# Extract Youtube Transcript > Use this skill when the user wants to extract audio from a YouTube video and transcribe it. Triggers include: 'extract transcript from youtube', 'transcribe this video', 'get transcript from YouTube', 'extract audio and transcribe', or any message that includes a YouTube URL and asks for a transcript or transcription. - Skill: `rtemm/extract-youtube-transcript` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add rtemm/extract-youtube-transcript` - Raw SKILL.md: https://api.skillmd.com/api/skills/rtemm/extract-youtube-transcript/raw - Safety review: pending (external: skill-scanner WARNING, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: Rtemm (https://skillmd.com/u/rtemm) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/rtemm/extract-youtube-transcript --- # extract-transcript Skill When invoked, follow these steps in order. Do not skip steps. --- ## Step 0 — Verify and install dependencies Check each tool and install if missing. Stop and report clearly if any install fails. ### youtube-transcript-api ```bash pip install youtube-transcript-api ``` Always needed — used in Step 3 to check for an existing YouTube transcript. ### yt-dlp ```bash yt-dlp --version ``` If not found: ```bash pip install yt-dlp ``` ### Whisper ```bash whisper --help ``` If not found: ```bash pip install openai-whisper ``` > Whisper is only used if no YouTube transcript is found in Step 3. Install it regardless — it is fast and needed as a fallback. ### ffmpeg ```bash ffmpeg -version ``` If not found, install via winget (Windows 11): ```bash winget install --id Gyan.FFmpeg -e ``` If winget fails, stop and tell the user: > "ffmpeg is required but could not be installed automatically. Please download it from https://ffmpeg.org/download.html, extract it, and add the `bin/` folder to your system PATH. Then re-run this skill." ### python-docx ```bash pip install python-docx ``` Always needed — used in Step 6 to generate the summary `.docx`. After any install, re-verify the tool is available before continuing. If it still fails, stop and report the error. --- ## Step 1 — Get the YouTube URL If the user already provided a URL in their message, use it. If not, ask: > "Please provide the YouTube URL you'd like to transcribe." --- ## Step 2 — Create the base output folder Create `doc summaries output` inside the current working directory if it does not already exist: ```bash mkdir -p "doc summaries output" ``` The per-video subfolder will be created in Step 3 once the video title is known. --- ## Step 3 — Resolve video title and check for a YouTube transcript This step does two things in one script: 1. Fetches the video title from YouTube metadata (always needed for the subfolder name and filenames). 2. Checks whether YouTube already provides a transcript. If yes, saves it to disk and skips audio download entirely. Write and run this Python script: ```python import re, os, subprocess, json from youtube_transcript_api import YouTubeTranscriptApi from youtube_transcript_api._errors import NoTranscriptFound, TranscriptsDisabled url = "" base_output = "doc summaries output" # --- Fetch video title --- meta = subprocess.run( ["yt-dlp", "--dump-json", "--skip-download", url], capture_output=True, text=True ) raw_title = json.loads(meta.stdout)["title"] # Sanitize title for use as a folder/file name (strip Windows-illegal characters) illegal = r'\/:*?"<>|' title = "".join(c for c in raw_title if c not in illegal).strip() # --- Create per-video subfolder --- video_folder = os.path.join(base_output, title) os.makedirs(video_folder, exist_ok=True) # --- Check for YouTube transcript --- video_id = re.search(r"(?:v=|youtu\.be/)([^&?/]+)", url).group(1) ytt = YouTubeTranscriptApi() try: transcript_list = ytt.list(video_id) # Prefer manual English captions → auto-generated English → anything available transcript_obj = None source = None for t in transcript_list: if not t.is_generated and t.language_code.startswith("en"): transcript_obj = t source = "manual" break if transcript_obj is None: for t in transcript_list: if t.is_generated and t.language_code.startswith("en"): transcript_obj = t source = "auto-generated" break if transcript_obj is None: transcript_obj = next(iter(transcript_list)) source = "auto-generated" if transcript_obj.is_generated else "manual" fetched = transcript_obj.fetch() text = " ".join(entry.text for entry in fetched) txt_path = os.path.join(video_folder, f"{title}.txt") with open(txt_path, "w", encoding="utf-8") as f: f.write(text) print(f"TRANSCRIPT_FOUND|{source}|{title}|{video_folder}|{txt_path}") except (NoTranscriptFound, TranscriptsDisabled): print(f"NO_TRANSCRIPT|{title}|{video_folder}") except Exception as e: print(f"NO_TRANSCRIPT|{title}|{video_folder}|error: {e}") ``` **If the output starts with `TRANSCRIPT_FOUND`:** - Parse: `source`, `title`, `video_folder`, `txt_path` from the pipe-delimited output. - Tell the user: "YouTube transcript found ({source}). Skipping audio download and Whisper." - **Skip Steps 4 and 5. Go directly to Step 6.