digest-skill

内容来源：SKILL.md（标准 Skill 格式） · 原始地址 · 查看安装指南

原始内容

---
name: digest
description: Weekly digest of newsletter and podcast insights — novel, high-impact, contrarian ideas for business, marketing, and LinkedIn. Generates audio version via ElevenLabs.
---

# Weekly Digest

Scan the past week's newsletters and podcast episodes, extract what matters, and deliver a single digest of insights that are novel, high-impact, contrarian, or directly useful for building your business, driving marketing success, and sharing on LinkedIn. Then generate an audio version.

## Important: Writing Style

Follow the writing preferences in your project's `CLAUDE.md`. Direct. Specific. No filler.

## Setup

Before using this skill, create a `config.json` file in the `.claude/skills/digest/` directory. See `config.example.json` for the required structure. The skill reads API keys and service credentials from this file at runtime.

## Step 1: Gather Sources (run in parallel)

Run all of the following simultaneously.

### 1A: Newsletters

Search Gmail for newsletters from the last 7 days:

| Query | Purpose | maxResults |
|-------|---------|------------|
| `label:Substack newer_than:7d` | All Substack newsletters | 100 |
| `label:Newsletter newer_than:7d` | Non-Substack newsletters (beehiiv, etc.) | 100 |

Use your Gmail MCP's `gmail_search_messages` tool for each. (Load the schema first via ToolSearch if needed.)

**Priority senders** (always include if present) — customize this list with your own subscriptions:
- Product Growth (Aakash Gupta) — `aakashgupta@substack.com`
- Lenny's Newsletter — `lenny@substack.com`
- Growth Unhinged (Kyle Poyar) — beehiiv / `mail.beehiiv.com`
- Mostly Growth — substack
- a16z — `a16z@substack.com`
- Not Boring (Packy McCormick) — `notboring@substack.com`
- MKT1 (Emily Kramer) — `mkt1@substack.com`
- GTM Strategist (Maja Voje) — `gtmstrategist@substack.com`
- OnlyCFO — `onlycfo@substack.com`
- Wes Kao — `weskao@substack.com`
- Tech Brew — `techbrew@morningbrew.com`
- The Frontier (Product Hunt) — `hi@deeperlearning.producthunt.com`
- Alex / basicarts — `alex@basicarts.org`

### 1B: Podcast Episodes

Use a **hybrid approach** to find episodes published in the last 7 days.

#### Source 1: RSS Transcript Feeds (preferred — gives full transcripts directly)

These podcasts have transcripts embedded in their RSS feeds. Curl the feed and extract `<podcast:transcript>` tags from recent episodes:

| Podcast | RSS Feed URL | Transcript Format |
|---------|-------------|-------------------|
| Cheeky Pint | `https://feeds.transistor.fm/cheeky-pint-with-john-collison` | `text/plain` (.txt) |
| Think Fast, Talk Smart | `https://feeds.transistor.fm/think-fast-talk-smart-communication-techniques` | `text/vtt` (.vtt) |

For each feed:
```bash
python3 -c "
import xml.etree.ElementTree as ET
from datetime import datetime, timedelta, timezone
import sys, json

xml_data = sys.stdin.read()
# Register namespaces
ns = {
    'itunes': 'http://www.itunes.com/dtds/podcast-1.0.dtd',
    'podcast': 'https://podcastindex.org/namespace/1.0',
}
root = ET.fromstring(xml_data)
channel = root.find('channel')
cutoff = datetime.now(timezone.utc) - timedelta(days=7)
results = []
for item in channel.findall('item'):
    pub_date_str = item.find('pubDate').text if item.find('pubDate') is not None else ''
    try:
        from email.utils import parsedate_to_datetime
        pub_date = parsedate_to_datetime(pub_date_str)
    except:
        continue
    if pub_date >= cutoff:
        title = item.find('title').text if item.find('title') is not None else ''
        transcript_el = item.find('podcast:transcript', ns)
        transcript_url = transcript_el.get('url') if transcript_el is not None else None
        results.append({'title': title, 'published': pub_date_str, 'transcript_url': transcript_url})
print(json.dumps(results, indent=2))
"
```

If a transcript URL is found, fetch it directly with curl. For VTT files, strip timestamps and keep just the text.

#### Source 2: YouTube RSS Feeds (for remaining podcasts)

Fetch YouTube RSS feeds for episodes published in the last 7 days:

```bash
python3 -c "
import xml.etree.ElementTree as ET
from datetime import datetime, timedelta, timezone
import sys, json

xml_data = sys.stdin.read()
ns = {'atom': 'http://www.w3.org/2005/Atom', 'yt': 'http://www.youtube.com/xml/schemas/2015', 'media': 'http://search.yahoo.com/mrss/'}
root = ET.fromstring(xml_data)
cutoff = datetime.now(timezone.utc) - timedelta(days=7)
results = []
for entry in root.findall('atom:entry', ns):
    published = entry.find('atom:published', ns).text
    pub_date = datetime.fromisoformat(published.replace('Z', '+00:00'))
    if pub_date >= cutoff:
        vid_id = entry.find('yt:videoId', ns).text
        title = entry.find('atom:title', ns).text
        results.append({'id': vid_id, 'title': title, 'published': published})
print(json.dumps(results))
"
```

