li-transcript

Video transcript extraction + benchmarking archive

Extract the transcript from a video link, and after AI proofreading, archive it to the benchmarking blogger directory.

Workflow

Step 0: Environment preparation before first use (only needed the first time)

If this is a skill just installed from an open-source repository, you need to prepare the runtime environment first:

# 1. 系统依赖（macOS）
brew install yt-dlp ffmpeg

# 2. Python 虚拟环境 + 腾讯云 ASR SDK
python3 -m venv .claude/skills/li-transcript/scripts/.venv
.claude/skills/li-transcript/scripts/.venv/bin/pip install tencentcloud-sdk-python-asr

3. Create a .env file in the project root directory and fill in the Tencent Cloud keys (apply in the console https://console.cloud.tencent.com/cam/capi ) Format: TENCENT_SECRET_ID=your_id TENCENT_SECRET_KEY=your_key


环境准备好之后，跳到 Step 1。

### Step 1：转录视频

运行转录脚本（脚本和 venv 都在 skill 目录下，自包含）：

```bash
.claude/skills/li-transcript/scripts/.venv/bin/python3 .claude/skills/li-transcript/scripts/video2text.py "视频链接"

The script will output:

stderr: progress information + [标题] xxx(original video title)
stdout: plain text transcript with timestamps removed

Troubleshooting order when the script reports an error:

Whether the .env in the project root directory contains TENCENT_SECRET_ID and TENCENT_SECRET_KEY(the script will search upward from the current directory for up to 6 levels)
.claude/skills/li-transcript/scripts/.venv/ whether it is normal (rebuild:python3 -m venv .claude/skills/li-transcript/scripts/.venv && .claude/skills/li-transcript/scripts/.venv/bin/pip install tencentcloud-sdk-python-asr）
Whether the video link is supported by yt-dlp

Step 2: AI proofreading

Proofread the raw transcript, changing only typos and not the content:

Homophone/near-homophone correction: infer from context (e.g. "艺人公司" → "一人公司", "四动会" → "私董会", "体校" → "提效")
Proper noun correction: technical terms, product names, personal names, English words
Obvious ASR garbled text: replace with reasonable inference

Keep colloquial expressions ("啊", "嗯", "就是说"), do not rewrite sentence structure, do not beautify the style.

Proofreading is performed directly, without listing differences one by one—if users want to verify, they can compare with the original text themselves.

Step 3: Identify the author & archive

Identify the author—by priority:

The user has already told you → use it directly
There is a self-introduction in the transcript ("Hello everyone, I am XXX", "I am XXX") → extract it
If neither → ask the user

Match benchmarking—read the 05-选题研究/对标博主/短视频/ directory list:

Matched an existing blogger → archive to that directory
No match → ask the user "[name] is not yet in the benchmarking list, create it?", and if agreed, create the directory

The author and whether to create a benchmarking entry can be combined into one question to reduce back-and-forth.

Determine the title—prefer the original title obtained by yt-dlp (the [标题] line in stderr); if unavailable, summarize one from the transcript content.

Step 4: Save the file

Path：05-选题研究/对标博主/短视频/[博主名]/[YYYY年M月]/[视频标题].md

If the time directory does not exist, create it.

Format—adjust the YAML field names according to the video source platform:

---
[平台]数据:
  点赞数: 
  收藏数: 
  评论数: 
  分享数: 
  观看数: 
tags: 
---

# 标题

[视频标题]

# 封面花字



# 视频脚本

[校对后的完整逐字稿]

Platform determination rules:

URL contains xiaohongshu.com → 小红书数据
URL contains bilibili.com → B站数据
URL contains douyin.com → 抖音数据
URL contains youtube.com or youtu.be → YouTube数据
Other → 平台数据

After saving, tell me the path in one sentence.