** **If the output starts with `NO_TRANSCRIPT`:** - Parse: `title`, `video_folder` from the output. - Tell the user: "No YouTube transcript available — downloading audio and transcribing with Whisper." - Continue to Step 4. --- ## Step 4 — Download audio *(Only reached if Step 3 found no transcript.)* Download the MP3 directly into the video subfolder resolved in Step 3: ```bash yt-dlp -x --audio-format mp3 "" -o "doc summaries output//%(title)s.%(ext)s" ``` Replace `<title>` with the sanitized title from Step 3. If yt-dlp reports an error (private video, age-restricted, unavailable), tell the user and stop. --- ## Step 5 — Transcribe *(Only reached if Step 3 found no transcript.)* Find the `.mp3` file inside `doc summaries output/<title>/`. Run Whisper on it, outputting the transcript into the same subfolder: ```bash whisper "doc summaries output/<title>/<filename>.mp3" --output_format txt --output_dir "doc summaries output/<title>" --device cuda ``` If CUDA is unavailable (no NVIDIA GPU), retry with CPU: ```bash whisper "doc summaries output/<title>/<filename>.mp3" --output_format txt --output_dir "doc summaries output/<title>" --device cpu ``` Let the user know transcription is running. GPU typically finishes in under a minute; CPU may take several minutes for longer videos. After Whisper finishes, locate the `.txt` it created inside the subfolder and set `txt_path` to that file's full path for use in Step 6. --- ## Step 6 — Generate summary .docx and .txt Read the transcript from `txt_path`. Generate **two** summary files inside `doc summaries output/<title>/` using a single Python script: | File | Purpose | |---|---| | `<title> - Summary.txt` | Plain text — readable in VS Code, easy to copy-paste | | `<title> - Summary.docx` | Formatted Word doc — upload to Google Drive to open as a Google Doc | Both files must contain the same content in these sections: 1. **Title** — video title 2. **Overview** — 2-3 sentence summary of what the video is about 3. **Detailed Summary** — subsections matching the main topics covered in the video 4. **Key Points** — numbered list, each with a bold label (in the .docx) and a 1-2 sentence explanation 5. **Main Takeaways** — bulleted list of the most actionable/important conclusions 6. **Tools or Concepts Referenced** — 2-column table (Name | Description) of every tool, platform, or concept mentioned ### Plain text formatting (.txt) Use plain ASCII so the file is clean in any editor: ``` ================================================ VIDEO TITLE HERE ================================================ OVERVIEW -------- Paragraph text here. DETAILED SUMMARY ---------------- Topic Heading ~~~~~~~~~~~~~ Paragraph text here. KEY POINTS ---------- 1. Bold Label — Explanation sentence here. 2. Bold Label — Explanation sentence here. MAIN TAKEAWAYS -------------- • Takeaway one • Takeaway two TOOLS AND CONCEPTS REFERENCED ------------------------------ Name | Description ----------------------|--------------------------------------------- Tool name | Description here. ``` Save both files to `doc summaries output/<title>/`. --- ## Step 7 — Report results When complete, tell the user: 1. The full path of the video subfolder: `<cwd>/doc summaries output/<title>/` 2. List every file inside it: - `<title>.mp3` — downloaded audio *(only present if Whisper path was taken)* - `<title>.txt` — raw transcript - `<title> - Summary.txt` — plain text summary (open in VS Code) - `<title> - Summary.docx` — formatted summary (upload to Google Drive) 3. Whether the transcript came from YouTube (manual or auto-generated) or was generated by Whisper. 4. Print the first ~300 characters of the raw transcript so the user can verify it looks correct. 5. Remind the user they can upload the `.docx` to Google Drive and it will open automatically as a Google Doc. --- ## Error handling reference | Error | Action | |---|---| | pip not found | Tell user to install Python from https://www.python.org/downloads/ | | Private / unavailable video | Tell user yt-dlp cannot download this video | | ffmpeg not in PATH after install | Instruct user to restart terminal or add ffmpeg to PATH manually | | Whisper CUDA out of memory | Retry with `--model base` or `--model tiny` to reduce VRAM usage | | Whisper CUDA not available | Retry with `--device cpu` (will be slower) | | No audio track / unsupported format | yt-dlp will report this; relay the message to the user | | youtube-transcript-api import error | Run `pip install youtube-transcript-api` and retry Step 3 | | Illegal characters in folder name | The Step 3 script sanitizes the title — if mkdir still fails, strip all non-alphanumeric characters except spaces and dashes | | yt-dlp title parse fails (empty stdout) | The video may be private or geo-blocked; tell the user and stop |