**YouTube feeds — customize this list with your preferred podcasts, in priority order:**

| Priority | Podcast | Feed URL |
|----------|---------|----------|
| 1 | Founders (David Senra) | `https://www.youtube.com/feeds/videos.xml?channel_id=UCy2FPslt0LLPsIV0iukvHpQ` |
| 2 | Lenny's Podcast | `https://www.youtube.com/feeds/videos.xml?channel_id=UC6t1O76G0jYXOAoYCm153dA` |
| 3 | 20VC | `https://www.youtube.com/feeds/videos.xml?channel_id=UCf0PBRjhf0rF8fWBIxTuoWA` |
| 4 | Mostly Growth | `https://www.youtube.com/feeds/videos.xml?channel_id=UC6WQpg2OxjHSpYRHnsM90bw` |
| 5 | a16z | `https://www.youtube.com/feeds/videos.xml?channel_id=UCQ1VQj-37kl2yS_VUhfQHsw` |
| 6 | Masters of Scale | `https://www.youtube.com/feeds/videos.xml?channel_id=UCiemDAS1bXMBTx3jIIOukFg` |
| 7 | This Week in Startups | `https://www.youtube.com/feeds/videos.xml?channel_id=UCkkhmBWfS7pILYIk0izkc3A` |

Curl all feeds in parallel (feed XML fetches are fine in parallel — rate limiting only happens during transcript fetches in Step 3).

**IMPORTANT — Filter out YouTube Shorts:** Before fetching transcripts, check video duration using `yt-dlp`:

```bash
yt-dlp --print duration VIDEO_ID 2>/dev/null
```

Skip any video shorter than 180 seconds (3 minutes) — these are Shorts/clips, not full episodes.

## Step 2: Read Newsletter Content

For each newsletter email from Step 1A, use your Gmail MCP's `gmail_read_message` tool to get the full content.

**Skip these — don't read or include:**
- Substack verification codes
- "Welcome" / subscription confirmation emails
- Substack "Weekly Stack" digest emails (from `no-reply@substack.com`)
- Newsletters clearly outside your business interests (geopolitics, poetry, macro econ, faith/devotional, general news)

**Keep and read everything about:** product, growth, marketing, GTM, startups, AI/tech, SaaS, PLG, building businesses, fundraising, leadership.

If there are more than 15 relevant newsletters, prioritize the priority senders listed in Step 1A.

## Step 3: Get Podcast Transcripts

### For RSS transcript podcasts:
Transcripts were already fetched in Step 1B. Use them directly.

### For YouTube-sourced podcasts:
For each new episode from Step 1B (that passed the duration filter), fetch the English transcript using `youtube-transcript-api`:

```bash
python3 -c "
from youtube_transcript_api import YouTubeTranscriptApi
api = YouTubeTranscriptApi()
transcript = api.fetch('VIDEO_ID')
text = ' '.join([t.text for t in transcript.snippets])
print(text)
" 2>&1
```

Replace `VIDEO_ID` with each video's ID. **Fetch in priority order, not all in parallel** — YouTube rate-limits after ~8 transcript requests in quick succession. If you hit a rate limit (`IpBlocked` error), stop fetching more YouTube transcripts and work with what you have.

If a transcript is very long (60+ minute episode), read just the first 15,000 words.

If the transcript fetch fails due to rate limiting, note the title as "transcript unavailable — rate limited" and move on. Do not retry.

## Step 4: Synthesize the Digest

Analyze all newsletter content and podcast transcripts. Extract and organize insights.

**What counts as signal:**
- Novel frameworks or mental models people haven't seen before
- Contrarian takes that challenge conventional wisdom
- Specific, tactical plays for growing a business (not generic advice)
- Marketing strategies, GTM plays, or positioning insights
- Data points or trends that reveal where markets are heading
- Ideas that would make strong LinkedIn content (provocative, specific, backed by reasoning)
- Concrete examples with names, numbers, and outcomes

**What to filter out:**
- Obvious takes everyone already agrees on
- Self-promotional content from the newsletter author
- Generic advice ("focus on your customer," "iterate fast," "build trust")
- Content outside your domain (pure engineering, policy, macro politics)
- Anything that sounds like a LinkedIn influencer cliche

**Also generate the episode title here.** Format:

```
[Mon DD–DD] | [Strongest claim] · [Topic 2] · [Topic 3]
```

Rules:
- Date range: abbreviated month + day range of the period covered (e.g., `Mar 12–18`)
- First segment: the single sharpest claim from The Signal — short declarative, 4–7 words, no hedging
- Two more topic slugs: specific and concrete, 3–5 words each
- Total title: under 80 characters
- No "weekly digest", no "episode", no generic filler — every word earns its place

Example: `Mar 12–18 | Growth is a trust problem · Kalshi vs. the CFTC · The 8% conversion floor`

This title is used in Step 7B for the RSS feed. Carry it forward.

