podcast-digest-template

A weekly podcast digest pipeline for people who'd rather read than listen.

Pulls YouTube auto-subtitles from a list of podcast channels, extracts structured signal (summary, themes, quotes, tools mentioned, trends) via Gemini 2.5 Flash, and posts a weekly synthesis to Slack. Per-episode details land in threaded replies so the main channel stays skimmable.

Runs on macOS via launchd (cron also works). Costs pennies per week at Flash pricing.

What it does

YouTube channels + playlists
        ↓
 Daily fetcher (launchd, 6am)
        ↓  (dedup via state.json + filters)
 youtube-transcript-api → auto-subs
        ↓
 Gemini 2.5 Flash → structured JSON per episode
        ↓
 Per-episode JSON cache (extracts/)
        ↓
     ... Sunday 2pm ...
        ↓
 Weekly digest compiler → synthesis pass via Flash
        ↓
 Slack post — main message + threaded per-episode replies

Setup

Prerequisites

• Python 3.9+
• A Google AI Studio API key with billing enabled (docs). Free tier limits (20 req/day) will choke on real volume.
• A Slack bot token with chat:write scope and the bot invited to your target channel.
• macOS if you want launchd scheduling (otherwise cron or any scheduler).

Install

git clone https://github.com/bonus414/podcast-digest-template.git
cd podcast-digest-template
pip install -r requirements.txt
cp feeds.example.json feeds.json

Configure feeds

Edit feeds.json. Each entry is either a YouTube channel or playlist:

{
  "id": "my-feed-id",
  "name": "Display Name",
  "source_type": "channel",
  "channel_id": "UCxxxxxxxxxxxxxxxxxxxx",
  "enabled": true,
  "last_seen_video_id": null
}

To resolve a YouTube handle (e.g. @LatentSpacePod) to a channel ID, use the included helper:

./scripts/resolve_channel_ids.sh @LatentSpacePod @AIDailyBrief @NoPriorsPodcast

Optional per-feed filters:

• "keyword_gate": ["ai","llm",...] — only pick episodes whose title+description contain at least one keyword. Useful for broad feeds where only some episodes are relevant.
• "publish_day_filter": ["Mon","Tue","Wed","Thu","Fri"] — only pick episodes published on these weekdays. Useful for daily feeds where you want weekdays only.

Environment

Set these in your shell env or a .env file:

export GOOGLE_API_KEY="your-google-ai-studio-key"
export SLACK_BOT_TOKEN="xoxb-..."
export SLACK_CHANNEL_ID="C01234567"   # destination for the weekly digest

Run manually

# Dry run — show what would be picked up, no side effects
python3 fetch_episodes.py --dry-run

# First run — backfill last 7 days
python3 fetch_episodes.py --backfill 7

# Pull transcripts + extract
python3 get_transcripts.py
python3 extract_episode.py

# Compile and post weekly digest
python3 compile_digest.py --since 7
# or preview without posting
python3 compile_digest.py --since 7 --no-slack

Outputs:

• transcripts/ — raw auto-subs (gitignored)
• extracts/ — per-episode JSON (one file per episode)
• digests/ — weekly markdown copies of what gets posted
• state.json — per-feed seen-video-id state

Schedule

See launchd/ for example plists and wrapper scripts. Install with:

cp launchd/com.example.podcast-fetch.plist ~/Library/LaunchAgents/
cp launchd/com.example.podcast-digest.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.example.podcast-fetch.plist
launchctl load ~/Library/LaunchAgents/com.example.podcast-digest.plist

Edit the plists first to point to your install path.

Design choices worth calling out

• YouTube auto-subs, not Whisper. Free and fast, but this means audio-only podcasts (lots of the MIT Sloan / Changelog Network kind) are invisible to the pipeline. See "Deferred" below.
• Short-clip filter. Transcripts under 3KB are skipped as likely YouTube Shorts. Keeps the digest from being polluted with 60-second teasers.
• Per-feed dedup via seen-video-ID set. Simple and robust. State capped at 500 IDs per feed.
• Two-pass extraction. Per-episode pass produces raw JSON; a second Flash pass at digest time deduplicates themes across shows and flags cross-show convergence. You could skip the synthesis pass and get a raw dump instead — it's roughly one API call away.
• gemini-2.5-flash, deliberately. Flash cost ~$0.002/episode and returned clean JSON. Gemma 4 31B was 9x slower with thinking-mode leaking into output and hit capacity 503s. See BAKEOFF-RESULTS.md.

Deferred / things you'll hit

• Audio-only feeds. A get_transcripts_whisper.py path would unlock podcasts that live on RSS but not YouTube. Not included because Whisper is its own setup story.
• Channel handle drift. YouTube handles occasionally change. Re-run resolve_channel_ids.sh if a feed goes cold.
• Rate limits. Flash paid tier is fine; free tier will throttle. Enable billing on your GCP project.
• Synthesis quality varies with episode volume. 3-episode weeks feel thin; 15-episode weeks may need a tighter prompt. Tune the compile_digest.py prompt to your taste.

Cost

At ~$0.002/episode (Flash pricing at time of writing) and ~25 episodes/week across 10-ish feeds, weekly cost is under a dollar. The Slack API and YouTube auto-subs are free.

License

MIT. Take what's useful.