一个 Claude Code 斜杠命令，自动完成口播/教程类视频的 A-roll 粗剪。
丢一个素材文件夹进去 → 自动转写、识别好坏 take、去掉静音 → 在达芬奇里生成可直接编辑的 时间线 + SRT 字幕。
原始素材文件夹
↓ FFmpeg 提取音频
↓ Whisper 语音转写（逐字时间戳）
↓ 静音检测
↓ LLM 分析：保留/丢弃 take，去重复段落
↓ 生成达芬奇 Resolve Python 脚本
↓ 生成 SRT 字幕（对齐剪辑后时间线）
↓ 执行 → 达芬奇里直接出现时间线
核心逻辑：很多创作者录制时会在满意的 take 后说 "OK"，不满意时说 "pass"/"重来"。这个工具能识别这些口头标记，自动只保留最佳 take。如果你录制时不说标记词也没关系——它会保留所有语音段落，只切掉静音。
- Claude Code（CLI 或 IDE 插件）
- FFmpeg — 音频提取和静音检测
- OpenAI Whisper — 语音转写（
pip install openai-whisper） - DaVinci Resolve（免费版或 Studio 版）— 执行脚本时需要打开
把 aroll.md 复制到你项目的 .claude/commands/ 目录：
mkdir -p .claude/commands
cp aroll.md .claude/commands/搞定，/aroll 命令就可以用了。
在 Claude Code 中输入：
/aroll ~/Videos/产品评测第12期 landscape
或者直接输入 /aroll，助手会主动问你：
- 素材文件夹路径
- 横屏还是竖屏
- 项目名称（可选）
- 帧率（可选，默认 23.976）
- 语言（可选，默认自动检测）
助手会依次：
- 扫描并列出素材文件
- 用 Whisper 转写全部音频
- 分析剪辑点，展示决策列表
- 等你确认后才执行
- 生成达芬奇时间线 + SRT 字幕
| 录制时说的话 | 结果 |
|---|---|
| "OK" / "好" / "keep" / "good" | 前面那段 保留 |
| "pass" / "不要" / "cut" / "again" / "redo" | 前面那段 丢弃 |
| （什么都没说，一直在讲） | 保留所有语音段，只切静音 |
同一句话录了多遍？只保留最后一个好 take（离 keep 标记最近的那个）。
直接用自然语言告诉助手就行：
- "用 29.97 帧率和 1080p"
- "我的保留标记是 'nice'，丢弃标记是 'nope'"
- "静音阈值 -40dB，我这环境比较吵"
- "用 whisper large 模型，精度高一些"
完整配置表见 aroll.md。
- macOS：开箱即用，默认路径适配标准达芬奇安装
- Windows / Linux：生成的脚本中有注释，标明了需要修改路径的位置
A Claude Code slash command that automates A-roll rough cuts for talking-head and voiceover videos.
Give it a folder of clips → it transcribes, detects good/bad takes, removes silence, and builds a ready-to-edit DaVinci Resolve timeline with synced SRT subtitles.
Raw footage folder
↓ FFmpeg extract audio
↓ Whisper transcribe (word-level timestamps)
↓ Silence detection
↓ LLM analyzes: keep/drop takes, remove repeats
↓ Generates DaVinci Resolve Python script
↓ Generates SRT subtitles (aligned to edited timeline)
↓ Executes → timeline ready in DaVinci Resolve
The key insight: many creators say "OK" after a good take and "pass" / "again" after a bad one. This skill understands those spoken markers and automatically keeps only the best takes. If you don't use markers, it still works — it'll keep all speech segments and just cut the silence.
- Claude Code (CLI or IDE extension)
- FFmpeg — for audio extraction and silence detection
- OpenAI Whisper — for transcription (
pip install openai-whisper) - DaVinci Resolve (Free or Studio) — must be running when the script executes
Copy aroll.md into your project's .claude/commands/ directory:
mkdir -p .claude/commands
cp aroll.md .claude/commands/That's it. The /aroll command is now available in Claude Code.
In Claude Code, type:
/aroll ~/Videos/product-review-ep12 landscape
Or just type /aroll and the assistant will ask you for:
- Footage folder path
- Landscape or portrait
- Project name (optional)
- Frame rate (optional, default 23.976)
- Language (optional, default auto-detect)
The assistant will:
- Scan and list your footage files
- Transcribe everything with Whisper
- Analyze edit points and show you the decision list
- Wait for your confirmation
- Generate and run the DaVinci Resolve script
- Output an SRT subtitle file
| You say during recording | What happens |
|---|---|
| "OK" / "好" / "keep" / "good" | Previous segment is kept |
| "pass" / "不要" / "cut" / "again" / "redo" | Previous segment is dropped |
| (nothing — just keep talking) | All speech segments are kept, silence is cut |
When you record multiple takes of the same line, the skill keeps the last good take (closest to a keep marker).
You can adjust defaults by telling the assistant in natural language:
- "Use 29.97 fps and 1080p"
- "My keep marker is 'nice', drop marker is 'nope'"
- "Silence threshold -40dB, I'm in a noisy environment"
- "Use whisper large model for better accuracy"
See the full configuration table in aroll.md.
- macOS: Works out of the box with standard DaVinci Resolve installation paths
- Windows/Linux: The generated script includes comments showing where to adjust DaVinci Resolve paths
MIT