## Step 5: Present the Written Digest

```
## Weekly Digest — [Date Range]

**Sources scanned:** X newsletters, X podcast episodes

---

### The Signal

[3-7 of the most important insights from this week. Each one gets 2-3 sentences max.
State the insight directly — don't build up to it. Attribute the source.]

---

### Contrarian / Provocative

[1-3 takes that challenge conventional wisdom. These are the ones worth
turning into LinkedIn posts. State why they're contrarian.]

---

### Tactical Takeaways

[3-5 specific, actionable ideas you could apply to your business or share.
The more concrete, the better. "Do X" not "consider X."]

---

### LinkedIn Angles

[2-3 specific content angles from this week's sources that could become a post or carousel.
For each: the hook line, the core argument, and which source it draws from.]

---

### Worth a Full Read

[Any newsletter or podcast episode that was especially dense with insight this week.
One-line reason why.]

---

### Skipped / Thin This Week

[List source names that had nothing relevant. No detail needed.]
```

Do NOT pad the digest. If there's only one contrarian take, list one. If no podcasts published, skip that section. Empty sections get cut entirely.

## Step 6: Generate Audio Digest

After presenting the written digest, generate an audio version covering **Signal + Contrarian + Tactical** sections only.

### 6A: Format for spoken delivery

Reformat those three sections into a spoken script. Rules:
- Remove markdown formatting (headers become spoken transitions like "Here's what mattered this week." / "Now, the contrarian takes." / "Tactical takeaways.")
- Remove attribution brackets — weave source names naturally ("Elena Verna on 20VC made the case that...")
- Keep sentences short and punchy — same voice as the written digest but optimized for ear, not eye
- No bullet points — convert to flowing paragraphs with natural transitions
- Total script should be under 10,000 characters (fits in one ElevenLabs Turbo v2 request)
- If over 10,000 characters, trim the least important tactical takeaway

### 6B: Generate audio via ElevenLabs

Read the config file to get the ElevenLabs API key and voice settings:

```bash
python3 -c "
import json, urllib.request, sys

config = json.load(open('CONFIG_PATH'))
api_key = config['elevenlabs_api_key']

# Read script from stdin
script = sys.stdin.read()

voice_id = config.get('elevenlabs_voice_id', 'NNl6r8mD7vthiJatiJt1')
url = f'https://api.elevenlabs.io/v1/text-to-speech/{voice_id}/stream'

payload = json.dumps({
    'text': script,
    'model_id': 'eleven_turbo_v2',
    'output_format': 'mp3_44100_128',
    'voice_settings': {
        'stability': 0.6,
        'similarity_boost': 0.8,
        'style': 0.15,
        'use_speaker_boost': True,
    },
}).encode()

req = urllib.request.Request(url, data=payload)
req.add_header('xi-api-key', api_key)
req.add_header('Content-Type', 'application/json')
req.add_header('Accept', 'audio/mpeg')

from datetime import date
output_dir = config.get('output_dir', './digests')
output_path = f'{output_dir}/digest-{date.today().isoformat()}.mp3'

import os
os.makedirs(os.path.dirname(output_path), exist_ok=True)

with urllib.request.urlopen(req) as resp:
    with open(output_path, 'wb') as f:
        while True:
            chunk = resp.read(1024)
            if not chunk:
                break
            f.write(chunk)

print(f'Audio saved: {output_path}')
char_count = len(script)
print(f'Characters used: {char_count} (~{char_count} credits)')
" <<'SCRIPT'
INSERT_SPOKEN_SCRIPT_HERE
SCRIPT
```

Replace `CONFIG_PATH` with the path to your `config.json` and `INSERT_SPOKEN_SCRIPT_HERE` with the actual spoken script from Step 6A.

After generating, report:
- File path of the MP3
- Character count / credits used
- Approximate duration (estimate ~150 words per minute)

## Step 7: Publish to Podcast RSS Feed

Upload the MP3 to Cloudflare R2 and update the podcast RSS feed so Spotify and Apple Podcasts auto-ingest the new episode.

### 7A: Upload MP3 to R2

```bash
python3 -c "
import boto3, json
from datetime import date

config = json.load(open('CONFIG_PATH'))
r2 = config['r2']

s3 = boto3.client('s3',
    endpoint_url=r2['endpoint'],
    aws_access_key_id=r2['access_key_id'],
    aws_secret_access_key=r2['secret_access_key'],
    region_name='auto'
)

output_dir = config.get('output_dir', './digests')
today = date.today().isoformat()
local_path = f'{output_dir}/digest-{today}.mp3'
r2_key = f'episodes/digest-{today}.mp3'

s3.upload_file(local_path, r2['bucket'], r2_key, ExtraArgs={'ContentType': 'audio/mpeg'})

import os
file_size = os.path.getsize(local_path)
print(f'Uploaded: {r2[\"public_url\"]}/{r2_key}')
print(f'Size: {file_size} bytes')
"
```

