Facebook Group Monitor Skill

Overview

Playwright-based headless browser scraper for Facebook groups. Default behavior: captures a stitched "feed strip" (scrolls 3× viewport, crops to feed column, stitches into 1 JPEG) → agent calls vision model once with a custom prompt to extract and interpret all new posts in a single LLM call.

Requires one-time manual login via terminal to establish a persistent browser session.

Setup

See references/SETUP.md for installation and first-time login instructions.

File Locations (after install)

• Script: scripts/fb-group-monitor.py
• Shell wrapper: scripts/fb-group-monitor.sh
• Browser session: auto-created at scripts/.browser-data/ (persistent login)
• Seen posts: auto-created at scripts/.seen-posts.json (dedup tracking)
• Screenshots: scripts/screenshots/ or custom --shots-dir

Commands

1. Check login session

scripts/fb-group-monitor.sh status

Output:

{"success": true, "action": "status", "message": "Session active — logged in to Facebook."}

2. Scrape new posts (with feed strip screenshot)

scripts/fb-group-monitor.sh scrape <GROUP_URL> [--limit N] [--shots-dir <PATH>]

Parameters:

• GROUP_URL: Full URL (https://www.facebook.com/groups/123456) or just group ID
• --limit N: Max posts to scrape (default: 10)
• --shots-dir <PATH>: ⚠️ REQUIRED for vision — save screenshots in agent workspace so image tool can read them
• --no-shots: Skip screenshots (faster, text-only mode)

Example:

scripts/fb-group-monitor.sh scrape "https://www.facebook.com/groups/123456789" --limit 10 --shots-dir ./temp-screenshots

Output JSON (with feed strip screenshot):

{
  "success": true,
  "action": "scrape",
  "group_name": "Example Group Name",
  "group_url": "https://www.facebook.com/groups/123456789",
  "total_scraped": 6,
  "new_count": 3,
  "feed_screenshot": "/path/to/temp-screenshots/feed_abc12345_1741800000.jpg",
  "posts": [
    {
      "author": "Poster Name",
      "text": "Post content (may be truncated by Facebook's 'See more')...",
      "url": "https://facebook.com/groups/123456/posts/789",
      "images": 3
    },
    {
      "author": "Another Poster",
      "text": "Text-only post, no images...",
      "url": "https://facebook.com/groups/123456/posts/790",
      "images": 0
    }
  ],
  "message": "Found 3 new posts / 6 total."
}

Note: feed_screenshot is a single stitched JPEG covering the full feed (3 viewport scrolls). Individual posts no longer have a screenshot_path — use feed_screenshot (top-level) instead.

3. Clean old screenshots

scripts/fb-group-monitor.sh clean-shots

Auto-removes screenshots older than 48h and caps at 100 max. Script also auto-cleans before each scrape.

4. Login (one-time, from terminal)

scripts/fb-group-monitor.sh login

Opens browser, login manually, press Enter to save session. Session lasts weeks/months.

Agent Workflow

When triggered by cron or user request to check a group:

Step 1: Scrape

scripts/fb-group-monitor.sh scrape "<GROUP_URL>" --limit 10 --shots-dir ./temp-screenshots

⚠️ Use --shots-dir pointing to a path within the workspace so the image tool can access screenshots.

Step 2: Analyze feed screenshot + combine with text

If feed_screenshot is present (1 stitched image covering the full feed):

• Compose a dynamic prompt using the author list from the JSON, so vision focuses on the correct new posts (the feed strip also contains already-seen posts)
• Call image tool with both image and prompt parameters:

  image tool:
    image: <feed_screenshot path>
    prompt: "Facebook group feed. Find and analyze posts by these people (new posts to process):
    - [Author 1]: "[first few words of text from JSON]" — N images
    - [Author 2]: "[first few words of text from JSON]"
    ...
    For EACH post, do 2 things:
    1. EXTRACT: book/product name, author/publisher (read from cover), price (including handwritten prices), condition (new/used/marked/yellowed).
    2. INTERPRET: is the deal worth it (price vs market if known), is the item rare/valuable, actual condition assessment from photos."
  

• ⚠️ NEVER call image tool with empty {} — default prompt just describes the image generically
• ⚠️ Feed strip includes old posts — always include author + text snippet so vision targets the right posts
• After getting vision result: match by author, add url from JSON
• Prioritize image content over truncated text field ("see more" posts)

If feed_screenshot is absent (screenshot failed):

• Read directly from text field of each post
• Summarize available information, note "no screenshot available"

Step 3: Report to user

MUST include the original post link in each item for verification.

Format per new post:

📌 *[Title/Summary]*
👤 Author: [author]
💰 Price: [price or "unknown"]
📝 Condition: [description]
💡 Note: [deal assessment or insight, if any]
🔗 [post URL]

Step 4: If no new posts → stay silent (no notification)

Step 5: If success == false → report error briefly

⚠️ Anti-duplicate rule

• Send Telegram only ONCE per run, regardless of whether vision succeeded or failed
• If message tool returns an error → report once and stop, do NOT retry
• If vision fails → fallback to text, send once with note "⚠️ Vision unavailable, text only"

Important Notes

• Rate limiting: Recommended cron interval ≥ 2 hours — safe for Facebook
• Session expired: If error "Not logged in" → run login from terminal
• UI changes: Facebook updates DOM frequently → selectors may need updates
• "See more": Text is truncated — feed strip screenshot usually has complete content
• Feed strip: Scrolls 3× viewport height, crops to feed column, stitches 1 JPEG — reduces vision calls from N to 1
• Screenshot quality: JPEG 82% — sufficient for LLM vision to read text and identify product images
• Vision model: Use a vision-capable model (e.g. Gemini Flash) for the image tool call