Adapted from the GoMarble Paid Media Claude Skills repository (https://github.com/gomarble-ai/gomarble-ai-paid-media-claude-skills) by its author, shared under GoMarble's attribution license. Works with any data source: a connected ads MCP such as GoMarble, CSV exports, screenshots, or pasted tables.

A consolidated set of skills for analyzing and optimizing Meta Ads (Facebook + Instagram) accounts. These skills are framework-agnostic - the analytical logic (Pareto, Hook/Hold scenarios, decision matrices, guardrails) applies regardless of where the data comes from.

The skills define what to know, what data to look for, how to interpret it, and how to recommend action. They do not depend on any specific tool integration.

Read-and-recommend only. These skills never make changes to the user's Meta Ads account. Claude reads data and produces recommendations in plain language; the user applies the recommended changes themselves in Meta Ads Manager.

---

Data Sources Are Not All Equal

Data can come from four kinds of sources, but they have very different completeness. Some skill rules silently break on weaker sources. Always identify the source first and consult the Data Source Compatibility Matrix below before applying any rule.

Priority	Source	Notes
1 (preferred)	A connected MCP server exposing Meta Ads data (e.g. GoMarble, Meta official)	Returns raw API fields, action arrays, full targeting, breakdowns. All rules work.
2	A CSV/Excel export from Ads Manager	Most rules work, but a few are inert - see compatibility matrix. Several signals require separate exports (breakdowns, video metrics, ad-level vs ad-set-level views).
3 (narrow questions only)	Screenshots of report tables	Use only for spot questions on metrics that are visibly captured in the screenshot. Skip deep analysis.
4 (last resort)	Direct copy-paste of metric values	Treat as user-asserted; ask for the underlying export when stakes are real.

---

How to Use These Skills

1. Identify the user's intent - are they asking for an audit, a performance review, a creative deep-dive, or a targeted question? Use the routing table below.
2. Run the Data Inventory step (Step 0 below) - confirm what source is being used and what's actually available before invoking any skill.
3. Gather the required data - ask the user to share the data the skill needs in their preferred source. The "Minimum required data" section of each skill lists what's needed.
4. Apply the analytical framework - classify, diagnose, decide based on the rules in the relevant skill.
5. Always check guardrails before recommending any action.
6. Output recommendations in the structured format defined per skill.

If required data is missing or unverifiable, say so explicitly and ask the user for it. Never fabricate metrics or invent thresholds.

Always work from current data - campaigns change frequently (budget edits, paused ads, new creatives). If the user shared data more than a day or two ago, ask for a fresh pull before making recommendations.

---

Step 0: Data Inventory (Mandatory Before Any Skill)

Run this once at the start of every analysis. It takes one short turn and prevents wasted effort.

1. Identify the data source - MCP / CSV / screenshot / copy-paste. If MCP is connected, use it. If the user attached a file, inspect its columns/fields before asking for more.
2. State what's available and what's missing for the requested skill, based on the Data Source Compatibility Matrix below.
3. For each missing signal, decide:
   • Blocking - cannot proceed without it; ask user to fetch.
   • Downgrade - can proceed with caveats; flag as low-confidence in the output.
   • Skip - rule doesn't apply to this source (e.g. Purchase De-Duplication on CSV); note it and move on.
4. Confirm the date range in the data matches what the skill expects (typically 30 days; 7d for ad set analysis; 60–90d for trend or Hold Rate baseline).

Output one short paragraph before applying the skill. Example:

"Using your GoMarble MCP. I have ad-level insights with actions and purchase_roas for the last 30 days - sufficient for Performance Analysis (Skill 5). The EMQ check (Skill 1) requires you to look in Events Manager directly; I'll flag it as a manual step."

---

Data Source Compatibility Matrix

Use to plan which signals you can get from which source before running any skill.