### 7B: Update RSS feed

Download the existing `feed.xml` from R2, add the new episode entry, and re-upload. The feed uses raw string templating (not `xml.etree`) to avoid the duplicate `xmlns:itunes` bug that breaks Spotify/Apple ingestion.

```bash
python3 -c "
import boto3, json, os, re
from datetime import date, datetime, timezone
from email.utils import format_datetime

config = json.load(open('CONFIG_PATH'))
r2 = config['r2']
podcast = config['podcast']

s3 = boto3.client('s3',
    endpoint_url=r2['endpoint'],
    aws_access_key_id=r2['access_key_id'],
    aws_secret_access_key=r2['secret_access_key'],
    region_name='auto'
)

output_dir = config.get('output_dir', './digests')
today = date.today().isoformat()
mp3_key = f'episodes/digest-{today}.mp3'
pub_url = r2['public_url']
mp3_url = f'{pub_url}/{mp3_key}'
local_mp3 = f'{output_dir}/digest-{today}.mp3'
file_size = os.path.getsize(local_mp3)

duration_seconds = int(file_size / 16000)
minutes = duration_seconds // 60
seconds = duration_seconds % 60

pub_date = format_datetime(datetime.now(timezone.utc))

# Build new <item> block
new_item = f'''    <item>
      <title>INSERT_EPISODE_TITLE_HERE</title>
      <description>INSERT_EPISODE_DESCRIPTION_HERE</description>
      <pubDate>{pub_date}</pubDate>
      <guid isPermaLink=\"true\">{mp3_url}</guid>
      <enclosure url=\"{mp3_url}\" length=\"{file_size}\" type=\"audio/mpeg\"/>
      <itunes:duration>{minutes}:{seconds:02d}</itunes:duration>
      <itunes:explicit>false</itunes:explicit>
    </item>'''

# Download existing feed and insert new item after opening <channel> metadata
feed_path = '/tmp/feed.xml'
try:
    s3.download_file(r2['bucket'], 'feed.xml', feed_path)
    with open(feed_path, 'r') as f:
        feed = f.read()
    # Insert new item before the first existing <item>, or before </channel>
    if '<item>' in feed:
        feed = feed.replace('<item>', new_item + '\n    <item>', 1)
    else:
        feed = feed.replace('</channel>', new_item + '\n  </channel>')
except:
    # Create new feed from scratch
    feed = f'''<?xml version=\"1.0\" encoding=\"UTF-8\"?>
<rss version=\"2.0\" xmlns:itunes=\"http://www.itunes.com/dtds/podcast-1.0.dtd\" xmlns:content=\"http://purl.org/rss/1.0/modules/content/\" xmlns:atom=\"http://www.w3.org/2005/Atom\">
  <channel>
    <title>{podcast['title']}</title>
    <link>{pub_url}</link>
    <atom:link href=\"https://pubsubhubbub.appspot.com\" rel=\"hub\"/>
    <atom:link href=\"{pub_url}/feed.xml\" rel=\"self\" type=\"application/rss+xml\"/>
    <description>{podcast['description']}</description>
    <language>en-us</language>
    <itunes:author>{podcast['author']}</itunes:author>
    <itunes:owner>
      <itunes:name>{podcast['author']}</itunes:name>
      <itunes:email>{podcast['email']}</itunes:email>
    </itunes:owner>
    <itunes:explicit>false</itunes:explicit>
    <itunes:image href=\"{pub_url}/cover.jpg\"/>
    <itunes:category text=\"Business\"/>
    <itunes:type>episodic</itunes:type>
{new_item}
  </channel>
</rss>'''

with open(feed_path, 'w') as f:
    f.write(feed)

s3.upload_file(feed_path, r2['bucket'], 'feed.xml', ExtraArgs={'ContentType': 'application/rss+xml'})
print(f'RSS feed updated: {pub_url}/feed.xml')
print(f'New episode: {today} ({minutes}:{seconds:02d})')
"
```

Replace `CONFIG_PATH` with the path to your `config.json`. Replace `INSERT_EPISODE_TITLE_HERE` with the episode title generated in Step 4. Replace `INSERT_EPISODE_DESCRIPTION_HERE` with a 1-2 sentence episode description from the Signal and Contrarian sections. Keep it under 250 characters.

### 7C: Ping WebSub hub

After uploading the feed, ping Google's PubSubHubbub hub so subscribers (Spotify, Apple, YouTube Music) get notified immediately instead of waiting for their next poll cycle.

```bash
curl -s -o /dev/null -w "HTTP %{http_code}" -X POST https://pubsubhubbub.appspot.com/ \
  -d "hub.mode=publish" \
  -d "hub.url=YOUR_FEED_URL"
```

Replace `YOUR_FEED_URL` with the public URL to your `feed.xml` on R2. A 204 response means the hub accepted the ping. Any other code means the ping failed — log it but don't block.

