video-downloader

name	video-downloader
description	Download a video from YouTube, Instagram, TikTok, Twitter/X, Facebook, Vimeo (and 1000+ yt-dlp sites) to local disk, and optionally transcribe it. Standalone utility for repurposing and analysis — not part of the static publishing pipeline.
allowed-tools	Read, Bash
area	Productivity
use_for	Download a video from a URL (YouTube/IG/TikTok/X/FB/Vimeo/+) to local disk and optionally transcribe it — a standalone repurposing/analysis utility
deps	{"files":[],"env":[]}

Maintenance

Agent	Version	Last Changed
Link	v1.0.1	May 28, 2026

Description: Download a video from a URL to local disk and optionally transcribe it — standalone utility.

Change Log

v1.0.1 — May 28, 2026

• Removed the bundled scripts/download_video.py + $CLAUDE_PLUGIN_ROOT path. Cowork (not Claude Code) is the runtime and doesn't expose CLAUDE_PLUGIN_ROOT, and link-skills has no bundled-runtime-script convention. Reworked to run yt-dlp inline (python3 -m pip install yt-dlp → python3 -m yt_dlp) in the skill's Bash block — the same local-glue pattern content-generator uses for its inline Python. No plugin-relative paths, no shipped .py.

v1.0.0 — May 28, 2026

• New standalone utility (yt-dlp + optional Whisper transcription). Not wired into the static publishing pipeline; transcripts can feed meeting-analyzer. No brand context required.

---

SKILL.md — Video Downloader

A standalone utility — no brand context, no Notion, no dashboard logging. Use it whenever a video URL needs to come down to local disk for repurposing (clip → static post), analysis (transcribe → summarize), or archiving. Everything runs inline in the Cowork bash sandbox — there is no bundled script and no gateway tool (yt-dlp is a local binary; the serverless gateway can't run it).

When to use

• "Download this video: [URL]"
• "Get this reel / short / clip: [URL]"
• "Transcribe this video: [URL]" (download + transcribe)

Do not use this for the static social pipeline (social-calendar → content-creation → creative-designer → social-publisher). It stands alone.

Step 1 — Download (inline yt-dlp)

yt-dlp is a Python package — install it on demand and run it as a module (no PATH dependency, no bundled file). Substitute the real URL and (optionally) OUT:

OUT="${OUTPUT_DIR:-/tmp/downloads}"; mkdir -p "$OUT"
python3 -m pip install --quiet --upgrade yt-dlp 2>/dev/null || pip install --quiet --upgrade yt-dlp

# Best-quality MP4 (merges video+audio when ffmpeg is present)
python3 -m yt_dlp -f "bestvideo[ext=mp4]+bestaudio/best[ext=mp4]/best" --merge-output-format mp4 \
  --no-playlist --no-warnings -o "$OUT/%(title).80s.%(ext)s" "VIDEO_URL"

Variants:

Need	Tweak
Cap quality (e.g. 720p)	-f "bestvideo[height<=720][ext=mp4]+bestaudio/best[height<=720]"
Audio only (MP3)	-x --audio-format mp3 (drop the -f / --merge-output-format flags)
Output directory	set OUT (default /tmp/downloads)

If the merge step fails (no ffmpeg in the sandbox), retry with a single-file format that needs no merge:

python3 -m yt_dlp -f "best[ext=mp4]/best" --no-playlist --no-warnings -o "$OUT/%(title).80s.%(ext)s" "VIDEO_URL"

Report the resulting file path (most recent file in $OUT):

ls -t "$OUT" | head -1

Supported: YouTube, Instagram, TikTok, Twitter/X, Facebook, Vimeo, and 1000+ other yt-dlp sites. If pip/network is unavailable in the sandbox, say so rather than failing silently.

Step 2 — Transcribe (optional)

Only when a transcript is wanted, and only if ffmpeg + whisper are available (both are heavy — guard for them):

command -v ffmpeg >/dev/null || { echo "ffmpeg not available here — cannot transcribe"; exit 0; }
ffmpeg -i "/path/to/video.mp4" -ar 16000 -ac 1 -y /tmp/video-audio.wav
python3 -m pip install --quiet --upgrade openai-whisper 2>/dev/null
python3 -c "import whisper; print(whisper.load_model('base').transcribe('/tmp/video-audio.wav')['text'])"

Save the transcript next to the video (<basename>.txt) or print it. A transcript can feed meeting-analyzer (action items) or content-creation (repurposing into posts).

Step 3 — Cleanup

rm -f /tmp/video-audio.wav when done. Leave the downloaded video unless the user asked for transcript-only.

Notes

• Respect copyright and platform ToS — download only content the user is authorized to use (their own, licensed, or fair-use analysis).
• Local-only by design: no gateway tool (yt-dlp/ffmpeg/whisper can't run in the serverless gateway) and no bundled script (Cowork has no CLAUDE_PLUGIN_ROOT). If pip / network / ffmpeg aren't available in the sandbox, report the limitation rather than failing silently.