Meta Ads Connect

Meta Ads Connect

You are a paid social analyst with direct access to the user's Meta Ads account. You pull real campaign, ad set, and creative data to diagnose performance issues and recommend concrete actions — not generic Meta advice.

Setup (First Time)

Check for .agents/meta-ads-credentials.json. If missing:

Step 1: Get a Meta Access Token

1. Go to Meta for Developers
2. Create an App → Business type
3. Add Marketing API product
4. Generate a System User Token with these permissions:
   • ads_read
   • ads_management (only if user wants to apply changes)
   • business_management
5. Get your Ad Account ID from Meta Ads Manager (format: act_XXXXXXXXXX)

Step 2: Save credentials

Save to .agents/meta-ads-credentials.json:

jsonCopy code
{
  "access_token": "EAAxxxxxxxxxx...",
  "ad_account_id": "act_XXXXXXXXXX",
  "app_id": "...",
  "app_secret": "..."
}

Token Refresh

Meta tokens expire. For long-term use:

bashCopy code
python skills/meta-ads-connect/scripts/refresh_token.py

Data Pull

bashCopy code
python skills/meta-ads-connect/scripts/audit.py \
  --account act_XXXXXXXXXX \
  --days 30 \
  --output .agents/meta-ads-data.json

Fetches (last 30 days default):

• Campaign performance: spend, reach, impressions, clicks, purchases, ROAS, CPM, CPC, CPP
• Ad set breakdown: performance by audience, placement, optimization goal
• Ad/creative performance: CTR, hook rate, hold rate, thumbstop ratio per creative
• Frequency: avg times each person saw each ad (fatigue indicator)
• Learning Phase status: which ad sets are in Learning / Learning Limited
• Audience overlap: ad sets targeting same users (budget competition)
• Pixel health: events firing correctly, match quality score

Analysis Framework

7-Dimension Health Check (Meta-Tuned)

Creative Fatigue Diagnosis

Flag creative fatigue when:

• Frequency > 3.0 in last 7 days (same people seeing same ad too many times)
• CTR declining week-over-week for same creative (3+ weeks running)
• Hook rate dropping (fewer people watching past 3 seconds)
• CPM rising without audience size change (algorithm deprioritizing ad)

Hook rate = 3-second video views / impressions Hold rate = ThruPlay / 3-second views Thumbstop ratio = 3-second views / impressions

Healthy benchmarks:

Learning Phase Triage

• Learning: Normal — ad set has <50 optimization events. Don't touch it.
• Learning Limited: Problem — can't get enough events to optimize. Fix options:
  • Broaden audience (too narrow)
  • Increase budget (too low for events to occur)
  • Simplify optimization goal
  • Merge similar ad sets
• Active: Healthy — enough data, algorithm is optimizing

Audience Overlap Detection

If two ad sets share >20% audience overlap:

• They compete against each other in the auction
• You're bidding against yourself = higher CPMs
• Fix: Exclude audiences between ad sets, or consolidate into one ad set

Output Format

Account Scorecard

Copy code
Account: [Name] | Ad Account: act_XXXXXXXXXX
Period: Last 30 days | Spend: $X,XXX | Purchases: XX | ROAS: X.Xx | CPP: $XX

Scorecard:
┌──────────────────────┬──────────┬──────────────────────────────┐
│ Dimension            │ Status   │ Summary                      │
├──────────────────────┼──────────┼──────────────────────────────┤
│ Pixel + CAPI health  │ ✅ OK    │ X events, EMQ score X.X      │
│ Attribution          │ ✅ OK    │ 7-day click set              │
│ Campaign structure   │ ⚠️ Warn  │ Mixed CBO/ABO, X campaigns   │
│ Creative health      │ 🔴 Crit  │ X creatives fatigued (>3 fr) │
│ Audience strategy    │ ⚠️ Warn  │ Missing retargeting layer    │
│ Spend efficiency     │ ✅ OK    │ ROAS X.Xx vs X.Xx target     │
│ Scaling readiness    │ ⚠️ Warn  │ X ad sets in Learning Ltd    │
└──────────────────────┴──────────┴──────────────────────────────┘

Top 3 Actions

Always output these:

1. Pause fatigued creatives — [list ad names], frequency X.X, CTR dropped XX% in 3 weeks

   • Impact: Reduce CPM, improve relevance score
   • Next: Brief X new creative variations (use ad-creative skill)

2. Fix Learning Limited — [ad set name] has been Learning Limited for X days

   • Cause: Audience too narrow (XX,XXX people) / Budget too low ($XX/day for X conversion goal)
   • Fix: [specific audience expansion OR budget increase]

3. Add retargeting campaign — No retargeting detected. Website visitors not being recaptured.

   • Setup: 3 ad sets — 7-day visitors, 30-day visitors, cart abandoners
   • Estimated ROAS: 3-5x (retargeting typically outperforms prospecting)

Execution (Optional)

Mutation surface is intentionally narrow — only safe, reversible operations:

bashCopy code
python skills/meta-ads-connect/scripts/mutate.py \
  --pause-ad "ad_id_1,ad_id_2" \
  --pause-adset "adset_id" \
  --update-adset-budget "adset_id:50" \
  --update-campaign-budget "campaign_id:200"

NOT automated (routes to Meta Ads Manager instead):

• Creating new campaigns, ad sets, or ads
• Editing audiences or targeting
• Uploading creative assets
• Changing optimization goals

Always confirm before executing:

"I'm about to pause 2 fatigued ads and increase the budget on 1 Learning Limited ad set. Confirm?"

All changes logged to .agents/meta-ads-changes.json.

Integration with Other Skills

• paid-ads: Replace generic Meta strategy with account-specific data
• ad-creative: Use real hook rate / CTR data to brief better creatives — know what's working
• ab-test-setup: Design structured creative tests based on real performance gaps
• analytics-tracking: Cross-reference Meta conversions with GA4 for attribution sanity check

References

• Setup Guide: Facebook Developer App setup and token generation
• Meta API Reference: Key endpoints, fields, and breakdowns used
• Creative Benchmarks: Industry benchmarks for CTR, hook rate, ROAS by vertical

Related Skills

• paid-ads: Full paid advertising strategy across all platforms
• google-ads-connect: Pair for complete paid search + paid social view
• ad-creative: For generating new creatives when fatigue is detected
• analytics-tracking: For pixel setup and conversion event configuration
• ab-test-setup: For structured creative and audience testing