blink-new Watch Video
Analyze a public video URL and produce a vivid, detailed report of what was seen on screen AND said aloud — precise enough for an engineer, PM, or designer to act on without watching it.
Workflow
Step 1: Determine video platform
| Platform | Strategy |
|---|---|
| Loom | Extract metadata + thumbnail from page; use Gemini on thumbnail + description |
| YouTube | Use yt-dlp to download audio → Whisper transcription; or Gemini with thumbnail |
| Other | Try yt-dlp first; fall back to page scrape + thumbnail |
Step 2: Extract what you can without downloading
Always try to get free data first (no disk needed):
# Get Loom metadata (title, description, duration, transcript status)
python3 scripts/loom_meta.py "https://www.loom.com/share/VIDEO_ID"
# Get YouTube metadata
yt-dlp --dump-json --no-download "YOUTUBE_URL" | python3 -c "
import sys, json
d = json.load(sys.stdin)
print('Title:', d['title'])
print('Description:', d['description'][:500])
print('Duration:', d['duration'])
"
Step 3: Get the thumbnail
Thumbnails are always public and reveal UI state, screen content, numbers.
# Loom: thumbnail URL is in oembed response
curl -s "https://www.loom.com/v1/oembed?url=LOOM_URL" | python3 -c "
import sys, json
d = json.load(sys.stdin)
print(d['thumbnail_url']) # animated GIF — download this
"
# YouTube: predictable URL
# https://img.youtube.com/vi/VIDEO_ID/maxresdefault.jpg
Download the thumbnail:
curl -s "THUMBNAIL_URL" -o /tmp/video_thumb.gif # or .jpg
Step 4: Analyze with Gemini
Use scripts/analyze_video.py — pass the thumbnail + all metadata.
python3 scripts/analyze_video.py \
--thumb /tmp/video_thumb.gif \
--title "Video title" \
--description "Auto-generated description text" \
--duration 143
Or call the Gemini API directly (see gemini-api.md).
Step 5: If disk space allows — download + full analysis
# Check disk first
df -h ~ | awk 'NR==2 {print $4}'
# Download with yt-dlp (works for Loom + YouTube + 1000+ sites)
yt-dlp "VIDEO_URL" -o /tmp/video.mp4
# Option A: Gemini video analysis (best for screen recordings)
python3 scripts/analyze_video.py --video /tmp/video.mp4
# Option B: Whisper transcription only (fast, audio-only)
curl -X POST https://api.openai.com/v1/audio/transcriptions \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-F "file=@/tmp/video.mp4" \
-F "model=whisper-1"
Output Format
Produce this structure every time:
## Video Analysis: [Title]
**URL:** [link] | **Duration:** [X:XX]
### What's on screen
[Vivid description of every visible UI element, number, text, state shown in the video/thumbnail. Be exhaustive — mention exact values, UI panel names, error messages, data shown.]
### What was said
[Transcript or paraphrase of the narration, with key quotes verbatim]
### Problem / Topic Summary
[1-2 sentences: what this video is about or demonstrating]
### Key Data Points
| Item | Value |
|------|-------|
| [exact numbers, names, states observed] | [value] |
### Action Items / Root Cause (if bug report)
[Specific things to investigate or do, based on the video content]
Platform Notes
See platforms.md for platform-specific quirks (Loom CDN auth, YouTube age-gate, etc.)
Disk Full Fallback
If disk is full (df -h shows < 500MB free):
- Use thumbnail + metadata only (Steps 1–4 above)
- Loom auto-descriptions are highly detailed — combine with thumbnail for good coverage
- Never fail silently — always produce a report from whatever data is available