Claude 'Watch' Plugin Turns Videos into Queryable AI Assets
Install free 'watch' Claude plugin using yt-dlp/FFmpeg to extract 80 timestamped frames + transcripts from videos, enabling NotebookLM-style analysis of sales calls, Looms, and tutorials for instant playbooks and automations.
Video-to-Data Pipeline Unlocks Hidden Business Knowledge
Feed any public video URL (YouTube, Twitter/X, Loom, Instagram MP4s) to Claude's 'watch' plugin, which uses yt-dlp to download, FFmpeg to pull 80 evenly spaced timestamped frames, and YouTube captions or OpenAI Whisper for transcripts. Costs stay low: free on Claude Max (token budget), ~$1/video via API at Opus pricing. Claude processes frames + text natively, answering like PDFs—e.g., a 12-minute video processes in 1+ minute. Install in 30 seconds via Claude Code (IDE like Cursor or desktop app): /plugin marketplace add https://github.com/.../claude-video then /plugin install watch@claude-video. Caps frames to prevent runaway costs, sampling sparsely for long videos (e.g., 43 minutes gets same 80 frames spread thinner), sufficient for business spines but not frame-perfect debugging.
Private/paywalled content fails without accessible URLs; works on local files too. Output saves as timestamped files for follow-ups, turning unqueryable video knowledge (sales calls, onboardings) into analyzable assets.
Analyze Archives to Fill Content Gaps and Build Instantly
Paste 28 YouTube URLs into channel.txt, prompt Claude: "Read channel.txt, run /watch on each, save outputs named after video, process one-by-one." Generates 28 files (transcripts + frame insights). Follow-up: "Read all outputs, extract core frameworks/claims/audience per video; identify top 3 repeated frameworks, uncovered topics for agency owners/service operators (e.g., AI pricing/packaging, ROI proof, 30-day team rollout, when not to use AI), script outline in your voice for one gap." Reveals audience split (AI installers vs. sellers), never-covered topics like client firing, outputs ready-to-film script—automates self-audit without manual review.
For saved tutorials: /watch Twitter video (Whisper transcribes no-captions), prompt: "Extract steps as checklist in setup.md; scaffold/do codable steps (e.g., Claude.md, context/memory.md, skills for LinkedIn scraping/likes, lead qual agent via Unipile/Firecrawl, Notion push), stop for credentials." Builds full para-style repo in ~7 minutes: playbooks (intelligence loop post-50-100 messages), resources, campaign planner—handles risky actions only after approval, turns 2-week bookmark into deployable LinkedIn outreach bot needing just API keys (Firecrawl, Unipile, Amplify, Notion).
Four Playbooks from Video Inputs Scale Service Businesses
Looms to SOPs: Feed 20 team recordings, extract step-by-step playbooks + training docs—replaces $5K consultant. Sales calls to playbook: 30 calls yield real objection patterns killing close rates + proven openers (data over memory). Competitor gaps: Top 15 videos output hook patterns + audience-requested topics for instant content briefs. Courses to KB: All recordings become 24/7 searchable Q&A, ends repetitive DMs. Each doubles as sellable AI service; package/pricing via communities like skool.com/systems-to-scale.
Trade-offs: Public URLs only, no paywall bypass; frame sampling misses fine details. Delivers production ROI: query sales/ops video goldmine, build from tutorials, compete via analysis—ships what NotebookLM couldn't.