Usage

Auto-trigger: When a YouTube URL is detected in the user's message, run this skill automatically.

Manual trigger:

/youtube-summary https://youtube.com/watch?v=xxx
/youtube-summary https://www.youtube.com/shorts/xxx [optional title]

YouTube Summary + Gist Upload

Downloads a YouTube video, transcribes Taiwanese Hokkien with Breeze ASR 26 and all other languages with Whisper Turbo through the Breeze streaming project's Transformers + ROCm pipeline, then uploads the raw transcript and a timestamped section breakdown to a GitHub Gist.

No external LLM calls are made — the summary is assembled directly from the transcript segments.

The non-Taiwanese path must not call the standalone openai-whisper CLI: on the Strix Halo/gfx1151 host that path can select an incompatible ROCm/PyTorch stack and crash with invalid device function or a segmentation fault.

Workflow

Step 1: Extract YouTube Link

Supports:

• https://youtube.com/watch?v=...
• https://www.youtube.com/shorts/...
• https://youtu.be/...

Step 2: Download + Transcribe

Choose the ASR before running: use --asr breeze only when the video is Taiwanese Hokkien; otherwise leave the default Whisper Turbo.

# Default: Breeze-compatible Whisper Turbo for non-Taiwanese audio
python3 ~/.pi/agent/skills/media/youtube-summary/scripts/youtube-summary.py \
  "https://www.youtube.com/shorts/VIDEO_ID" \
  "Video Title"

# Taiwanese Hokkien: Breeze ASR 26
python3 ~/.pi/agent/skills/media/youtube-summary/scripts/youtube-summary.py \
  --asr breeze "https://www.youtube.com/shorts/VIDEO_ID" \
  "台語影片"

Output:

• Transcript printed to stdout (raw, continuous text)
• Transcript segments with timestamps (JSON format)
• Summary prompt for Qwen with section-by-section breakdown instructions

Step 3: Assemble Summary from Transcript

Build the summary directly from the transcript segments (no external model calls):

1. 完整摘要 — 3-5 key points extracted from the transcript text
2. 逐段內容拆解 — Section-by-section breakdown using timestamped segments from the transcript
3. 重點數據／事實 — Key data points extracted from the transcript
4. 結論／感想 — Conclusion from the transcript
5. 【全內容】最後一段的完整內容 — If the last segment is an interview, dialogue, Q&A, tutorial, code walkthrough, or any structured content type, reproduce the full verbatim content in this final section.

Step 4: Upload to Gist

The gist includes two files:

• summary.md — Structured summary built directly from the transcript (no external LLM)
• transcript.txt — Raw, unedited transcript (完整原始逐字稿)

Script Reference

youtube-summary.py

Full pipeline (download → extract → transcribe → output):

For non-Taiwanese audio, the script launches its hidden --whisper-worker using the Breeze streaming ROCm venv and WhisperForConditionalGeneration; it does not invoke the whisper executable. The worker uses bfloat16, timestamped HF pipeline chunks, and repairs missing/reversed timestamps before summary generation.

# Args: youtube_url [title]
python3 scripts/youtube-summary.py "https://youtube.com/watch?v=..." "Title"

Output format:

📄 TRANSCRIPT:
[raw continuous transcription text]

📋 TRANSCRIPT SEGMENTS (JSON):
[[start_s, end_s, "text"], ...]

upload-gist.py

Upload to GitHub Gist — auto-detects file paths vs raw content:

# Auto-detects: if arg is an existing file, reads its content; otherwise uses raw text
python3 scripts/upload-gist.py <summary_file_or_content> <transcript_file_or_content> <title>

Output:

✅ Gist URL: https://gist.github.com/.../xxxxx

Gist structure:

• summary.md — Formatted summary (full summary at top)
• transcript.txt — Raw transcript (完整原始逐字稿)

Note: The script always reads file content into memory before sending to the API — raw paths are never passed to GitHub. This prevents shell variable expansion issues when passing large text content as arguments.

Agent Usage Notes

The entire pipeline is self-contained: download → transcribe → assemble summary from transcript segments → upload to Gist. No calls to Qwen or any other LLM.

