Imported from ychenjk-sudo/xiaoyuzhou-transcription-skill (SKILL.md). Install upstream with npx skills add ychenjk-sudo/xiaoyuzhou-transcription-skill. Copyright stays with the author.

小宇宙播客转录与总结

Transcribe 小宇宙 podcast episodes into structured Markdown: speaker-segmented verbatim transcript with timestamps, 5–8 key-point summaries grouped by theme, and 8–10 Q&A pairs covering core content.

Prerequisites

• Environment variable: QWEN_API_KEY — 阿里云 DashScope API Key (required)
• Dependencies: curl, jq, python3

Workflow

Step 1: Transcribe audio

./scripts/transcribe.sh "https://www.xiaoyuzhoufm.com/episode/xxxxx" /tmp/transcript_raw.json

The script extracts the audio URL from the 小宇宙 page, submits an async transcription task to Qwen ASR (qwen3-asr-flash-filetrans, supports up to 8-hour episodes), polls until complete (typically 2–5 min), and downloads the JSON result.

Validation: Confirm /tmp/transcript_raw.json exists and contains a transcripts array before proceeding.

Step 2: Format verbatim transcript

python3 ./scripts/format_transcript.py /tmp/transcript_raw.json /tmp/transcript_formatted.md

Outputs Markdown with **Speaker** [HH:MM:SS] headers, auto-splitting paragraphs longer than 300 characters. Speaker detection uses content markers (e.g. "我是XXX"), not model diarization.

Validation: Confirm the output file is non-empty and contains at least one **...** speaker header.

Step 3: AI-generated summary (Agent)

Read the formatted transcript, then generate:

1. 核心要点 (Key points): 5–8 most important insights, grouped by theme
2. 关键 Q&A (Key Q&A): 8–10 question-answer pairs covering core content
3. Write in Chinese, keep tone concise and professional

Step 4: Assemble final output

Merge the summary and verbatim transcript into one Markdown file saved to /workspace/podcasts/[播客名称].md. See README.md for the full output template.

Validation: Ensure the final file contains all three sections (核心要点, 关键问答, 逐字稿) before reporting success.

Notes

• Long episodes (1 h+) may take 2–10 min to transcribe; the script polls with a 1-hour timeout
• Speaker identification relies on content markers, not model-based diarization
• Summary quality depends on the OpenClaw session model in use