After publishing, report:
- RSS feed URL
- Episode MP3 URL
- WebSub ping status (204 = success)

## Notes

- `boto3` is used for R2 uploads. If missing, install with `pip3 install --user boto3`.
- `youtube-transcript-api` is installed via pip3. If missing, install with `pip3 install --user youtube-transcript-api`.
- `yt-dlp` is used to check video duration. If missing, install with `pip3 install --user yt-dlp`.
- Some podcast channels (especially 20VC and This Week in Startups) publish multiple times per week. Include all episodes but keep summaries tight.
- If a YouTube feed returns a 404 or empty result, the channel ID may have changed. Flag it and move on.
- For a weekly digest with more sources, focus synthesis on quality over quantity. A week with 20 newsletters should still produce the same tight digest format — just with better signal extracted.
- Audio generation costs ~11,000 ElevenLabs credits per episode. At once per week, that's ~44,000 credits/month on the Creator plan.

name: digest description: Weekly digest of newsletter and podcast insights — novel, high-impact, contrarian ideas for business, marketing, and LinkedIn. Generates audio version via ElevenLabs.

Weekly Digest

Scan the past week's newsletters and podcast episodes, extract what matters, and deliver a single digest of insights that are novel, high-impact, contrarian, or directly useful for building your business, driving marketing success, and sharing on LinkedIn. Then generate an audio version.

Important: Writing Style

Follow the writing preferences in your project's CLAUDE.md. Direct. Specific. No filler.

Setup

Before using this skill, create a config.json file in the .claude/skills/digest/ directory. See config.example.json for the required structure. The skill reads API keys and service credentials from this file at runtime.

Step 1: Gather Sources (run in parallel)

Run all of the following simultaneously.

1A: Newsletters

Search Gmail for newsletters from the last 7 days:

Query	Purpose	maxResults
`label:Substack newer_than:7d`	All Substack newsletters	100
`label:Newsletter newer_than:7d`	Non-Substack newsletters (beehiiv, etc.)	100

Use your Gmail MCP's gmail_search_messages tool for each. (Load the schema first via ToolSearch if needed.)

Priority senders (always include if present) — customize this list with your own subscriptions:

Product Growth (Aakash Gupta) — aakashgupta@substack.com
Lenny's Newsletter — lenny@substack.com
Growth Unhinged (Kyle Poyar) — beehiiv / mail.beehiiv.com
Mostly Growth — substack
a16z — a16z@substack.com
Not Boring (Packy McCormick) — notboring@substack.com
MKT1 (Emily Kramer) — mkt1@substack.com
GTM Strategist (Maja Voje) — gtmstrategist@substack.com
OnlyCFO — onlycfo@substack.com
Wes Kao — weskao@substack.com
Tech Brew — techbrew@morningbrew.com
The Frontier (Product Hunt) — hi@deeperlearning.producthunt.com
Alex / basicarts — alex@basicarts.org

1B: Podcast Episodes

Use a hybrid approach to find episodes published in the last 7 days.

Source 1: RSS Transcript Feeds (preferred — gives full transcripts directly)

These podcasts have transcripts embedded in their RSS feeds. Curl the feed and extract <podcast:transcript> tags from recent episodes:

Podcast	RSS Feed URL	Transcript Format
Cheeky Pint	`https://feeds.transistor.fm/cheeky-pint-with-john-collison`	`text/plain` (.txt)
Think Fast, Talk Smart	`https://feeds.transistor.fm/think-fast-talk-smart-communication-techniques`	`text/vtt` (.vtt)

For each feed:

python3 -c "
import xml.etree.ElementTree as ET
from datetime import datetime, timedelta, timezone
import sys, json

xml_data = sys.stdin.read()
# Register namespaces
ns = {
    'itunes': 'http://www.itunes.com/dtds/podcast-1.0.dtd',
    'podcast': 'https://podcastindex.org/namespace/1.0',
}
root = ET.fromstring(xml_data)
channel = root.find('channel')
cutoff = datetime.now(timezone.utc) - timedelta(days=7)
results = []
for item in channel.findall('item'):
    pub_date_str = item.find('pubDate').text if item.find('pubDate') is not None else ''
    try:
        from email.utils import parsedate_to_datetime
        pub_date = parsedate_to_datetime(pub_date_str)
    except:
        continue
    if pub_date >= cutoff:
        title = item.find('title').text if item.find('title') is not None else ''
        transcript_el = item.find('podcast:transcript', ns)
        transcript_url = transcript_el.get('url') if transcript_el is not None else None
        results.append({'title': title, 'published': pub_date_str, 'transcript_url': transcript_url})
print(json.dumps(results, indent=2))
"

If a transcript URL is found, fetch it directly with curl. For VTT files, strip timestamps and keep just the text.

Source 2: YouTube RSS Feeds (for remaining podcasts)

Fetch YouTube RSS feeds for episodes published in the last 7 days:

python3 -c "
import xml.etree.ElementTree as ET
from datetime import datetime, timedelta, timezone
import sys, json

xml_data = sys.stdin.read()
ns = {'atom': 'http://www.w3.org/2005/Atom', 'yt': 'http://www.youtube.com/xml/schemas/2015', 'media': 'http://search.yahoo.com/mrss/'}
root = ET.fromstring(xml_data)
cutoff = datetime.now(timezone.utc) - timedelta(days=7)
results = []
for entry in root.findall('atom:entry', ns):
    published = entry.find('atom:published', ns).text
    pub_date = datetime.fromisoformat(published.replace('Z', '+00:00'))
    if pub_date >= cutoff:
        vid_id = entry.find('yt:videoId', ns).text
        title = entry.find('atom:title', ns).text
        results.append({'id': vid_id, 'title': title, 'published': published})
print(json.dumps(results))
"

