← Back to Skills Marketplace
sgw2285265753-design

红果短剧审片

by sgw2285265753-design · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
30
Downloads
1
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install video-review
Description
视频基础质量审核技能,对视频进行批量或单条的基础问题检测,包括画面质量、构图、音频、黑边、分辨率等,输出审核报告。
README (SKILL.md)

视频初审 Skill

本技能对视频进行基础质量审核内容安全检测,输出标准化的检测报告。每次执行请严格遵循以下流程。


执行流程总览

步骤 内容 工具
1 视频信息提取 + 基础质量检测 ffprobe
2 抽帧(1fps,全部帧) ffmpeg
3 视觉安全检测(并行批量) image 工具
4 字幕转录 + 违禁词检测 haoone-cli + 正则
5 汇总生成完整报告

第一步:视频信息提取 + 基础质量检测

ffprobe -v error \
  -show_entries format=duration,size \
  -show_entries stream=width,height,codec_name,r_frame_rate \
  -of json "\x3C视频路径>"

基础质量检测项目:

检测项 合格标准 超标处理
分辨率/比例 1080×1920(9:16竖屏) 标记⚠️
帧率 统一(24/25/30fps) 标记⚠️
码率 8-15Mbps 标记⚠️
黑边 不存在 letterbox/pillarbox width/height 比值在0.5-0.6为正常
跳帧/黑帧 无明显卡顿 逐帧像素对比,超80%相似则无跳帧
人声响度 各段音量差异\x3C6dB 差异>10dB标记⚠️

跳帧检测脚本:

python3 ~/.openclaw/skills/video-review/scripts/jump_frame_detector.py \
  -i "\x3C视频路径>" --threshold 0.80

第二步:抽帧

mkdir -p /tmp/audit_frames
ffmpeg -y -i "\x3C视频路径>" -vf "fps=1" -q:v "2" /tmp/audit_frames/frame_%04d.png
  • 采样频率:每秒1帧(1fps)
  • 帧位置对应:frame_0001 = 第0秒,frame_0002 = 第1秒,frame_00N = 第N-1秒
  • 输出目录:/tmp/audit_frames/
  • 总帧数 ≈ 视频时长(秒)

第三步:视觉安全检测(并行批量)

使用 image 工具分批分析,每批10帧,同时并行提交所有批次(最多6批/轮)。

时间戳换算公式:

frame_NNNN = 第 (NNNN-1) // 60 分 (NNNN-1) % 60 秒

示例:

  • frame_0046 → 第 0 分 45 秒
  • frame_0091 → 第 1 分 30 秒
  • frame_0106 → 第 1 分 45 秒

每批 prompt(严格使用):

Review these 10 video frames for content safety compliance. 
OUTPUT ONE LINE PER FRAME. Format: frame_XXXX | [PASS or ISSUE_LETTER] | [SEVERE/MINOR/NONE]

Time mapping: frame_XXXX = (N-1) seconds (e.g. frame_0031 = 30s, frame_0046 = 45s)

Issues: A=revealing clothes, B=gov uniform, C=national emblem, D=money, E=blood/injury, F=dark/shocking, G=superstition, H=bad behavior

Safety work uniforms OK. Only mark police/military/gov uniform as B. Common work helmets and cleaning uniforms are NOT violations.

Output all 10 lines then end.

问题等级标准:

  • 严重(SEVERE):正面大面积暴露、警察/军人制服特写、人民币完整图案、血腥暴力特写
  • 轻微(MINOR):边缘暴露、暗示性穿着、仿真枪贴纸、轻微伤痕

并发策略: 同时最多提交6批(60帧),队列式接力,不等待第一批完成再提交下一批。


第四步:字幕检测

# 提取WAV
ffmpeg -y -i "\x3C视频路径>" -ar 16000 -ac 1 -acodec pcm_s16le /tmp/audit.wav

# 转录
haoone-cli transcribe /tmp/audit.wav

违禁词检测(比对 refs/banned_words.json):

类别 示例
脏话粗口 他妈的、傻逼、去死、装逼、草泥马、卧槽
极限词 最高、最低、第一、顶级、极品、绝对、全网最低
敏感国家名 中国、美国、英国、日本、俄罗斯、台湾

第五步:汇总生成完整报告

报告格式规范(严格遵循):

━━━ 视频初审检测报告 ━━━
文件:
时长:\x3CX>秒 | 分辨率:\x3C宽>x\x3C高> | 帧率:\x3Cfps>
通过:✅ / ⚠️ / ❌
问题数量:\x3CN>

━━━ 视觉问题 ━━━

【A类 · 低俗暴露】

| 时间戳 | 第几分第几秒 | 问题描述 | 严重程度 |
|--------|------------|---------|---------|
| frame_XXXX | 00:0X:XX-XX | \x3C问题描述> | **严重** |
| frame_XXXX | 00:0X:XX-XX | \x3C问题描述> | 轻微 |

━━━ 字幕/台词问题 ━━━
  无违禁词检出

━━━ 总结 ━━━
综合判定:⚠️ 需处理后通过
严重问题:\x3CN>帧(A类低俗暴露)
轻微问题:\x3CN>帧

【问题分布】
  00:0X:XX-XX(\x3C秒数>秒区间)━ A类约\x3CN>帧

━━━ 操作建议 ━━━
\x3C秒数>秒区间存在多帧女性暴露内容,建议:
  ① 剪辑删除该区间(\x3C起始秒>-\x3C结束秒>秒)
  ② 或局部遮挡+画面替换

