Twitter Bookmarks Processor

Automated Twitter bookmark processing inspired by Alex Hillman's JFDI system. Checks bookmarks periodically, categorizes via LLM, and routes to configured destinations.

Quick Start

# Build
go build -o twitter-bookmarks

# Process bookmarks
./twitter-bookmarks process

# Check status
./twitter-bookmarks status

How It Works

1. Fetch - Get new bookmarks via bird bookmarks
2. Analyze - Read full tweet + thread context
3. Classify - Send to Gemini with your category definitions from YAML config
4. Route - Execute action based on category mapping (save, summarize, notify, discard)
5. Track - Update state to avoid reprocessing

Configuration

Create ~/.twitter-bookmarks-config.yaml:

categories:
  # Define categories with descriptions and keywords (guides LLM classification)
  work:
    description: "Work-related tools, articles, and research"
    keywords: [enterprise, b2b, productivity, saas]
  
  personal:
    description: "Personal interests, hobbies, random finds"
    keywords: [hobby, interest, fun, learn]

routing:
  # Map categories to actions
  work:
    action: save_obsidian  # Options: save_obsidian, summarize, notify, save_file, unbookmark
    path: "Work/Research"  # Relative to Obsidian vault
    notify: true           # Send Telegram notification
  
  personal:
    action: summarize      # Extract & summarize URLs
    notify: true

Available Actions

Environment Variables

Required:

• GEMINI_API_KEY - For LLM classification

Optional:

• TWITTER_BOOKMARKS_CONFIG - Config file path (default: ~/.twitter-bookmarks-config.yaml)
• TWITTER_BOOKMARKS_STATE - State file (default: ~/.twitter-bookmarks-state.json)
• TWITTER_BOOKMARKS_OBSIDIAN - Obsidian vault path (if using save_obsidian)
• BIRD_BIN - bird CLI path (default: bird)
• SUMMARIZE_BIN - summarize CLI path (optional)
• TWITTER_BOOKMARKS_QUIET_START - Quiet hours start (default: 23:00)
• TWITTER_BOOKMARKS_QUIET_END - Quiet hours end (default: 08:00)

Automation

Cron (every 20 minutes)

*/20 * * * * cd /path/to/twitter-bookmarks && ./twitter-bookmarks process

Clawdis Cron

from clawdis_cron import add_job
add_job({
    'id': 'twitter-bookmarks',
    'schedule': '*/20 * * * *',
    'command': 'cd /path/to/twitter-bookmarks && ./twitter-bookmarks process'
})

Example Workflows

Workflow 1: Research Collection

categories:
  research:
    description: "Academic papers, research tools, datasets"
    keywords: [research, paper, study, dataset, academic]

routing:
  research:
    action: save_obsidian
    path: "Research/Papers"
    notify: true

Workflow 2: Auto-discard spam

categories:
  spam:
    description: "Unwanted promotional content"
    keywords: [crypto, nft, airdrop, presale]

routing:
  spam:
    action: unbookmark  # Remove automatically

Workflow 3: Recipe collection

categories:
  recipes:
    description: "Cooking recipes and food content"
    keywords: [recipe, cook, bake, ingredient]

routing:
  recipes:
    action: save_file
    path: "/Users/you/recipes.md"
    notify: false

State Tracking

File: ~/.twitter-bookmarks-state.json

{
  "lastProcessed": "2026-01-04T02:20:00Z",
  "processedIds": ["1234567890", "9876543210"],
  "categories": {
    "work": 15,
    "personal": 8,
    "spam": 3
  }
}

Integration with Clawdis

When running as a Clawdis skill, the processor can:

• Send notifications to configured Telegram accounts
• Save to Obsidian vault (if path configured)
• Create Codex prompts for AI-assisted implementation
• Trigger other Clawdis skills based on content

Requirements

• bird CLI (https://github.com/steipete/bird) - Authenticated Twitter access
• Gemini API key - For LLM classification
• Obsidian vault (optional) - Only if using save_obsidian
• summarize CLI (optional) - For URL content extraction
• Telegram bot (optional) - For notifications

Tips

• Start with 3-5 categories max, refine based on actual bookmarks
• Use discard category + unbookmark action to auto-clean spam
• Set notify: false for high-volume categories
• Keywords guide LLM but aren't strict filters
• Check ./twitter-bookmarks status to see category distribution

Inspired by Alex Hillman's JFDI system