YouTube feeds — customize this list with your preferred podcasts, in priority order:

Priority	Podcast	Feed URL
1	Founders (David Senra)	`https://www.youtube.com/feeds/videos.xml?channel_id=UCy2FPslt0LLPsIV0iukvHpQ`
2	Lenny's Podcast	`https://www.youtube.com/feeds/videos.xml?channel_id=UC6t1O76G0jYXOAoYCm153dA`
3	20VC	`https://www.youtube.com/feeds/videos.xml?channel_id=UCf0PBRjhf0rF8fWBIxTuoWA`
4	Mostly Growth	`https://www.youtube.com/feeds/videos.xml?channel_id=UC6WQpg2OxjHSpYRHnsM90bw`
5	a16z	`https://www.youtube.com/feeds/videos.xml?channel_id=UCQ1VQj-37kl2yS_VUhfQHsw`
6	Masters of Scale	`https://www.youtube.com/feeds/videos.xml?channel_id=UCiemDAS1bXMBTx3jIIOukFg`
7	This Week in Startups	`https://www.youtube.com/feeds/videos.xml?channel_id=UCkkhmBWfS7pILYIk0izkc3A`

Curl all feeds in parallel (feed XML fetches are fine in parallel — rate limiting only happens during transcript fetches in Step 3).

IMPORTANT — Filter out YouTube Shorts: Before fetching transcripts, check video duration using yt-dlp:

yt-dlp --print duration VIDEO_ID 2>/dev/null

Skip any video shorter than 180 seconds (3 minutes) — these are Shorts/clips, not full episodes.

Step 2: Read Newsletter Content

For each newsletter email from Step 1A, use your Gmail MCP's gmail_read_message tool to get the full content.

Skip these — don't read or include:

Substack verification codes
"Welcome" / subscription confirmation emails
Substack "Weekly Stack" digest emails (from no-reply@substack.com)
Newsletters clearly outside your business interests (geopolitics, poetry, macro econ, faith/devotional, general news)

Keep and read everything about: product, growth, marketing, GTM, startups, AI/tech, SaaS, PLG, building businesses, fundraising, leadership.

If there are more than 15 relevant newsletters, prioritize the priority senders listed in Step 1A.

Step 3: Get Podcast Transcripts

For RSS transcript podcasts:

Transcripts were already fetched in Step 1B. Use them directly.

For YouTube-sourced podcasts:

For each new episode from Step 1B (that passed the duration filter), fetch the English transcript using youtube-transcript-api:

python3 -c "
from youtube_transcript_api import YouTubeTranscriptApi
api = YouTubeTranscriptApi()
transcript = api.fetch('VIDEO_ID')
text = ' '.join([t.text for t in transcript.snippets])
print(text)
" 2>&1

Replace VIDEO_ID with each video's ID. Fetch in priority order, not all in parallel — YouTube rate-limits after ~8 transcript requests in quick succession. If you hit a rate limit (IpBlocked error), stop fetching more YouTube transcripts and work with what you have.

If a transcript is very long (60+ minute episode), read just the first 15,000 words.

If the transcript fetch fails due to rate limiting, note the title as "transcript unavailable — rate limited" and move on. Do not retry.

Step 4: Synthesize the Digest

Analyze all newsletter content and podcast transcripts. Extract and organize insights.

What counts as signal:

Novel frameworks or mental models people haven't seen before
Contrarian takes that challenge conventional wisdom
Specific, tactical plays for growing a business (not generic advice)
Marketing strategies, GTM plays, or positioning insights
Data points or trends that reveal where markets are heading
Ideas that would make strong LinkedIn content (provocative, specific, backed by reasoning)
Concrete examples with names, numbers, and outcomes

What to filter out:

Obvious takes everyone already agrees on
Self-promotional content from the newsletter author
Generic advice ("focus on your customer," "iterate fast," "build trust")
Content outside your domain (pure engineering, policy, macro politics)
Anything that sounds like a LinkedIn influencer cliche

Also generate the episode title here. Format:

[Mon DD–DD] | [Strongest claim] · [Topic 2] · [Topic 3]

Rules:

Date range: abbreviated month + day range of the period covered (e.g., Mar 12–18)
First segment: the single sharpest claim from The Signal — short declarative, 4–7 words, no hedging
Two more topic slugs: specific and concrete, 3–5 words each
Total title: under 80 characters
No "weekly digest", no "episode", no generic filler — every word earns its place

Example: Mar 12–18 | Growth is a trust problem · Kalshi vs. the CFTC · The 8% conversion floor

This title is used in Step 7B for the RSS feed. Carry it forward.

Step 5: Present the Written Digest

## Weekly Digest — [Date Range]

**Sources scanned:** X newsletters, X podcast episodes

---

### The Signal

[3-7 of the most important insights from this week. Each one gets 2-3 sentences max.
State the insight directly — don't build up to it. Attribute the source.]

