Watch Any Video With Claude
Paste any video URL and Claude reads the transcript plus smart-selected frames to write clean notes in seconds. Works on YouTube, Loom, TikTok, and 1600+ sites.
Before & after
Same prompt, same agent — the only difference is the skill.
Transcribing and summarizing a long video is definitely doable. To give you the best summary, I'd need to know a bit about the video—what's the topic, who's speaking, and what format would work best for you? I could grab the transcript if it's available on YouTube, but if there's no transcript, that's more complicated and might require external services. I can also pull some key frames if visuals are important, though that adds complexity. Let me know what you're working with, and I can suggest an approach. Do you have a preference for summary length or format—bullet points, narrative, or something else?
VIDEO PROCESSED: 90-min YouTube interview MODE USED: Efficient (40x speedup vs balanced mode) PROCESSING TIME: 87 seconds TOKEN COST: 2,100 tokens (vs 84,000 for full transcript) TRANSCRIPT STATUS: Built-in captions extracted FRAMES SELECTED: 12 (scene changes + deictic moments detected) KEY FINDINGS: SPEAKER: Dr. Jane Chen, AI researcher TOPIC: Scaling laws in neural networks—why bigger models get smarter TOP 3 INSIGHTS: 1. Scaling doesn't plateau (2:15-8:40 in video) → Prediction: 10x parameters = 15-20% performance boost → Implication: No evidence of diminishing returns yet 2. Training data quality now > quantity (23:45-31:20) → Key moment: "We wasted 2 years on raw internet scrapes" → Lesson: Curated datasets 3x more efficient 3. Emergent abilities appear without planning (52:15-59:30) → Models suddenly gain reasoning skills during scaling → Prediction paradox: Can't forecast which abilities emerge SCREEN SHARE MOMENTS: 3 detected (graphs at 34:50, live demo at 66:20, cited papers at 79:10) CITATIONS FOUND: → "Chinchilla scaling laws" (reference, year noted) → 3 arXiv papers mentioned (links auto-filed) FILE: ~/notes/chen-scaling-interview-summary.md FORMAT: Timestamped outline with embedded speaker quotes NEXT: Add to notion integration
About this skill
name: universal-video-watcher description: Paste any video URL and Claude reads the transcript plus smart-selected frames to write clean notes in seconds. Works on YouTube, Loom, TikTok, and 1600+ sites. Use when working on watch any video with claude.
Watch Any Video With Claude
Ships four modes (transcript, efficient, balanced, token-burner), scene-change detection, deictic gesture detection ('look here', 'as you can see'), and duplicate-frame culling. Efficient mode pulls only baked-in key frames for 40x speedup at low token cost. Falls back to Groq or OpenAI for transcription on captionless videos and Looms. Use every time you need to consume a video without watching it end to end.
What you get
- Timestamped transcript
- Smart-selected frames aligned to scene changes and deictic moments
- Structured notes
- Per-creator tracking log
- Auto-file destination
- Summary at top
Customize your output
- Mode: transcript, efficient, balanced, token-burner
- Frame cap: 20, 50, 100, all
- Transcription provider: Whisper, Groq, OpenAI
- Auto-file location: dir
- Language: en, es, fr, de, it, pt
- Summary length: paragraph, bullet, exec brief
Example output
'Full digest of a 2-hour Boris Cherny interview in 90 seconds including screen-share moments the raw transcript would miss.'
Best for
Creators, researchers, execs, and PMs who track many video sources weekly.
Note: Groq key recommended over OpenAI for free faster fallback transcription. High confidence.
SKILL.md preview
---
name: universal-video-watcher
description: Use this skill when a video URL needs to be turned into clean notes by reading its transcript and smart-selected frames instead of watching it end to end.
version: 1.0.0
category: Content Creation
author: AgentVolt
license: proprietary
tags:
- content-creation
- p4
---
# Watch Any Video With Claude
Consumes a video from its URL by combining transcript and key-frame analysis into clean notes without watching the full runtime.
## When to use
… (sign up to view the full skill)