详细帧位置已标记,可对应原视频时间码精确定位剪辑点。

格式规范说明:

  • 通过:行:✅(通过)/ ⚠️(需处理后通过)/ ❌(不通过)
  • 表格使用 ┌─┬─┐ 框线字符,字段:时间戳 | 第几分第几秒 | 问题描述 | 严重程度
  • 严重程度为"严重"的行,文字加粗处理
  • 第几分第几秒格式:00:0X:XX(分:秒),区间范围 00:0X:XX-XX
  • 问题描述:简洁明确,说明违规内容和视觉特征
  • 问题分布:标注具体时间区间和帧数
  • 操作建议:给出具体处理方案(剪辑删除/局部遮挡)
  • 报告末尾不加结尾符号,直接结束

检测维度(完整)

检测维度(完整)

A. 基础质量检测

检测项 说明
分辨率/比例 9:16竖屏(1080x1920等)
帧率 统一(24/25/30fps)
码率 8-15Mbps
黑边 不存在 letterbox/pillarbox
音频 音轨正常
跳帧/黑帧 无明显卡顿花屏
人声响度 各段音量差异\x3C6dB

B. 内容安全/违规检测

类别 说明
A·低俗暴露 深V、低胸、透视、过于紧身、性暗示动作
B·公职人员 警察/军人制服、警徽/军徽、法官/检察官、国徽/法槌/法院标识
C·机关标志 国徽、警局/公安局标识、检察院、政府楼外观
D·人民币 人民币纸币/硬币完整图案(含变形涂鸦)
E·血腥暴力 割腕/自残血痕(中景及以上特写)、虐待、家暴殴打特写
F·阴暗画风 整体色调极度阴暗压抑、灵异惊悚氛围
G·封建迷信 现实题材中算命/看相/做法事/巫蛊/香灰
H·未成年保护 未成年人吸烟/饮酒/纹身、校园霸凌、早恋
脏话粗口 字幕/台词中粗鄙辱骂性语言
极限词 含"最"字的夸大宣传语
敏感国家名 中国、美国、英国、日本、俄罗斯等(真实背景设定语境)

违禁词库

文件:refs/banned_words.json


常用脚本

单条视频审核

python3 ~/.openclaw/skills/video-review/scripts/content_audit.py \x3C视频路径>

跳帧检测

python3 ~/.openclaw/skills/video-review/scripts/jump_frame_detector.py \
  -i \x3C视频路径> --threshold 0.80 --output-json /tmp/jump_report.json

批量审核文件夹

for f in /path/to/folder/*.mp4; do
  echo "▶ $f"
  python3 ~/.openclaw/skills/video-review/scripts/content_audit.py "$f"
done

注意事项

  1. 抽帧频率:统一使用1fps,不用场景感知抽帧
  2. 并行检测:视觉检测必须并行分批,不逐个等待
  3. 帧号换算:frame_NNNN = 第 (NNNN-1)//60(NNNN-1)%60
  4. 假阳性:AI 视觉判断有误差,报告仅供参考,最终由人判断
  5. 审核尺度:以《动画微短剧内容创作建议》和《剧目修改意见统计汇总》为基准
Usage Guidance
Install only if you want uploaded or local videos to be checked for both technical quality and content-policy issues, including extracted frames and possible speech transcription. Avoid using it on confidential videos unless you are comfortable with OpenClaw image analysis and the referenced transcription tool processing derived media.
Capability Assessment
Purpose & Capability
The full SKILL.md and changelog disclose quality review, visual safety review, transcription, and banned-word checks, but the manifest/frontmatter description frames the skill mainly as basic video quality review.
Instruction Scope
Generic triggers such as video review or video detection can route broad user requests into a workflow that analyzes visual policy categories and spoken content without an explicit confirmation step.
Install Mechanism
No package installer, startup hook, background service, or privilege-changing install mechanism was found.
Credentials
The workflow extracts frames and audio, invokes OpenClaw image analysis, and may invoke a hard-coded external transcription CLI; those are plausible for video auditing but sensitive enough to require clearer disclosure and user control.
Persistence & Privilege
No persistence or privilege escalation was found, but temporary media artifacts are written under /tmp or beside the source video and cleanup is not guaranteed on all failure paths.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install video-review
  3. After installation, invoke the skill by name or use /video-review
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
video-review v1.0.0 - Initial release providing automated video quality and content safety review. - Supports batch or single video checks for resolution, frame rate, bit rate, black edges, frame drops, audio level, and visual content issues. - Automated frame extraction (1fps), parallel visual safety checks, and subtitle transcription with prohibited word detection. - Standardized, detailed audit report output with pinpointed issue timestamps and actionable edit advice. - Designed for consistent, efficient first-pass video auditing based on industry content standards.
Metadata
Slug video-review
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is 红果短剧审片?

视频基础质量审核技能,对视频进行批量或单条的基础问题检测,包括画面质量、构图、音频、黑边、分辨率等,输出审核报告。 It is an AI Agent Skill for Claude Code / OpenClaw, with 30 downloads so far.

How do I install 红果短剧审片?

Run "/install video-review" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is 红果短剧审片 free?

Yes, 红果短剧审片 is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does 红果短剧审片 support?

红果短剧审片 is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created 红果短剧审片?

It is built and maintained by sgw2285265753-design (@sgw2285265753-design); the current version is v1.0.0.

💬 Comments