Save hours every week
Skip videos that aren't relevant. Read a 3-line summary, decide if you want to watch the full thing.
Running every morning at 7 AM
Wake up to a daily summary of YouTube videos you actually care about
A personal robot that watches your favorite YouTube channels overnight, writes a clean summary of each new video, and drops them into your Notion before you finish your coffee.
Why this exists
YouTube subscriptions pile up. Half of what's posted isn't relevant. Watching a 30-minute video to find out it's not useful is a waste. This system reads everything for you and gives you the gist.
Never miss the good stuff
A trusted channel posts something important and you scrolled past it? Not anymore. Every video gets a glance.
One inbox for everything
All summaries land in one Notion page. Search, browse, share with your team, archive old ones.
How it works
Five steps, fully automated, every morning. You don't touch anything — your computer doesn't even need to be on.
flowchart TD
A([⏰ 7 AM trigger
cloud server wakes up]) --> B[📥 STEP 1: Collect
Visit every channel
Find videos posted yesterday] B --> C[🤖 STEP 2: Summarize
Read each video's transcript
Write a 3-line summary] C --> D[✅ STEP 3: Review
Check formatting
Fix any issues] D --> E[📝 STEP 4: Publish
Create a new Notion page
Save a backup to GitHub] E --> F([📧 You wake up
open Notion
read summaries]) style A fill:#FF6B35,stroke:#FF6B35,color:#fff style F fill:#4ECDC4,stroke:#4ECDC4,color:#000 style B fill:#141414,stroke:#FF6B35,color:#EDEDED style C fill:#141414,stroke:#FF6B35,color:#EDEDED style D fill:#141414,stroke:#FF6B35,color:#EDEDED style E fill:#141414,stroke:#FF6B35,color:#EDEDED
cloud server wakes up]) --> B[📥 STEP 1: Collect
Visit every channel
Find videos posted yesterday] B --> C[🤖 STEP 2: Summarize
Read each video's transcript
Write a 3-line summary] C --> D[✅ STEP 3: Review
Check formatting
Fix any issues] D --> E[📝 STEP 4: Publish
Create a new Notion page
Save a backup to GitHub] E --> F([📧 You wake up
open Notion
read summaries]) style A fill:#FF6B35,stroke:#FF6B35,color:#fff style F fill:#4ECDC4,stroke:#4ECDC4,color:#000 style B fill:#141414,stroke:#FF6B35,color:#EDEDED style C fill:#141414,stroke:#FF6B35,color:#EDEDED style D fill:#141414,stroke:#FF6B35,color:#EDEDED style E fill:#141414,stroke:#FF6B35,color:#EDEDED
1
Collect new videos
A tool called yt-dlp visits each YouTube channel on your list, finds videos uploaded yesterday, and grabs their titles, descriptions, and captions.
yt-dlp + cookies2
Summarize with AI
Claude (an AI assistant) reads each video's content and writes a structured summary: 2–4 sentences of the gist, key timestamps, and one-line takeaway.
Claude Haiku 4.53
Quality check
A small program checks each summary follows the right format. Catches missing sections, fixes timestamps, makes sure nothing's broken before publishing.
Node.js validator4
Publish to Notion
Creates a new page inside your "YouTube Summary" parent page. Each video becomes a clickable link. Channel names show up in red so you can scan fast.
Notion API5
Save a backup
Same content also gets saved to GitHub so you have a permanent archive. Useful if Notion has issues or you want to grep old summaries later.
Git commitWhat you actually see
A new Notion page lands every morning. Here's exactly what one entry looks like — clickable channel header, clickable video title, one-line takeaway, key summary, and a clickable timeline that jumps to the exact moment in the video.
2026-05-02
Building Conversational Agents with Gemini Live API
💡 Real-time multimodal agents are getting cheap and fast — worth a serious look for any voice or IoT product.
Key Summary
Google DeepMind engineers walk through building a real-time voice assistant with the new Live API. The session covers WebSocket setup, multimodal streaming input (audio + screen share), and how to invoke tools mid-conversation without breaking the audio loop.
