/reel.mp4 -lavfi "showspectrumpic=s=1000x500:legend=0:scale=log:stop=11000" W/cut-<n>.png

。

节拍：在频谱图上统计 10 秒内规律出现的低位纵向笔画（或播放时跟着打拍）：每 10 秒的笔画数 × 6 = BPM。

人声下的音乐电平（可选，需要在

video-agent/.env

中配置

ELEVENLABS_API_KEY

，会消耗少量 credits，开始前先告知用户）：

bash

set -a; . video-agent/.env; set +a
ffmpeg -v error -y -i W/reel.mp4 -vn -ac 1 -ar 44100 W/full.wav
curl -s -X POST https://api.elevenlabs.io/v1/audio-isolation -H "xi-api-key: $ELEVENLABS_API_KEY" -F audio=@W/full.wav -o W/voice.mp3
ffmpeg -v error -y -i W/voice.mp3 -ac 1 -ar 44100 W/voice.wav
for f in full voice; do ffmpeg -v error -i W/$f.wav -af "asetnsamples=882,astats=metadata=1:reset=1,ametadata=print:key=lavfi.astats.Overall.RMS_level:file=-" -f null - | grep -o "RMS_level=.*" | cut -d= -f2 > W/rms-$f.txt; done
paste W/rms-voice.txt W/rms-full.txt | awk '$1<-40{print $2}' | sort -n | awk '{a[NR]=$1}END{print "music while the voice is silent:", a[int(NR/2)+1], "dB over", NR/50, "s"}'
awk '$1>-30' W/rms-voice.txt | sort -n | awk '{a[NR]=$1}END{print "voice while talking:", a[int(NR/2)+1], "dB"}'

这些命令每 20 ms 测量一次电平。分离人声处于静音状态时，音乐的电平就是完整混音的电平。
人声下的音乐 = 人声电平减去音乐电平。
如果静音不足 1 秒（每次停顿都被剪掉的剪辑），则不足以据此判断：应结合频谱图评估。
绝不要从混音中减去分离人声：二者的相位与原片不同相减，结果会比人声更响。
没有 API key 时，根据频谱图判断电平。

记录音乐从哪里开始，以及每个事件对应哪种声音：剪切、贴纸、变焦、冲击。

Zooms

On the sheets, measure the face height as a share of the frame height on 10 tiles or so, and list the jumps in face size from one tile to the next:

a jump at a cut = a change of framing between two shots
a jump inside a shot = a punch-in
a slow growth over a shot = a push

Write the range (e.g. face 12 to 25 % of the frame), the punch-ins per 10 s, and whether there is a push. These numbers are the ceiling for your copy.

在接触表上选取约 10 个图块，测量脸部高度占画面高度的比例，并列出相邻图块之间的脸部尺寸跳变：

剪切点处出现跳变 = 两个镜头之间的构图发生变化
镜头内部出现跳变 = punch-in（快速推近）
镜头过程中缓慢增大 = push（缓慢推近）

记录范围（例如，脸部占画面高度的 12% 到 25%）、每 10 秒的快速推近次数，以及是否存在缓慢推近。这些数值是复刻时的上限。

3. Write the style

3. 编写风格

video-agent/styles/<name>.md

, in the user's language, with the same sections as

short-editing

's style.md. Each line gives a value and the measurement behind it:

markdown

undefined

video-agent/styles/<name>.md

应使用用户的语言，并采用与 short-editing 的 style.md 相同的章节。每一行都应给出取值及其测量依据：

markdown

undefined

Style: <name>

Copied from <link>, measured on <date>. Applied on top of short-editing's style.md: what is not here stays as there.

Cuts

hard cut every 1.4 s on average (35 cuts in 50 s); no fade, no transition

hard cut every 1.4 s on average (35 cuts in 50 s); no fade, no transition

Framing

face 12 to 25 % of the frame; punch-in 1.1 at most, about every 4 s, held until the next cut; slow push 0.6 %/s

face 12 to 25 % of the frame; punch-in 1.1 at most, about every 4 s, held until the next cut; slow push 0.6 %/s

Captions

Stickers and images

Sound

boom bap beat ~90 BPM from the hook impact, 12 dB under the voice, drops out 0.3 s on punch-ins
a dry high click on most cuts (sfx/cut.wav), a pop on each sticker (sfx/pop.wav)

boom bap beat ~90 BPM from the hook impact, 12 dB under the voice, drops out 0.3 s on punch-ins
a dry high click on most cuts (sfx/cut.wav), a pop on each sticker (sfx/pop.wav)