Signal	MCP (GoMarble / Meta / etc.)	CSV (Ads Manager export)	Screenshot
Spend, impressions, clicks, CTR, CPM, reach, frequency	✅ direct	✅ default columns	✅ if shown
Purchase count (deduplicated)	✅ via actions array - apply Purchase De-Duplication rule	✅ "Purchases" column (Meta already deduplicates - skip the dedup rule)	⚠️ if shown
Purchase ROAS	✅ purchase_roas field	✅ "Purchase ROAS" column (must be enabled)	⚠️
Revenue / conversion value	✅ action_values	✅ "Purchases Conversion Value" column	⚠️
Video Plays	✅ video_play_actions	✅ "Video Plays" column	❌ rarely shown
3-Second Video Plays	✅ via actions (video_view action)	✅ "3-Second Video Plays" column (already pre-computed - skip the action-name distinction)	❌ rarely shown
ThruPlays	✅ via actions	✅ "ThruPlays" column	❌ rarely shown
Hook Rate / Hold Rate	❌ must compute from the above	❌ must compute from the above	❌
Promoted object / custom_event_type	✅ from ad set object	❌ NOT in default CSV exports - ask user to confirm per ad set / campaign	❌
Campaign objective, ad set bid strategy, bid amount	✅ direct	⚠️ requires "Bid Strategy" + "Bid Amount" columns enabled (often not default)	⚠️
Full targeting object (audiences, interests, custom audiences)	✅ via ad set object	❌ NOT in CSV - must be described by user or pulled via MCP	⚠️ if visible
EMQ Score	❌ not in standard insights APIs - manual check in Events Manager	❌ same	❌ same
Audience Segment breakdown (New / Engaged / Existing)	✅ insights with breakdowns=['user_segment_key']	⚠️ requires a separate breakdown export	❌
Placement breakdown (Feed / Stories / Reels / etc.)	✅ breakdowns=['publisher_platform','platform_position']	⚠️ requires separate breakdown export	❌
Age / gender breakdown	✅ breakdowns=['age','gender']	⚠️ separate export	❌
Geographic / device breakdown	✅ via breakdowns	⚠️ separate export	❌
Ad creative content (video file, image, copy)	✅ via creative endpoint (e.g. facebook_get_ad_creative_details)	❌ never in CSV - must be shared separately	⚠️
Comments / sentiment	❌ requires manual check in Ads Manager or the Page	❌ same	⚠️ if shown
Attribution window settings	✅ via ad set attribution_spec	⚠️ requires "Attribution Setting" column	❌
Daily time-series (for trend / fatigue detection)	✅ insights with time_increment=1	⚠️ requires day-by-day export	❌
3rd-party tracker config (Profitmetrics / Triple Whale etc.)	✅ inferred from custom_event_type on PURCHASE campaign objective	❌ must ask user explicitly - see upgraded check in Skill 5	❌

✅ = available and reliable. ⚠️ = available but requires extra export / setup. ❌ = not available from this source.

---

Meta CSV Column Translation

When the source is a CSV export from Ads Manager, the column names rarely match the API field names used throughout the skills. Use this table to translate. (Column names vary slightly by language and export template - match by meaning, not by exact string.)

Skill / API reference	Typical Ads Manager CSV column
spend	"Amount spent (USD)" / "Amount Spent" / "Spend"
impressions	"Impressions"
reach	"Reach"
frequency	"Frequency"
cpm	"CPM (cost per 1,000 impressions)"
ctr	"CTR (all)" or "CTR (link click-through rate)"
clicks	"Link clicks" or "Clicks (all)"
purchase / omni_purchase (deduped)	"Purchases" (Meta has already deduplicated)
purchase_roas	"Purchase ROAS (return on ad spend)"
Revenue (conversion value)	"Purchases Conversion Value"
video_view action (3-second views)	"3-Second Video Plays" (already pre-computed)
ThruPlays	"ThruPlays"
Video Plays	"Video Plays"
Leads	"Leads" or "On-Facebook Leads" + "Off-Facebook Leads" - confirm with user which is the conversion event
Campaign objective	"Campaign Objective" (often not in default export - must enable)
Bid strategy	"Bid Strategy" (often not in default export - must enable)
Daily / lifetime budget	"Budget" or "Campaign Budget" / "Ad Set Budget"
Attribution window	"Attribution Setting"
custom_event_type / promoted_object	❌ NOT in CSV - ask user to confirm

Rule for Claude: if a CSV column maps to a pre-computed metric (Purchases, 3-Second Video Plays, ThruPlays), use it directly and skip the upstream dedup / action-array rules. Those rules exist to clean raw API data.

---

Routing - Which Skill to Use

User asks about	Use Skill
"Audit my account" / full optimization request	Run Initial Health Check → Conversion Metric Focus → Performance Analysis → Creative Analysis in that order
Pixel / tracking setup, audience segments setup	Initial Health Check
"What's my account type?" / which KPI to use	Conversion Metric Focus
Campaign structure, primary vs testing, bid splits, audience splits	Campaign Structure Analysis
Ad set targeting, budget utilization, audience comparison	Ad Set Analysis
Top performers, Pareto analysis, which ads to keep/pause	Performance Analysis
Hook rate, hold rate, video diagnostics, why creative isn't working	Creative Analysis
Change requests (pause campaign, increase budget, change CTA, etc.)	Apply the relevant analysis skill first, then recommend the change in plain language with rationale. Claude does not execute changes on the user's ad account - the user applies recommendations manually in Ads Manager.