When uploading to Gist, always pass the actual content, not file paths. However, both scripts now auto-detect:

Arg type	Behavior
Existing file path	Content is read into memory, then uploaded
Any other text	Treated as raw content and uploaded directly

This means the agent can safely pass either form — file paths are resolved automatically, and raw content is used as-is. Shell variable expansion issues (e.g., ugur-spark-vllm-docker being treated as a command) are avoided because the scripts read content internally.

When uploading to Gist, always pass the actual content, not file paths. However, both scripts now auto-detect:

Gist Output Format

summary.md

# 🎬 [Video Title]

**來源**: YouTube
**原始連結**: https://youtube.com/watch?v=xxx
**上傳時間**: YYYY/MM/DD HH:MM:SS

---

## 一、完整摘要

[3-5 個要點總結核心內容，讓讀者不用往下看就能掌握重點]
- **核心觀點**：...
- **關鍵發現**：...
- **結論**：...

---

## 二、逐段內容拆解

### ⏱️ 00:00 - 01:30 【第一段主題】
- [重點 1]
- [重點 2]

### ⏱️ 01:30 - 03:45 【第二段主題】
- [重點 1]
- [重點 2]

---

## 三、重點數據／事實
- [數據 1]
- [數據 2]

---

## 四、結論
[結論]

---

## 五、【全內容】最後一段的完整內容

[如果影片最後一段是訪談、對談、Q&A、教程、程式碼演示等結構化內容，在此完整還原。]

### 1. 對談／訪談類型

**問**：[問題原文]
**答**：[回答原文，含所有細節]

### 2. 教程／程式碼類型

[完整程式碼、步驟說明、輸出結果]

### 3. Q&A 問答集

**Q1**：[觀眾提問]
**A1**：[完整回答]

**Q2**：[觀眾提問]
**A2**：[完整回答]

### 4. 產品評測／開箱類型

| 項目 | 詳細說明 |
|------|---------|
| 規格參數 | [完整規格表] |
| 優點 | [逐項說明] |
| 缺點 | [逐項說明] |
| 評分 | [各面向評分 + 總分] |
| 購買建議 | [目標族群、價格區間] |

### 5. 新聞報導／深度解析類型

**事件背景**：[完整背景說明]

**關鍵事實**：
- [事實 1]
- [事實 2]

**各方反應**：
- [各方立場與原話]

**影響評估**：[完整分析]

### 6. 演講／Keynote 類型

**主題**：[演講主題]

**核心論點**：
1. [論點 1] — [完整論述]
2. [論點 2] — [完整論述]
3. [論點 3] — [完整論述]

**金句／重點摘錄**：
> [直接引用原話]

### 7. 理财／投資分析類型

**分析標的**：[股票/基金/加密貨幣等]

**基本面分析**：
- [營收、獲利、成長率等完整數據]

**技術分析**：
- [支撐位、壓力位、均線、指標]

**操作建議**：
- [進場/出場策略、停損停利點]

### 8. 運動賽事／紀錄類型

**賽事資訊**：[對陣雙方、日期、地點]

**逐場/逐回合記錄**：

#### 第 1 節／回合
[完整過程、關鍵時刻、得分]

#### 第 2 節／回合
[完整過程、關鍵時刻、得分]

**賽後分析**：[完整戰術分析、教練選手原話]

### 9. 健康／健身教學類型

**教學目標**：[鍛鍊部位、預期效果]

**動作分解**：

| 步驟 | 動作說明 | 時間/次數 | 注意事項 |
|------|---------|----------|---------|
| 1 | [完整動作描述] | [時數/次數] | [關鍵要點] |
| 2 | [完整動作描述] | [時數/次數] | [關鍵要點] |

**常見錯誤**：
- [錯誤 1] → [正確做法]

### 10. 歷史紀錄／紀錄片類型

**時代背景**：[完整歷史脈絡]

**關鍵人物／事件**：
- [人物 1]：[完整生平／貢獻]
- [人物 2]：[完整生平／貢獻]

**時間軸**：
- [年份]：[事件 + 詳細說明]
- [年份]：[事件 + 詳細說明]

### 11. 音樂／歌詞解析類型

**歌曲資訊**：[歌名、歌手、專輯、發行日期]

**完整歌詞**：

[第 1 段]
[完整歌詞]