Grade

ffmpeg: eq=contrast=1.08:saturation=0.9,vignette=PI/5,unsharp=5:5:0.4

ffmpeg: eq=contrast=1.08:saturation=0.9,vignette=PI/5,unsharp=5:5:0.4

Left out

the AI video clips: replaced by your own footage or a callout

the AI video clips: replaced by your own footage or a callout

How short-editing applies it

the components to add to the composition and the props they take


**Sound effects** (in `video-agent/styles/<name>/sfx/`):

- With an ElevenLabs key, generate 2 or 3 takes per sound and keep the one closest to the original on the spectrogram:
  ```bash
  curl -s -X POST https://api.elevenlabs.io/v1/sound-generation -H "xi-api-key: $ELEVENLABS_API_KEY" -H "content-type: application/json" -d '{"text":"single dry short high click, tight transient, no reverb","duration_seconds":0.5}' -o video-agent/styles/<name>/sfx/cut-a.mp3

Otherwise, use a royalty-free library.

Stickers:

Generate them with
```
/codex-images
```
. Use the user's REAL photos as the face reference, never a generated face. Ask for a flat #00FF00 background, no shadow, no text.

Key the background out to a transparent PNG:

bash

ffmpeg -v error -y -i sticker.png -vf "colorkey=0x00FF00:0.3:0.08,despill=type=green" -pix_fmt rgba sticker-cut.png

Logos: every app named on screen shows its real logo, never a letter in a circle.

the components to add to the composition and the props they take


**音效**（位于 `video-agent/styles/<name>/sfx/`）：

- 配置 ElevenLabs key 后，为每个音效生成 2 或 3 个版本，并保留频谱图上最接近原声的一版：
  ```bash
  curl -s -X POST https://api.elevenlabs.io/v1/sound-generation -H "xi-api-key: $ELEVENLABS_API_KEY" -H "content-type: application/json" -d '{"text":"single dry short high click, tight transient, no reverb","duration_seconds":0.5}' -o video-agent/styles/<name>/sfx/cut-a.mp3

否则，使用免版税素材库。

贴纸：

使用
```
/codex-images
```
生成。必须使用用户的真实照片作为人脸参考，绝不能使用生成的人脸。要求使用纯色 #00FF00 背景，不要阴影，不要文字。

抠除背景并输出透明 PNG：

bash

ffmpeg -v error -y -i sticker.png -vf "colorkey=0x00FF00:0.3:0.08,despill=type=green" -pix_fmt rgba sticker-cut.png

Logo：屏幕上提到的每个 App 都必须显示其真实 Logo，绝不能只用圆圈里的字母代替。

4. Test it

4. 测试

Pick a 10 to 20 s take of the user from
```
video-agent/takes/
```
.
Run
```
/short-editing <take> style <name>
```
. It reads the style file and adds what it describes to the composition: stickers on the word that names them, sound effects on the cuts, the grade.
The slug is
```
<take slug>-<name>
```
.

从
```
video-agent/takes/
```
中选择一段 10 到 20 秒的用户拍摄素材。
运行
```
/short-editing <take> style <name>
```
。它会读取风格文件，并将其中描述的内容加入合成：在点明名称的单词上添加贴纸、在剪切点添加音效，并应用调色。
slug 为
```
<take slug>-<name>
```
。

5. Compare before showing

5. 展示前对比

Put the original on top and the copy below, at the same seconds:

bash

ffmpeg -v error -y -i W/reel.mp4 -i video-agent/edits/<slug>/final.mp4 -filter_complex "[0:v]fps=1,scale=216:384,trim=end_frame=15,tile=15x1[a];[1:v]fps=1,scale=216:384,trim=end_frame=15,tile=15x1[b];[a][b]vstack" -frames:v 1 W/compare.jpg

Go tile by tile:

type size and weight
sticker and logo size (they fill the space like in the original)
grade
no fade where the original cuts hard
captions off the mouth
nothing upside down

Run the spectrogram on the copy too: the hits must land on the cuts, the beat must sit at the same level under the voice. Measure the face size on its sheets: no more zooms than the original, the face never bigger.

With sub-agents, give the compare sheet, the key stills and the list above to a fresh-eyes sub-agent while you check the sound and the zooms: it reports every tile that differs. Fix every gap before showing. The user should never be the one to spot it.

在同一时间点将原片放在上方、复刻视频放在下方：

bash

ffmpeg -v error -y -i W/reel.mp4 -i video-agent/edits/<slug>/final.mp4 -filter_complex "[0:v]fps=1,scale=216:384,trim=end_frame=15,tile=15x1[a];[1:v]fps=1,scale=216:384,trim=end_frame=15,tile=15x1[b];[a][b]vstack" -frames:v 1 W/compare.jpg

逐图块检查：

字体大小和字重
贴纸和 Logo 的尺寸（占满空间的效果应与原片一致）
调色
原片采用硬切的位置不能出现淡入淡出
字幕不得遮挡嘴部
不得出现上下颠倒的内容

也要对复刻视频运行频谱图分析：音效命中点必须落在剪切点上，节拍在人声下的电平必须与原片一致。在复刻视频的接触表上测量脸部尺寸：变焦次数不得超过原片，脸部不能比原片更大。

使用 sub-agent 时，将对比表、关键静帧和上述检查项交给一个具备全新视角的 sub-agent，同时你检查声音和变焦；由它报告每一个存在差异的图块。展示前修复所有差异，绝不能让用户来发现问题。

6. Deliver

6. 交付

Open

final.mp4

, then tell the user in a few lines:

what was measured (seconds per shot, BPM, music level, zooms)
what was copied, and what was left out and why
where the style lives

From now on,

/short-editing <take> style <name>

uses it. Each time the user rejects something, write it in the style file. When the lesson is general, add it to references/learnings.md.

打开

final.mp4

，然后用几行文字告诉用户：

量化了哪些内容（每个镜头的时长、BPM、音乐电平、变焦）
复制了哪些内容，以及省略了哪些内容和原因
风格文件保存在哪里

从此以后，

/short-editing <take> style <name>

就会使用该风格。每次用户否决某项内容时，都将其写入风格文件。如果某条经验具有普遍性，则将其添加到 references/learnings.md。