If the user's intent is ambiguous, ask which to focus on first.

---

Universal Guardrails - Apply Before Every Recommendation

Read these before generating any recommendation.

Budget Reallocation Prohibition

You cannot allocate, shift, increase, reduce, or scale budget for: individual ads (only campaigns and ad sets have budgets), placements (Meta controls distribution), demographics (age, gender, location breakdowns), user segments (new, existing, engaged).

Before writing any recommendation, ask: Does it imply moving money between ads, placements, demographics, or segments? If yes → rewrite using the valid controls below.

Valid budget controls (only three):

1. Campaign budget (CBO) or ad set budget (ABO)
2. Bid caps / cost caps
3. Turning ads / ad sets on or off

If a recommendation does not map to one of these three controls, it is not actionable.

Mandatory rewrites:

• ❌ "Reallocate $X from [ad A] to [ad B]"

• ✅ "Pause [ad A]. Increase the campaign/ad set budget by 15–20% to give [ad B] more spend opportunity."

• ❌ "Allocate more to [placement / demographic / segment]"

• ✅ "Create a new ad set targeting [dimension] with its own budget, OR exclude underperforming [dimension] via ad set placement settings."

Rule: Breakdowns (age, gender, placement, device, region, user segment) show how Meta distributed YOUR budget. They are not levers the user controls. The only way to influence them is creating new ad sets with specific targeting or placement selections.

Metric Selection by Conversion Type

• Ecommerce (purchases / revenue): Use ROAS, revenue, AOV, cost per purchase. Never use ROAS for leadgen.
• Lead gen (form submits, leads): Use CPL, cost per qualified lead. Never use ROAS unless offline revenue is mapped.
• App (installs, in-app events): Use CPI, cost per activation. Never use ROAS unless in-app revenue tracking is confirmed.
• Custom conversion: Use cost per result for that specific event. Do not mix in other conversion types.
• Unclear: Ask the user. Do not assume.

User Segment Targeting Notes

Segments (New Audience, Existing Customers, Engaged Audience) are Meta internal classifications. They cannot be targeted without custom audiences.

• Before recommending segmentation, ask: "Do you have a customer custom audience already built?"
• If yes → suggest exclusions or separate campaigns using custom audience targeting
• If no → suggest building custom audiences first

Top Performer Protection (CRITICAL)

Never recommend pausing the top-converting ad in an ad set. If Meta's algorithm allocates 60–80% of an ad set's budget to one ad and that ad produces the majority of conversions, that's Meta's optimization working correctly - not a problem.

Pause decision matrix:

Condition	Action
Ad has highest conversion volume in ad set	KEEP - regardless of budget share
Ad has 0 conversions AND spend > 2× ad set CPA target	PAUSE
Ad has conversions but CPA > 3× ad set average AND ≥ 3 conversions (statistically significant)	PAUSE
Ad has 1–2 conversions with high CPA	WATCH - insufficient data, do not pause yet
Ad is < 7 days old	WATCH - let it exit learning phase

Self-check before any pause recommendation: "Am I about to recommend pausing the ad that drives the most conversions in this ad set?" If yes → DO NOT recommend pausing. Instead recommend testing new creatives alongside it.

Methodology Consistency

Once you establish a methodology for evaluating performance in a conversation, do NOT change it unless the user explicitly asks. Changing criteria mid-analysis destroys trust. If you realize your methodology was wrong, say so clearly, explain what you're changing and why, and apply the new methodology consistently from that point.

STOP Rules

Pre-check before any recommendation: Can the user point to the exact Ads Manager UI field that this change maps to? If no → do not recommend.

• No ad-level budget moves
• No demographic or placement budget allocation
• No frequency caps for conversion objectives (Sales, Leads, App install)
• No scaling individual ads (only campaign / ad set level)
• No scaling > 20% at once
• Do not pause ads or ad sets younger than 7 days without enough signals
• Do not increase budget on entities below ROAS/CPL/CPA targets - diagnose first
• Do not call something "scalable" with < 10% revenue share of total account or statistically weak conversion volume
• Minimum signal: 3–5 conversions per ad set per week before performance judgments
• Avoid suggestions that restart the learning phase unless absolutely required
• Do not rely on the first 24–48 hours of data as reliable

Required Diagnostics Before Action

Before any scale / pause / duplicate / structural recommendation:

1. User segments: New vs Existing vs Engaged distribution
2. Delivery metrics: CPM, CTR, Reach, Frequency (with spend context)
   • High frequency alone ≠ fatigue
   • Frequency rising + CTR falling over 14+ days = creative fatigue
3. Engagement relevance: offer-aligned comments vs spam / backlash
4. Attribution windows: 1-day click vs 7-day click vs incremental, if platform / external data mismatch
5. Confirm results not dominated by warm / small pools before scaling

