AI Content Curator & Publisher

🔥 Generated by this AI Agent → Blog · Substack · Telegram

Stateless pipeline: reads your Telegram channels, LLM curates a digest, publishes to Telegram, Ghost, and Substack.

How it works

1. Collect — Reads all subscribed broadcast channels → /tmp/telegram_messages.json
2. Curate — Two parallel LLM calls generate platform-specific digests:
   • Messenger digest (short, emoji, links) for Telegram
   • Blog digest (long-form HTML) for Ghost and Substack
3. Publish — Posts to all platforms in parallel:
   • Ghost (main blog) — Lexical format, newsletter emails via Mailgun
   • Telegram — plain text to channel
   • Substack (secondary) — ProseMirror format

Runs daily at 6:00 UTC via GitHub Actions. No persistent state.

Setup

1. Telegram credentials

1. Go to https://my.telegram.org → API development tools
2. Create a new application, copy api_id and api_hash
3. Generate a session string:

pip install telethon
export TELEGRAM_API_ID="your_api_id"
export TELEGRAM_API_HASH="your_api_hash"
export TELEGRAM_PHONE="+your_phone_number"
python agent/integrations/telegram/setup_session.py

4. Create a target channel: Menu → New Channel → make it Public → set a username

2. Ghost setup

1. Deploy Ghost using infra/ghost/deploy-ghost.sh (Docker Compose on any Linux server)
2. Create an admin account at https://your-domain/ghost
3. Settings → Integrations → Add custom integration → copy the Admin API key
4. (Optional) Settings → Mailgun → configure for newsletter emails

3. Substack credentials

See agent/integrations/substack/README.md for how to get your connect.sid cookie.

4. GitHub secrets & variables

Secrets:

Secret	Description
TELEGRAM_API_ID	Numeric API ID from my.telegram.org
TELEGRAM_API_HASH	API hash string from my.telegram.org
TELEGRAM_SESSION_STRING	Output from setup_session.py
CLAUDE_CODE_OAUTH_TOKEN	Claude Code OAuth token (primary LLM auth)
ANTHROPIC_API_KEY	Claude API key (fallback LLM auth)
GHOST_ADMIN_API_KEY	Ghost Admin API key (format: {id}:{secret})
SUBSTACK_COOKIE	Substack connect.sid cookie value

Variables:

Variable	Description
TELEGRAM_PUBLISH_CHANNEL	Target Telegram channel (e.g. @my_ai_digest)
GHOST_URL	Ghost domain without https:// (e.g. aicmo.blog)
SUBSTACK_PUBLICATION_URL	Substack publication URL (e.g. https://howai.substack.com)
LLM_PROVIDER	LLM provider: claude, openai, gemini, vertex (default: claude)
LLM_MODEL	Model name (default: claude-sonnet-4-6)
LLM_MAX_TOKENS	Max tokens (default: 12288)

5. Run

Daily digest: Runs automatically at 4:00 UTC. Manual: Actions → "Generate Daily Digest" → Run workflow. Supports workflow dispatch inputs: start_date, end_date, and per-platform publish toggles.

Weekly digest: Runs automatically Saturday at 6:00 UTC. Manual: Actions → "Generate Weekly Digest" → Run workflow. Reads published daily digests from the publish channel, resolves t.me source links, and synthesizes a thematic weekly roundup. Supports start_date, end_date, and per-platform publish toggles.

6. Backfill

To backfill past digests one day at a time:

.github/scripts/backfill-digests.sh 2026-02-08 2026-03-09

Project structure

agent/integrations/
  telegram/
    telegram.py              # Telegram I/O (--read, --post, --list-channels)
    setup_session.py         # one-time auth → StringSession
    requirements.txt         # telethon, cryptg
  ghost/
    ghost.py                 # Ghost publisher (--post, --draft, HTML→Lexical)
    requirements.txt         # requests
  substack/
    substack.py              # Substack publisher (--post, --draft, HTML→ProseMirror)
    requirements.txt         # requests
    README.md                # Substack auth setup guide
infra/
  hetzner-setup.sh           # one-click VPS setup (see below)
  ghost/
    deploy-ghost.sh          # deploy Ghost (Docker Compose)
    docker-compose.yml       # Caddy + Ghost + MySQL
  ubuntu_security_hardening.sh  # server hardening (SSH, UFW, fail2ban, sysctl)
.github/
  scripts/
    backfill-digests.sh           # backfill past digests sequentially
  prompts/
    generate-digest-daily-messenger.md   # Daily Telegram prompt (short, emoji, links)
    generate-digest-daily-blog.md        # Daily blog prompt (long-form HTML)
    generate-digest-weekly-messenger.md  # Weekly Telegram prompt (thematic roundup)
    generate-digest-weekly-blog.md       # Weekly blog prompt (thematic HTML, source links)
  workflows/
    generate-digest-daily.yml     # daily pipeline (4:00 UTC)
    generate-digest-weekly.yml    # weekly pipeline (Saturday 6:00 UTC)
    test-ghost.yml                # manual Ghost test
    test-substack.yml             # manual Substack test

Infrastructure

One-click hardened VPS setup with infra/hetzner-setup.sh:

export HCLOUD_TOKEN="..."
export TAILSCALE_API_KEY="..."
bash infra/hetzner-setup.sh my-server

Connection stack: Tailscale + Mosh/SSH + Termius + sshid.io hardware keys

• Provisions Ubuntu 24.04 on Hetzner Cloud via API + cloud-init
• No public SSH — all access through Tailscale mesh VPN
• Cloudflare-only HTTP/S, Docker + Compose, fail2ban, unattended upgrades
• Server auto-appears in Tailscale and ~/.ssh/config
• Connect from any device: mosh evios@my-server or via Termius