YT Digest

YouTube Transcript Analyzer — Automated highlights from your favorite channels.

YT Digest is a Python script that monitors a YouTube channel, fetches transcripts from long-form videos, and sends them to Google Gemini for top-5 highlights analysis. Everything is stored in a local SQLite database for easy retrieval.

Features

• RSS Feed Monitoring — Polls a YouTube channel's RSS feed for recent videos
• Smart Video Filtering — Skips Shorts, clips under 30 min, and members-only videos; no upper duration limit
• Self-Installing — On first run, copies itself into %APPDATA%\YTDigest so everything lives in one place
• Transcript Retrieval — Fetches auto-generated or manual transcripts via YouTube's API
• AI-Powered Highlights — Sends transcripts to Google Gemini 2.5 Flash for top-5 highlights
• Transcript Retry — Videos without transcripts are retried over 48 hours (9 attempts at 6-hour intervals) before giving up
• Manual Summary — Import a Whisper (or other) transcript via --summarize to generate highlights for videos with no auto-transcript
• SQLite Storage — Stores titles, dates, transcripts, and summaries permanently
• Rate Limiting — 10-minute minimum between transcript fetches, tracked persistently on disk (Gemini calls are not rate-limited)
• 3-Day Lookback — On first run (or after outage), catches up on recent videos chronologically
• Auto-Dependency Check — Verifies and auto-installs pip packages on startup
• External Config — All settings in config.ini (no hardcoded API keys)

Architecture

OpenClaw (Cron) → yt_digest.py
                    ├── YouTube RSS Feed
                    ├── Video Filter (Shorts / Duration)
                    ├── YouTube Transcript API
                    ├── Google Gemini 2.5 Flash
                    └── SQLite Database

Quick Start

1. Clone the repo

git clone https://github.com/Digitalgods2/YTdigest.git

2. Run it

cd YTdigest
python yt_digest.py

That's it. The script handles everything automatically:

1. Installs dependencies — auto-detects missing pip packages and installs them
2. Copies itself to AppData — installs yt_digest.py and requirements_ytdigest.txt into %APPDATA%\YTDigest\
3. Prompts for setup — asks for your YouTube Channel ID and Gemini API key, saves to config.ini
4. Starts processing — fetches videos, transcripts, and generates highlights

After the first run, everything lives in %APPDATA%\YTDigest\ and the clone folder can be deleted.

Finding a Channel ID:

yt-dlp --print channel_id --playlist-items 1 "https://www.youtube.com/@ChannelHandle"

Getting a Gemini API key: Sign up free at Google AI Studio

3. Schedule (optional)

Set up as a scheduled task in OpenClaw (Clawdbot) with cron: 0 */6 * * *

Run the command: python "%APPDATA%\YTDigest\yt_digest.py"

File Layout

%APPDATA%\YTDigest\
├── yt_digest.py              # Main script
├── requirements_ytdigest.txt # Dependencies
├── config.ini                # API keys & settings (auto-created)
├── ytdigest.db               # SQLite database
├── ytdigest.log              # Execution log
└── last_api_call.txt         # Rate limiter state

Database Schema

Column	Type	Description
video_id	TEXT (PK)	YouTube video ID
title	TEXT	Video title
published	DATETIME	Original publish date (ISO 8601)
processed_at	DATETIME	When YT Digest processed it
status	TEXT	done, pending_transcript, or no_transcript
summary	TEXT	Gemini's top 5 highlights
transcript	TEXT	Full transcript with timestamps
retry_count	INTEGER	Number of transcript fetch attempts

Video Statuses

Status	Meaning
done	Transcript fetched and Gemini highlights generated
pending_transcript	No transcript yet — will retry automatically every 6 hours
no_transcript	Gave up after 9 attempts (~48 hours) — use --summarize for manual import

Manual Summary (--summarize)

If a video ends up as no_transcript, you can provide your own transcript (e.g. from Whisper) and generate highlights:

python yt_digest.py --summarize VIDEO_ID path/to/transcript.txt

This imports the transcript, sends it to Gemini for analysis, and marks the video as done.

Video Filtering

Videos are classified through a multi-step pipeline:

1. Thumbnail check — hq2.jpg pattern indicates a Short
2. URL check — /shorts/ in the link
3. Duration check — Fetched from video page meta tag (PT1H30M45S format)

Result	Criteria
Skip	Shorts (any detection method)
Skip	Under 30 minutes
Skip	Members-only (title contains [member access] or [members only])
Process	30 minutes and up (no upper limit)

Dependencies

Package	Min Version	Purpose
feedparser	6.0.0	RSS feed parsing
requests	2.25.0	HTTP requests for duration detection
youtube-transcript-api	1.2.0	Transcript fetching
google-genai	1.0.0	Gemini API SDK

Documentation

See YT_Digest_Installation_Guide_v3.pdf for the full installation and reference guide, including architecture diagrams, OpenClaw scheduled task setup, and troubleshooting.

License

MIT