TLDR.skill

Turn high-density social media videos into readable knowledge.

Capture the transcript. Clean it. Summarize it. Reality-check it.

Typical moment: A friend, teacher, or coworker sends you a 10-minute video packed with ideas. You do not have time to watch it now — but you still need the substance.

TLDR.skill gives you the whole chain:

social video link -> transcript -> cleaned transcript -> summary -> Reality Check

---

TL;DR

When a friend, teacher, or coworker sends you a long, high-density video from social media, the problem is usually not "can I open the link?"

The real problem is:

• you don't have time to watch it now
• the video may contain useful ideas buried inside dense speech
• you still want to know what it says, what matters, and what to trust

TLDR.skill solves that full value chain:

social media video link -> transcript capture -> transcript cleanup -> summary -> Reality Check

That is the core value:

1. get the video into text from Xiaohongshu / Douyin / YouTube
2. clean the transcript so it becomes readable
3. summarize the content so you get the key takeaway fast
4. add a Reality Check so you know what sounds solid vs what needs caution

So the output is not just more readable — it is more usable.

---

What it does

Input: a video URL from

• Xiaohongshu / 小红书
• Douyin / 抖音
• YouTube

Output:

1. Summary — the distilled takeaway
2. Reality Check — what seems credible vs what needs caution
3. Cleaned Transcript — punctuated, paragraphized, easier to read than raw ASR

---

Why this is different

Traditional pipeline

video -> raw transcript -> summary

TLDR.skill pipeline

video -> raw transcript -> cleaned transcript -> summary + reality check

That extra cleanup stage matters a lot in practice, especially for:

• mixed Chinese / English terminology
• ASR mistakes like token / AIDC / Agent
• creator videos with poor punctuation after transcription
• short-form video content that is fast, noisy, and context-heavy

---

Best use cases

1. Someone sends you a long, dense video and you do not have time to watch it

This is the most common case:

• a friend sends you a creator video
• a teacher shares a lecture clip
• a coworker drops a social-media link in chat

The video may be valuable, but it is long, dense, and inconvenient to watch right now.

What you really need is:

• the transcript
• the cleaned version
• the summary
• the Reality Check

2. Research faster

You see a video about AI, startups, monetization, or distribution and want the point without spending 10 minutes watching it.

3. Share better inside teams

Instead of sending “watch this,” send:

• summary
• reality check
• transcript

4. Build a searchable knowledge base

Turn noisy video links into Markdown you can archive, grep, reuse, and export.

---

Supported platform strategy

YouTube

1. official / existing captions via youtube-transcript-api
2. yt-dlp auto subtitles
3. media download + ASR fallback

Douyin

1. Playwright + local Chrome
2. capture the real audio request
3. download audio
4. run STT

Xiaohongshu

1. yt-dlp download
2. STT

---

Example output shape

# 视频转录 Digest - <title>

- 平台：<platform>
- 链接：<url>
- 转录来源：<provider>

## 总结
### 一句话结论
...

### 核心要点
- ...

### 关键信号
- ...

## Reality Check
### 核心判断
...

### 哪些点相对可信
- ...

### 哪些点需要谨慎
- ...

### 最终结论
...

## 转录稿
...

---

Quick start

1. Install

python3.11 -m venv venv
source venv/bin/activate
pip install -e ".[video]"

2. Run

tldr-skill "https://www.youtube.com/watch?v=Mfzucn4f9Xk" --output /tmp/video_digest.md

3. JSON mode

tldr-skill "https://v.douyin.com/xxxx/" --format json

---

Repo structure

src/tldr_skill/
  cli.py
  llm.py
  transcription.py
  video_digest.py
tests/
  test_video_digest.py
skills/
  video-link-transcript-digest/SKILL.md

This repo is now intentionally lean. It contains the actual TLDR.skill implementation, not a full Hermes snapshot.

---

Environment notes

If a target platform needs browser cookies:

export HERMES_YTDLP_COOKIES_FROM_BROWSER=chrome
export HERMES_YTDLP_BROWSER_HOME=/Users/your-username

If you want to override the browser path for Douyin capture:

export TLDR_SKILL_CHROME_PATH="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"

If you want to override the LLM endpoint / model:

export OPENAI_API_KEY=...
export TLDR_SKILL_MODEL=gpt-4.1-mini
# optional
export TLDR_SKILL_BASE_URL=...

---

Test

pytest -q

---

Hermes trigger phrase

Inside Hermes, the intended trigger is:

• 转录 <url>
• 转录这个视频 <url>
• 总结这个视频 <url>

---

License

MIT.