In the live demo they swap a stock prompt for a tool-aware one and show latency dropping from ~900ms to ~250ms once partial audio streaming is enabled, then add a calendar tool call mid-turn to demonstrate uninterrupted speech.
In the live demo they swap a stock prompt for a tool-aware one and show latency dropping from ~900ms to ~250ms once partial audio streaming is enabled, then add a calendar tool call mid-turn to demonstrate uninterrupted speech.
Key Timeline
00:00:42Why streaming-first APIs change agent design00:04:18Setting up the WebSocket session and auth00:11:30Handling barge-in and turn detection00:18:55Tool calls during a live conversation00:25:10Production tips: cost, latency, fallbacks
Why Building Eval Platforms Is Hard
💡 If you ship AI features, a real eval pipeline matters more than your model choice.
Key Summary
A Braintrust engineer breaks down why "good" is so hard to define for AI outputs. He walks through the labeling, versioning, and team-alignment problems most eval teams hit by month three, and proposes a minimal scoring loop that scales without a full platform.
As a concrete example he shows two labelers scoring the same 50 outputs and disagreeing on 18 of them, then demonstrates how a tiny rubric and a paired-review step cut disagreement down to 4 — without buying any new tooling.
As a concrete example he shows two labelers scoring the same 50 outputs and disagreeing on 18 of them, then demonstrates how a tiny rubric and a paired-review step cut disagreement down to 4 — without buying any new tooling.
Key Timeline
00:01:20The illusion of "we'll just measure quality"00:07:45Labelers disagree more than you think00:14:02Versioning prompts vs. versioning datasets00:22:30A minimal scoring loop that actually works
The 200-LoC Skill That Replaced 12,000 Lines of Code
💡 Markdown-driven workflows are the new "config as code" — most internal tools are over-engineered.
Key Summary
A walkthrough of how an engineer rewrote a complex Git Worktrees feature as a Claude Code Skill — going from 12,000 lines of TypeScript to ~200 lines of plain markdown. Covers the trade-offs, what broke, and where the markdown approach hits its limits.
The demo shows the original TypeScript version handling a stale-lock edge case in 80 lines, vs. the markdown skill solving the same case with a single instruction line plus a tool call — a useful example of where prose beats code, and one case where it doesn't.
The demo shows the original TypeScript version handling a stale-lock edge case in 80 lines, vs. the markdown skill solving the same case with a single instruction line plus a tool call — a useful example of where prose beats code, and one case where it doesn't.
Key Timeline
What runs and when
Three different jobs you can use. The daily one is fully automatic; the other two run on demand.
AUTO
Every day · 7:00 AM
Daily Digest
Yesterday's videos from all 8 channels. Runs by itself in the cloud. You don't have to do anything.
MANUAL
On demand
Weekly Recap
Last 7 days, all channels, capped at 3 videos per channel. Click "Run workflow" when you want a Sunday-style review.
MANUAL
On demand
Single Channel Catch-up
Pick one channel (e.g., @unrealtech), get the most recent N videos summarized in one Notion page. Useful for new subscriptions.
The free vs paid trick
Most "AI summary" tools charge a monthly API fee. This one piggybacks on existing subscriptions, so the running cost is essentially zero.
Most automation tools
Need their own API key
- Sign up for an OpenAI / Anthropic API account
- Add a credit card
- Pay per summary (a few cents each)
- Monitor usage so you don't get billed too much
This system
Reuses your Claude subscription
- Already paying for Claude Pro? It's covered
- One-time setup: copy a token, paste into GitHub
- $0 added cost (uses subscription quota)
- GitHub Actions itself is free for public repos
Quick numbers
~5 min
to run end-to-end
8
channels watched daily
$0
monthly cost (with Claude Pro)
24/7
runs even when laptop is off
See the actual code
This whole system is open source. Fork it, customize the channels, change the format — it's yours.