YouTube-as-AI-curated-reading-feed (multi-stage cascade)
Hour-long videos are the most under-leveraged input for busy operators. Three-model cascade turns each into a 5-minute illustrated read.
What
Curate a list of YouTube channels you would watch if you had time. A cron scraper pulls new videos via Bright Data / SerpAPI. A cheap model (Haiku) parses subtitles for topic + structure. A mid-tier model (Gemini Flash) enriches with screenshots at key moments + a 200-word summary. Output: a markdown digest in Obsidian, with embedded images, sized so a 1-hour video becomes a 5-minute read.
Why it matters
Long-form video is the most under-leveraged channel for a busy operator. SaaS summarizers exist but cost per video and miss visual content. This cascade is cost-conscious (Haiku where cheap suffices, Gemini Flash for vision) and lands in the operator's KM substrate, not a third-party app.
End-to-end
- Channel list: 10–30 channels worth watching. Markdown file checked in.
- Cron scraper (Bright Data or SerpAPI) — pulls new uploads daily.
- Stage 1 (Haiku): fetch transcript / captions → classify topic + outline.
- Stage 2 (Gemini Flash with vision): for the top 5–10 moments by transcript signal, grab the screenshot + 2-sentence caption.
- Compose markdown: title, source, 200-word abstract, outline, 5–10 screenshots with captions, source link, runtime.
- Drop into Obsidian under
inbox/youtube/<date>/. - Daily morning digest: top 5 items by interestingness → Telegram.
Gotchas
- Stage 1 hallucinates aggressively on bad captions. Always retain the original transcript path so downstream agents can re-derive. Don't let the Haiku abstract become the ground truth.
Tools
- Bright Data, SerpAPI, or yt-dlp + a YouTube data API key
- Haiku + Gemini Flash via API (or whatever cheap-with-vision pair exists at run time)
- Obsidian (or equivalent) for output