---

### Contrarian / Provocative

[1-3 takes that challenge conventional wisdom. These are the ones worth
turning into LinkedIn posts. State why they're contrarian.]

---

### Tactical Takeaways

[3-5 specific, actionable ideas you could apply to your business or share.
The more concrete, the better. "Do X" not "consider X."]

---

### LinkedIn Angles

[2-3 specific content angles from this week's sources that could become a post or carousel.
For each: the hook line, the core argument, and which source it draws from.]

---

### Worth a Full Read

[Any newsletter or podcast episode that was especially dense with insight this week.
One-line reason why.]

---

### Skipped / Thin This Week

[List source names that had nothing relevant. No detail needed.]

Do NOT pad the digest. If there's only one contrarian take, list one. If no podcasts published, skip that section. Empty sections get cut entirely.

Step 6: Generate Audio Digest

After presenting the written digest, generate an audio version covering Signal + Contrarian + Tactical sections only.

6A: Format for spoken delivery

Reformat those three sections into a spoken script. Rules:

Remove markdown formatting (headers become spoken transitions like "Here's what mattered this week." / "Now, the contrarian takes." / "Tactical takeaways.")
Remove attribution brackets — weave source names naturally ("Elena Verna on 20VC made the case that...")
Keep sentences short and punchy — same voice as the written digest but optimized for ear, not eye
No bullet points — convert to flowing paragraphs with natural transitions
Total script should be under 10,000 characters (fits in one ElevenLabs Turbo v2 request)
If over 10,000 characters, trim the least important tactical takeaway

6B: Generate audio via ElevenLabs

Read the config file to get the ElevenLabs API key and voice settings:

python3 -c "
import json, urllib.request, sys

config = json.load(open('CONFIG_PATH'))
api_key = config['elevenlabs_api_key']

# Read script from stdin
script = sys.stdin.read()

voice_id = config.get('elevenlabs_voice_id', 'NNl6r8mD7vthiJatiJt1')
url = f'https://api.elevenlabs.io/v1/text-to-speech/{voice_id}/stream'

payload = json.dumps({
    'text': script,
    'model_id': 'eleven_turbo_v2',
    'output_format': 'mp3_44100_128',
    'voice_settings': {
        'stability': 0.6,
        'similarity_boost': 0.8,
        'style': 0.15,
        'use_speaker_boost': True,
    },
}).encode()

req = urllib.request.Request(url, data=payload)
req.add_header('xi-api-key', api_key)
req.add_header('Content-Type', 'application/json')
req.add_header('Accept', 'audio/mpeg')

from datetime import date
output_dir = config.get('output_dir', './digests')
output_path = f'{output_dir}/digest-{date.today().isoformat()}.mp3'

import os
os.makedirs(os.path.dirname(output_path), exist_ok=True)

with urllib.request.urlopen(req) as resp:
    with open(output_path, 'wb') as f:
        while True:
            chunk = resp.read(1024)
            if not chunk:
                break
            f.write(chunk)

print(f'Audio saved: {output_path}')
char_count = len(script)
print(f'Characters used: {char_count} (~{char_count} credits)')
" <<'SCRIPT'
INSERT_SPOKEN_SCRIPT_HERE
SCRIPT

Replace CONFIG_PATH with the path to your config.json and INSERT_SPOKEN_SCRIPT_HERE with the actual spoken script from Step 6A.

After generating, report:

File path of the MP3
Character count / credits used
Approximate duration (estimate ~150 words per minute)

Step 7: Publish to Podcast RSS Feed

Upload the MP3 to Cloudflare R2 and update the podcast RSS feed so Spotify and Apple Podcasts auto-ingest the new episode.

7A: Upload MP3 to R2

python3 -c "
import boto3, json
from datetime import date

config = json.load(open('CONFIG_PATH'))
r2 = config['r2']

s3 = boto3.client('s3',
    endpoint_url=r2['endpoint'],
    aws_access_key_id=r2['access_key_id'],
    aws_secret_access_key=r2['secret_access_key'],
    region_name='auto'
)

output_dir = config.get('output_dir', './digests')
today = date.today().isoformat()
local_path = f'{output_dir}/digest-{today}.mp3'
r2_key = f'episodes/digest-{today}.mp3'

s3.upload_file(local_path, r2['bucket'], r2_key, ExtraArgs={'ContentType': 'audio/mpeg'})

import os
file_size = os.path.getsize(local_path)
print(f'Uploaded: {r2[\"public_url\"]}/{r2_key}')
print(f'Size: {file_size} bytes')
"

7B: Update RSS feed

Download the existing feed.xml from R2, add the new episode entry, and re-upload. The feed uses raw string templating (not xml.etree) to avoid the duplicate xmlns:itunes bug that breaks Spotify/Apple ingestion.

