Fb Insight Miner

Facebook Fanpage Customer Insight Miner

Purpose

Given a public Facebook fanpage URL, extract deep customer insights from comments to inform content strategy, copywriting, and offer design.

When to trigger

Auto-trigger when user says (Vietnamese or English):

• "phân tích insight fanpage X"
• "đào comment fanpage X"
• "customer research fanpage FB"
• "pain points từ fanpage X"
• "voice of customer FB"
• Or user uses /fb-insight-miner <url> command

Required environment

• Apify MCP connected (verify with mcp__apify__* tools available)
• Python 3 installed (for scripts/ pipeline)
• Optional: Google Drive MCP (mcp__claude_ai_Google_Drive__*) if output=gdrive

⚠️ Windows environment note

When running Python scripts that print Vietnamese/emoji output, prefix with PYTHONIOENCODING=utf-8:

PYTHONIOENCODING=utf-8 python scripts/sample.py ...

Without this, script crashes on UnicodeEncodeError (cp1252 codec).

Inputs needed from user (ask if missing)

1. fanpage_url — e.g. https://www.facebook.com/brandname
2. depth — default standard:
   • light: 20 posts, 500 comments ($2)
   • standard: 50 posts, 1,500 comments ($4)
   • deep: 80 posts, 3,000+ comments ($6)
3. days — time window, default 90
4. output — chat | file | gdrive, default chat

Workflow — Execute in order

Phase 0 — Preflight (verify + cost transparency)

0.1 — Verify URL is a public fanpage, not profile/group

• Call WebFetch or Tavily extract on fanpage URL
• If response looks like personal profile (single name, "City" only, no page metadata) → STOP, ask user to confirm it's really a fanpage
• Apify fb-posts-scraper only works on public PAGES, not personal profiles

0.2 — Cost estimate + confirmation Calculate based on depth:

• posts × $0.005 + comments × $0.0025 + 2 × $0.001 actor starts Present to user and ask for go/no-go before burning credit.

Phase 1 — Collect (Apify MCP)

1.1 — Scrape posts

• Actor: apify/facebook-posts-scraper
• Key input fields (verified from real runs):

  {
    "startUrls": [{"url": "<fanpage_url>"}],
    "resultsLimit": <light=20 / standard=50 / deep=80>,
    "onlyPostsNewerThan": "<days> days",
    "captionText": false
  }
  

• Keep fields via get-actor-output with fields="url,text,likes,comments,shares" to minimize tokens.

1.2 — Rank posts by engagement

• Sort by comments (commentsCount) DESC.
• For light/standard/deep: keep top 20 posts → these give ~80%+ of total comment signal.
• Skip posts with 0 comments (no signal).

1.3 — Scrape comments (top 20 posts)

• Actor: apify/facebook-comments-scraper
• Input:

  {
    "startUrls": [<top20_post_urls>],
    "resultsLimit": 1500,
    "includeNestedComments": true,
    "viewOption": "RANKED_UNFILTERED"
  }
  

• Store dataset ID — Apify output will be >25K tokens, use Python pipeline in Phase 2.

Phase 2 — Sample + Compact (Python scripts)

Apify returns raw JSON often >25K tokens — cannot load directly. Use pipeline:

2.1 — Apply sampling rules

PYTHONIOENCODING=utf-8 python scripts/sample.py <raw.json> "<Page Owner Name>" <filtered.json>

• Page owner name = profile name of admin (scrape from profileName where posted by page)
• Outputs filtered JSON + prints sampling stats
• See references/sampling-rules.md for rule details

2.2 — Convert to compact text

PYTHONIOENCODING=utf-8 python scripts/compact.py <filtered.json> <compact.txt>

• Reduces token usage ~70%
• Groups by post, sorts comments by likes DESC
• Format: [likes|replies|depth|signals] author — text

Phase 3 — Analyze (3-pass AI)

Read the compact.txt into context, then run passes internally (no need to write pass files — Claude does inline):

Pass 1 — Breadth — follow templates/prompt-pass1-breadth.md Cluster into 9 dimensions (references/9-insight-dimensions.md), count frequency, representative quotes.

