home / skills / theonelee / theone_claude_skill / vlog_workflow

vlog_workflow skill

safe

/vlog_workflow

This skill intelligently trims vlog recordings by removing duplicates and missteps while preserving natural pacing and transitions.

This is most likely a fork of the video_edit skill from theonelee

npx playbooks add skill theonelee/theone_claude_skill --skill vlog_workflow

Review the files below or copy the command above to add this skill to your agents.

Files (12)

SKILL.md

4.3 KB

---
name: vlog_video_edit
description: 对录制的视频进行智能剪辑，去除重复和说错的内容，保持通顺自然
---

# Vlog Video Edit Skill

对用户录制的 vlog 视频进行智能剪辑：通过语音识别对比口播稿，自动标记需要剪辑的片段，经用户确认后执行精确剪切。

## 输入

- 原始视频文件路径（`raw.mp4`，支持 MP4/MOV 等常见格式）
- 口播稿文件路径（`speech.md`，用于对比参考）
- 特殊剪辑要求（可选）

## 前置条件

```bash
# 确保 speaches Docker 服务已启动
curl http://localhost:8000/v1/models
# 确保 ffmpeg 已安装
ffmpeg -version
```

## 核心流程

### Step 1: 语音识别（转录）

从视频中提取音频，调用 speaches API 进行语音识别：

```bash
python3 scripts/transcribe.py \
  --input ~/vlog_projects/{project}/raw.mp4 \
  --output ~/vlog_projects/{project}/transcript.json \
  --api-url http://localhost:8000
```

输出 `transcript.json` 包含带时间戳的逐段文字。

### Step 2: AI 智能分析（由 Agent 执行）

Agent 读取 `transcript.json` 和 `speech.md`，进行对比分析：

1. **对齐口播稿和实际录音**：找出两者的对应关系
2. **标记需剪辑的片段**：
   - 🔴 **重复内容**：同一段内容说了多次，保留最好的一次
   - 🔴 **说错/卡壳**：明显的口误或停顿过长
   - 🟡 **多余口头禅**：过于频繁的"嗯"、"那个"
3. **保留自然过渡**：
   - ✅ 保留必要的语气词（适量的"嗯"、"啊"、"这个"）
   - ✅ 保留章节间的自然过渡和停顿
   - ✅ 保留情感表达（笑声、感叹等）

### Step 3: 生成剪辑方案

Agent 生成 `cut_plan.json`，格式如下：

```json
{
  "keep_segments": [
    {"start": 0.0, "end": 45.2, "note": "开场白，表现自然"},
    {"start": 48.5, "end": 120.3, "note": "第一部分正文"},
    {"start": 125.0, "end": 180.0, "note": "第二部分，跳过了 120.3-125.0 的卡壳"}
  ],
  "removed_segments": [
    {"start": 45.2, "end": 48.5, "reason": "重复了开场白最后一句"},
    {"start": 120.3, "end": 125.0, "reason": "卡壳停顿 5 秒"}
  ]
}
```

### Step 4: 用户确认

向用户展示剪辑方案：
- 列出所有要删除的片段及原因
- 告知预计剪辑后时长
- 等待用户确认或调整

### Step 5: 执行剪辑

```bash
python3 scripts/cut_video.py \
  --input ~/vlog_projects/{project}/raw.mp4 \
  --plan ~/vlog_projects/{project}/cut_plan.json \
  --output ~/vlog_projects/{project}/edited.mp4
```

## 剪辑原则

### 必须剪掉

- 完全重复的段落（保留表现最好的一次）
- 明显的口误后重新开始的部分
- 超过 3 秒的无意义停顿

### 谨慎保留

- 短暂的思考停顿（1-2 秒）→ 保留，显得自然
- 偶尔的"嗯"、"那个" → 保留，避免机器感
- 语气加重、情感表达 → 一定保留
- 章节过渡处的自然停顿 → 保留

### 剪辑技巧

- **剪在语句间隙**：在句子的自然停顿处剪切，而非句中
- **保留呼吸音**：不要把呼吸声也剪掉，会显得不自然
- **前后 buffer**：每个保留片段前后各留 0.1-0.3 秒缓冲

## 输出

- 剪辑后的视频：`~/vlog_projects/{project}/edited.mp4`
- 剪辑报告：删除了哪些部分、原因、前后时长对比

Overview

This skill performs intelligent editing of recorded vlog footage to remove repeated lines, obvious mistakes, and long pauses while preserving a natural flow. It compares the spoken transcript with the prepared script, marks candidate cuts, and produces a precise cut plan for user review. After confirmation, it executes frame-accurate edits and outputs an edited video plus a cut report.

How this skill works

The skill extracts audio, transcribes speech with timestamps, and aligns the transcript to the reference script to detect repeats, errors, and filler words. It generates a detailed cut_plan.json with keep and removed segments, including short buffers around cuts to maintain natural transitions. The plan is presented to the user for confirmation or adjustment, then ffmpeg-based tools perform the exact trimming to produce the final edited file.

When to use it

You recorded a vlog with multiple takes and want the best performance kept.
You want to remove long pauses, stutters, or obvious mistakes without losing natural rhythm.
You have a prepared script and want the recording aligned and cleaned automatically.
You need a fast, deterministic cut plan for batch editing several episodes.

Best practices

Provide a clear reference script to improve alignment accuracy.
Run the local speech recognition service and verify transcript.json before analysis.
Review and confirm the generated cut_plan.json; automatic suggestions are conservative.
Keep nonverbal cues (laughs, emphasis) marked as keep to preserve authenticity.
Use small front/back buffers (0.1–0.3s) to avoid abrupt audio jumps.

Example use cases

Trim a 20-minute raw vlog to a polished 12-minute episode by removing repeats and long pauses.
Batch-process multiple recorded segments from the same shoot to unify pacing and remove false starts.
Prepare a talk-through where the spoken recording diverged from the script and you need aligned, minimal edits.
Generate a cut report for collaborators showing exactly which segments were removed and why.

FAQ

What input files are required?

A raw video file (MP4/MOV) and an optional reference script file; the transcript is generated from the video audio.

Can I adjust the suggested cuts?

Yes. The skill outputs a cut_plan.json for user review and adjustment before any destructive edits.