python3 -c "
import boto3, json, os, re
from datetime import date, datetime, timezone
from email.utils import format_datetime

config = json.load(open('CONFIG_PATH'))
r2 = config['r2']
podcast = config['podcast']

s3 = boto3.client('s3',
    endpoint_url=r2['endpoint'],
    aws_access_key_id=r2['access_key_id'],
    aws_secret_access_key=r2['secret_access_key'],
    region_name='auto'
)

output_dir = config.get('output_dir', './digests')
today = date.today().isoformat()
mp3_key = f'episodes/digest-{today}.mp3'
pub_url = r2['public_url']
mp3_url = f'{pub_url}/{mp3_key}'
local_mp3 = f'{output_dir}/digest-{today}.mp3'
file_size = os.path.getsize(local_mp3)

duration_seconds = int(file_size / 16000)
minutes = duration_seconds // 60
seconds = duration_seconds % 60

pub_date = format_datetime(datetime.now(timezone.utc))

# Build new <item> block
new_item = f'''    <item>
      <title>INSERT_EPISODE_TITLE_HERE</title>
      <description>INSERT_EPISODE_DESCRIPTION_HERE</description>
      <pubDate>{pub_date}</pubDate>
      <guid isPermaLink=\"true\">{mp3_url}</guid>
      <enclosure url=\"{mp3_url}\" length=\"{file_size}\" type=\"audio/mpeg\"/>
      <itunes:duration>{minutes}:{seconds:02d}</itunes:duration>
      <itunes:explicit>false</itunes:explicit>
    </item>'''

# Download existing feed and insert new item after opening <channel> metadata
feed_path = '/tmp/feed.xml'
try:
    s3.download_file(r2['bucket'], 'feed.xml', feed_path)
    with open(feed_path, 'r') as f:
        feed = f.read()
    # Insert new item before the first existing <item>, or before </channel>
    if '<item>' in feed:
        feed = feed.replace('<item>', new_item + '\n    <item>', 1)
    else:
        feed = feed.replace('</channel>', new_item + '\n  </channel>')
except:
    # Create new feed from scratch
    feed = f'''<?xml version=\"1.0\" encoding=\"UTF-8\"?>
<rss version=\"2.0\" xmlns:itunes=\"http://www.itunes.com/dtds/podcast-1.0.dtd\" xmlns:content=\"http://purl.org/rss/1.0/modules/content/\" xmlns:atom=\"http://www.w3.org/2005/Atom\">
  <channel>
    <title>{podcast['title']}</title>
    <link>{pub_url}</link>
    <atom:link href=\"https://pubsubhubbub.appspot.com\" rel=\"hub\"/>
    <atom:link href=\"{pub_url}/feed.xml\" rel=\"self\" type=\"application/rss+xml\"/>
    <description>{podcast['description']}</description>
    <language>en-us</language>
    <itunes:author>{podcast['author']}</itunes:author>
    <itunes:owner>
      <itunes:name>{podcast['author']}</itunes:name>
      <itunes:email>{podcast['email']}</itunes:email>
    </itunes:owner>
    <itunes:explicit>false</itunes:explicit>
    <itunes:image href=\"{pub_url}/cover.jpg\"/>
    <itunes:category text=\"Business\"/>
    <itunes:type>episodic</itunes:type>
{new_item}
  </channel>
</rss>'''

with open(feed_path, 'w') as f:
    f.write(feed)

s3.upload_file(feed_path, r2['bucket'], 'feed.xml', ExtraArgs={'ContentType': 'application/rss+xml'})
print(f'RSS feed updated: {pub_url}/feed.xml')
print(f'New episode: {today} ({minutes}:{seconds:02d})')
"

Replace CONFIG_PATH with the path to your config.json. Replace INSERT_EPISODE_TITLE_HERE with the episode title generated in Step 4. Replace INSERT_EPISODE_DESCRIPTION_HERE with a 1-2 sentence episode description from the Signal and Contrarian sections. Keep it under 250 characters.

7C: Ping WebSub hub

After uploading the feed, ping Google's PubSubHubbub hub so subscribers (Spotify, Apple, YouTube Music) get notified immediately instead of waiting for their next poll cycle.

curl -s -o /dev/null -w "HTTP %{http_code}" -X POST https://pubsubhubbub.appspot.com/ \
  -d "hub.mode=publish" \
  -d "hub.url=YOUR_FEED_URL"

Replace YOUR_FEED_URL with the public URL to your feed.xml on R2. A 204 response means the hub accepted the ping. Any other code means the ping failed — log it but don't block.

After publishing, report:

RSS feed URL
Episode MP3 URL
WebSub ping status (204 = success)

Notes

boto3 is used for R2 uploads. If missing, install with pip3 install --user boto3.
youtube-transcript-api is installed via pip3. If missing, install with pip3 install --user youtube-transcript-api.
yt-dlp is used to check video duration. If missing, install with pip3 install --user yt-dlp.
Some podcast channels (especially 20VC and This Week in Startups) publish multiple times per week. Include all episodes but keep summaries tight.
If a YouTube feed returns a 404 or empty result, the channel ID may have changed. Flag it and move on.
For a weekly digest with more sources, focus synthesis on quality over quantity. A week with 20 newsletters should still produce the same tight digest format — just with better signal extracted.
Audio generation costs ~11,000 ElevenLabs credits per episode. At once per week, that's ~44,000 credits/month on the Creator plan.