frame-extraction

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Frame Extraction

帧提取

Follow shared release-shell rules in:

```
postplus-shared
```
release-shell rules

Use this skill when the user needs frames, not just a text analysis of the video.

This is a general-purpose extraction skill. Do not narrow it to creator-face research only.

Follow shared routing rules in:

```
postplus-shared
```
research preferences

Use

skills/40-creative/video-analysis

before or alongside this skill when shot-level understanding already exists or would materially improve frame selection.

遵循以下共享的release-shell规则：

```
postplus-shared
```
release-shell rules

当用户需要帧而非仅视频文本分析时，使用此技能。

这是一个通用提取技能，请勿仅将其局限于创作者面部研究。

遵循以下共享路由规则：

```
postplus-shared
```
research preferences

当已具备镜头层面的理解，或该理解能显著提升帧选择效果时，可在使用此技能之前或同时使用

skills/40-creative/video-analysis

。

Use For

适用场景

persona or vibe research
creator appearance reference packs
shot or structure review
product visibility checks
UI or screen-demo capture
before/after frame pairs
opening hook frame sets
ending CTA or compliance frame sets
cover-frame or first-frame candidate pulls

人物或风格研究
创作者外观参考包
镜头或结构审查
产品曝光检查
UI或屏幕演示捕获
前后帧对比组
开场钩子帧集合
结尾CTA或合规帧集合
封面帧或首帧候选提取

Trigger Signals

触发信号

Use this skill when the user asks for things like:

抽帧
截关键帧
做 contact sheet
看人物长相 / vibe
看产品怎么露出
看 UI 怎么展示
抓开头 / 结尾 / CTA 画面
从视频里挑一些可做参考的画面

Do not use this skill when the user only wants:

hook analysis
spoken-line breakdown
adaptation ideas without frame output

Those are usually better routed to

skills/40-creative/video-analysis

当用户提出以下需求时，使用此技能：

抽帧
截关键帧
制作contact sheet
查看人物长相/vibe
看产品怎么露出
看 UI 怎么展示
抓开头 / 结尾 / CTA 画面
从视频里挑一些可做参考的画面

当用户仅提出以下需求时，请勿使用此技能：

钩子分析
台词拆解
无帧输出的改编思路

这些需求通常更适合路由至

skills/40-creative/video-analysis

。

Core Principle

核心原则

Do not default to uniform frame sampling.

Choose an extraction mode that matches the task intent.

The same video may need different frames depending on whether the user is studying:

people
products
UI
pacing
style
claims or CTAs

请勿默认使用均匀帧采样。

选择与任务意图匹配的提取模式。

根据用户研究的内容不同，同一视频可能需要提取不同的帧：

人物
产品
UI
节奏
风格
声明或CTA

Extraction Modes

提取模式

Use the smallest mode that matches the request.

See:

```
references/extraction-modes.md
```
```
references/output-contract.md
```

Default modes:

```
uniform-sample
```
```
scene-change
```
```
face-priority
```
```
object-priority
```
```
text-ui-priority
```
```
hook-first
```
```
cta-last
```
```
before-after-pair
```
```
style-board
```

使用与请求匹配的最小粒度模式。

参考：

```
references/extraction-modes.md
```
```
references/output-contract.md
```

默认模式：

```
uniform-sample
```
```
scene-change
```
```
face-priority
```
```
object-priority
```
```
text-ui-priority
```
```
hook-first
```
```
cta-last
```
```
before-after-pair
```
```
style-board
```

Input Types

输入类型

This skill should work with:

one local video file
a local folder of videos
a manifest that maps source ids to local files
a shortlist already linked to source metadata

If the user gives TikTok URLs but local video files are missing, first recover or download the local videos before extracting frames.

Do not lock this skill to one platform.

此技能支持以下输入类型：

单个本地视频文件
本地视频文件夹
映射源ID到本地文件的清单文件
已关联源元数据的候选清单

如果用户提供TikTok链接但缺少本地视频文件，需先恢复或下载本地视频，再进行帧提取。

请勿将此技能限制于单一平台。

Output Shapes

输出形式

Pick outputs based on the use case.

Common outputs:

selected frame folder
contact sheet
frame manifest
markdown summary
side-by-side frame comparison

Every extraction run should preserve enough metadata to trace each frame back to:

source video
timestamp
extraction mode
selection reason

根据使用场景选择输出形式。

常见输出形式：

选定帧文件夹
contact sheet
帧清单
Markdown摘要
帧并排对比

每次提取运行都应保留足够的元数据，以便追溯每帧的来源：

源视频
timestamp
提取模式
选择理由

Workflow

工作流程

1. Clarify the intent

1. 明确意图

Classify the ask into one of these buckets:

persona / vibe
shot / structure
product visibility
UI / text readability
before / after
hook
CTA / compliance
broad visual scan

If the request is ambiguous, ask one short question:

你这次抽帧主要是为了看人物、人设 vibe，还是看产品 / UI / 镜头结构？

将需求归类为以下类别之一：

人物/风格
镜头/结构
产品曝光
UI/文本可读性
前后对比
hook
CTA / compliance
全面视觉扫描

如果需求不明确，可提出一个简短问题：

你这次抽帧主要是为了看人物、人设 vibe，还是看产品 / UI / 镜头结构？

2. Select mode

2. 选择模式

Map the intent to one primary mode.