Pass 2 — Depth — follow templates/prompt-pass2-depth.md Per-dimension sub-themes, ≥10 verbatim quotes each, language patterns.

Pass 3 — Synthesis — follow templates/prompt-pass3-synthesis.md Cross-theme patterns, silent signals (references/silent-signals-guide.md), persona, temporal drift.

Phase 4 — Enrichment (Python + Claude narrative)

4.1 — Run enrichment script

PYTHONIOENCODING=utf-8 python scripts/enrich.py <raw.json> "<Page Owner>" <enrichment.txt>

This produces 6 micro-reports:

1. Admin voice decoding (page owner's own replies)
2. Top commenters — DM outreach list
3. Deep thread conversations
4. N-gram analysis — repeat VoC phrases
5. Tag-a-friend mechanic detection
6. Emoji-emotion fingerprint per post

4.2 — Synthesize novel findings Read enrichment.txt, identify insights NOT covered in Phase 3 report. Focus on:

• What repeat bigrams/trigrams reveal about VoC
• Superfans → DM outreach candidates
• Admin voice patterns (short vs teaching vs funnel)
• Tag-a-friend viral mechanics
• Emotional fingerprints per post type

Phase 5 — Report

Use templates/insight-report.md structure, filled with outputs from Phase 3 + Phase 4.

Save per output:

• chat: print full report in conversation
• file: write to CWD as fb-insight-<fanpage-slug>-<YYYYMMDD>.md
• gdrive: base64 encode + call mcp__claude_ai_Google_Drive__create_file

Cost transparency

• Always estimate before Phase 1
• Always report actual cost after Phase 1 (posts_count × $0.005 + comments_count × $0.0025 + actor_starts × $0.001)

Failure modes & edge cases

• Fanpage private/deleted → stop, report to user
• URL is personal profile → stop (actor can't scrape profiles)
• Fanpage has <30 posts in window → suggest depth=light (avoid waste)
• <50 comments total scraped → warn insight will be shallow; proceed with degraded confidence
• Apify MCP not connected → explain how to connect
• Apify output >25K tokens (always) → use Python pipeline in Phase 2 (never Read raw JSON directly)
• Claude context near limit during Phase 3 → reduce sampling cap from 1500 to 800
• ⚠️ Adult/sensitive-content fanpage → posts scrape OK but comments return near-zero (age-gated/adult-hint filtering). Detected in real run on sex-ed academy: 20 posts scraped fine, only 2/2000 comments returned from 2 different actors. Detection heuristic: if comment scrape returns <5% of expected comments on first 3 URLs → STOP and tell user. Niches that hit this: sexual wellness, dating-explicit, adult education. Workaround: none via public API. Need manual cookies-based scraper OR pivot to IG/TikTok where filtering differs.

Output language

Default Vietnamese unless fanpage content is primarily English. Keep verbatim quotes in ORIGINAL language — NEVER translate them.

Files in this skill

fb-insight-miner/
├── SKILL.md                        # This file
├── scripts/
│   ├── sample.py                   # Phase 2.1 — sampling rules
│   ├── compact.py                  # Phase 2.2 — JSON→text
│   └── enrich.py                   # Phase 4.1 — 6 micro-reports
├── templates/
│   ├── insight-report.md           # Phase 5 report structure
│   ├── prompt-pass1-breadth.md     # Phase 3 pass 1
│   ├── prompt-pass2-depth.md       # Phase 3 pass 2
│   └── prompt-pass3-synthesis.md   # Phase 3 pass 3
└── references/
    ├── 9-insight-dimensions.md     # 9-axis framework
    ├── sampling-rules.md           # Sampling logic details
    └── silent-signals-guide.md     # Advanced signals