[第 2 段]
[完整歌詞]

**創作背景**：[完整故事、靈感來源]

### 12. Cooking／食譜類型

**菜名**：[完整名稱]

**材料清單**：
| 食材 | 用量 | 備註 |
|------|------|------|
| [食材] | [精確用量] | [備註] |

**步驟**：
1. [完整步驟說明 + 時間 + 火候]
2. [完整步驟說明 + 時間 + 火候]

**關鍵技巧**：[完整秘訣說明]

### 12. 旅行／Vlog 類型

**地點資訊**：[完整地名、交通方式、行程日期]

**逐日行程**：

#### Day 1
- [08:00] [活動 + 詳細描述 + 費用]
- [12:00] [活動 + 詳細描述 + 費用]
- [18:00] [活動 + 詳細描述 + 費用]

**實用資訊**：
| 項目 | 詳細內容 |
|------|---------|
| 住宿 | [完整名稱、價格、評價] |
| 交通 | [詳細路線與費用] |
| 預算 | [總花費與分配] |
| 推薦 | [必去地點 + 理由] |

**其他結構化內容類型

[完整內容]

---
*Generated by youtube-summary skill (no external LLM calls)*

transcript.txt

[完整原始逐字稿，不做任何修改或摘要]

ASR Configuration

Audio language	Backend	Invocation
Taiwanese Hokkien (台語)	Breeze ASR 26	--asr breeze
All other languages	Whisper Turbo	default (--asr whisper)

Breeze ASR endpoint:

• URL: http://127.0.0.1:8025
• Max segment: 28 seconds (3000 mel feature limit)
• Audio format: Opus, 48kHz, mono, 16k
• Endpoint: POST /transcribe with file=@audio.ogg

Whisper Turbo is invoked through the Breeze streaming project's Transformers pipeline in /home/chihmin/models-work/flux2/.venv-rocm72/bin/python when available. The worker writes timestamped JSON used for the section breakdown and sets the Strix Halo ROCm variables HSA_OVERRIDE_GFX_VERSION=11.5.1, PYTORCH_HIP_ALLOC_CONF=expandable_segments:True, and TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1.

Override defaults when needed:

• YOUTUBE_SUMMARY_PYTHON — Python executable for the Whisper worker.
• YOUTUBE_SUMMARY_WHISPER_MODEL — local HF snapshot or model ID.
• YOUTUBE_SUMMARY_WHISPER_LANGUAGE — optional forced language such as english.
• HF_HOME — Hugging Face cache root.

Error Handling

Issue	Fix
Breeze ASR 500	Wrong audio format — ensure opus/48kHz/mono/16k
Long-form error	Audio >30s — split into ≤28s segments
yt-dlp fails	Check URL format or network
video path not found	The script now discovers video.*; check the yt-dlp output if no file was produced
Whisper GPU crash	Confirm the Breeze ROCm venv/model path; do not replace the worker with the standalone whisper CLI
invalid Whisper timestamps	The worker drops empty chunks and clamps reversed/missing timestamps
Gist upload fails	Token invalid or no gist scope
Empty transcript	Video has no speech (music only)
Gist content corrupted	Content was passed as shell args with variable expansion — now fixed: scripts auto-detect file paths vs raw content

Notes

• Always use Python for JSON (avoids shell escaping)
• Public gists by default
• Gist includes BOTH summary.md and raw transcript.txt
• Transcript is uploaded unmodified (完整原始逐字稿)
• Work dir: /tmp/youtube-summary (cleaned up after)
• Section breakdown uses timestamps: ⏱️ 00:00 - 01:30 【主題】
• Section 5 — 全內容：If the last segment is an interview, dialogue, Q&A, tutorial, code walkthrough, or any structured content type, reproduce the full verbatim content in this section. For interviews, use **問**： / **答**： format. For tutorials, include complete code/examples. This is the definitive reference section.
• No external LLM calls — everything is assembled from the raw transcript directly.