Good defaults:

persona / vibe ->
```
face-priority
```
shot / structure ->
```
scene-change
```
product visibility ->
```
object-priority
```
UI / text ->
```
text-ui-priority
```
hook review ->
```
hook-first
```
CTA / compliance ->
```
cta-last
```
broad scan ->
```
uniform-sample
```

Use a secondary mode only if the first one clearly misses the target.

将意图映射至一种主模式。

推荐默认配置：

persona / vibe ->
```
face-priority
```
shot / structure ->
```
scene-change
```
product visibility ->
```
object-priority
```
UI / text ->
```
text-ui-priority
```
hook review ->
```
hook-first
```
CTA / compliance ->
```
cta-last
```
broad scan ->
```
uniform-sample
```

仅当主模式明显无法满足需求时，才使用次要模式。

3. Choose scope

3. 选择范围

Do not always scan the full video.

Typical scope choices:

first 3-5 seconds
full video
final 3-5 seconds
manually specified timestamp range

请勿总是扫描完整视频。

典型范围选择：

前3-5秒
完整视频
最后3-5秒
手动指定的时间戳范围

4. Package outputs

4. 打包输出

Match the packaging to the downstream task:

persona research -> best frames + contact sheet + vibe notes
shot review -> scene-change frames + timestamp list
product review -> product-visible frames + proof notes
UI study -> readable UI frames + OCR or text notes if needed

根据下游任务匹配打包方式：

人物研究 -> 最优帧 + contact sheet + 风格说明
镜头审查 -> 场景变化帧 + 时间戳列表
产品审查 -> 产品曝光帧 + 验证说明
UI研究 -> 清晰可读的UI帧 + 必要时附带OCR或文本说明

Release-Shell Execution Contract

Release-Shell执行约定

keep extraction requests, frame manifests, contact-sheet build inputs, and other intermediate state under
```
<work-folder>/.postplus/frame-extraction/
```
keep only final user-facing frame exports, contact sheets, or review packs outside
```
.postplus/
```
use small, task-shaped extraction scopes before broad full-video pulls
if
```
ffmpeg
```
is missing, the user's agent should proactively install it with the host package manager already present on the machine before continuing
rerun a direct check such as
```
ffmpeg -version
```
after installation
if installation or verification fails, stop immediately instead of switching to ad hoc shell glue

将提取请求、帧清单、contact sheet构建输入及其他中间状态存储在
```
<work-folder>/.postplus/frame-extraction/
```
目录下
仅将面向用户的最终帧导出文件、contact sheet或审查包存储在
```
.postplus/
```
目录之外
在进行全面完整视频提取之前，先使用小型、任务特定的提取范围
如果缺少
```
ffmpeg
```
，用户的Agent应使用机器上已有的主机包管理器主动安装它，然后再继续操作
安装后重新运行直接检查，例如
```
ffmpeg -version
```
如果安装或验证失败，立即停止操作，而非切换到临时shell脚本

Default Sequence

默认流程

For TikTok benchmark work:

use platform research to shortlist videos first
use local video files when available
use
```
video-analysis
```
shot outputs if they already exist
only then extract frames

When shot-level outputs exist, prefer those as stronger evidence than blind timestamp sampling.

针对TikTok基准测试工作：

先通过平台研究筛选出候选视频
如有可用本地视频文件则优先使用
如果已存在
```
video-analysis
```
的镜头输出，则使用该输出
最后再提取帧

当存在镜头层面的输出时，优先使用该输出作为证据，而非盲目时间戳采样。

First-Version Boundary

第一版边界

The first version of this skill should stay pragmatic.

Prefer:

ffmpeg-based extraction
scene-change driven sampling
simple timestamp filtering
contact sheet generation
lightweight manifests

Do not require:

identity recognition
demographic inference
heavy CV pipelines
perfect semantic understanding of every frame

此技能的第一版应保持实用性。

优先选择：

基于ffmpeg的提取
场景变化驱动的采样
简单时间戳过滤
contact sheet生成
轻量级清单

无需实现：

身份识别
人口统计推断
复杂计算机视觉流水线
对每帧的完美语义理解

Keep These Assets

保留以下资产

For reusable benchmark work, keep:

the local source video
extracted frames
contact sheets
frame manifest
any markdown summary used in later persona or creative work

Do not treat extracted frames as disposable if they support future persona, visual-style, or benchmark decisions.

对于可复用的基准测试工作，保留以下内容：

本地源视频
提取的帧
contact sheets
帧清单
后续人物研究或创意工作中使用的所有Markdown摘要

如果提取的帧能为未来的人物研究、视觉风格或基准测试决策提供支持，请勿将其视为一次性资源。