Scaling Guardrails

When recommending a budget increase:

• Only campaign or ad set budget fields (never ad-level)
• Max increase: 20% at once
• Require before scaling: performance target hit, ad set revenue share ≥ 10% of account, not dominated by 80%+ warm/customer pocket
• If data is noisy or contradictory → say insufficient data and shift to a structured testing recommendation instead
• All scaling recommendations must include: deterioration-risk warning + review window + spend cap

Change-Recommendation Safety Language

When recommending changes the user will apply manually, always include these warnings where relevant:

• Budget increase > 100%: warn this is a major increase and ask the user to confirm intent
• Budget decrease > 50%: warn this significantly reduces reach and performance
• Daily vs lifetime budget: always distinguish which one you're talking about
• Pausing: traffic and spending stop immediately
• Deletion: permanent, cannot be undone - require explicit confirmation
• Archiving: reversible - can be unarchived back to PAUSED
• Bid strategy change: resets the learning phase - always warn
• Targeting change: if non-trivial, may reset learning phase

---

Metric Glossary - Canonical Definitions

Used across every skill below.

Currency Note

Meta API budget values are typically in cents (integer). $50 = 5000 cents. If you see raw integer budget values from an API or export, divide by 100 to display. Confirm account currency before showing monetary values - never assume USD.

Direct Metrics

Metric	Notes
Spend	Total spend in account currency
CPM	Cost per 1,000 impressions
CTR	Click-through rate (decimal or %)
Impressions
Clicks
Reach	Unique users reached
Frequency	Average impressions per user
Video Plays	Initial impressions of the video
3-Second Views	"video_view" actions - hook engagement (this is inside the actions array, not a separate field)
ThruPlays	15-second or completed video watches
Purchases	Use ONE purchase action type only - see Purchase De-Duplication below
Leads	Lead events
Revenue	Conversion value for purchases
Purchase ROAS	Returned directly by Meta

Calculated Metrics

Metric	Formula
Hook Rate	(3-second views / video plays) × 100 - API/MCP source: use the video_view action inside the actions array, NOT video_p25_watched_actions. CSV source: use the "3-Second Video Plays" column directly - it's already pre-computed; the action-name distinction does not apply.
Hold Rate	(ThruPlays / 3-second views) × 100 - for CSV, use "ThruPlays" ÷ "3-Second Video Plays" columns directly.
Cost Per Purchase	spend / purchases
Cost Per Lead (CPL)	spend / leads
Cost Per Result (CPR)	spend / specific custom-event count
Conversion Rate (CVR)	conversions / clicks × 100
Budget Utilization	spend / (daily budget × days in period)
Campaign Spend Share	entity spend / parent total spend (require ≥ 25% before comparing ad sets)
Pareto Set	Sort ads by spend desc; keep ads where cumulative spend ≤ 90%

Purchase De-Duplication (CRITICAL - API/MCP only)

This rule applies only when the source is an MCP / API that returns the raw actions array. If the source is a CSV export, the "Purchases" column has already been deduplicated by Meta - skip this rule entirely and use the column directly.

When working with raw API data: Meta returns overlapping purchase action types in the actions array: omni_purchase, purchase, offsite_conversion.fb_pixel_purchase, onsite_web_purchase, web_in_store_purchase, app_custom_event.fb_mobile_purchase.

Rule: Use ONLY ONE. Check omni_purchase first; if absent, fall back to purchase. Never sum multiple types - this double-counts revenue and inflates ROAS.

Source check before applying:

• MCP returning an actions array → apply the rule.
• CSV with a "Purchases" column → skip the rule; the column is already deduplicated.
• Screenshot showing a "Purchases" number → skip the rule; trust the figure but flag the source.

Account Type → Primary Conversion Metric (PCM)

Read the promoted object / conversion event from each ad set. Never infer from campaign or ad set names - names frequently don't match the actual config.

Custom Event Type	Account Type	PCM
PURCHASE	Ecommerce	Purchase ROAS
LEAD	Lead generation	Cost per Lead
COMPLETE_REGISTRATION	Registration	Cost per Registration
OTHER (custom event)	Custom conversion	Cost per Result for that specific event

If the promoted object is missing → ask the user. Never assume.

Mixed accounts: group ad sets by their custom event type. Each group gets its own PCM. Do not mix conversion types across groups.

Special Case: COMPLETE_REGISTRATION on PURCHASE Objective (Cross-Source Safety Check)

If an ad set has campaign objective OUTCOME_SALES (or legacy CONVERSIONS / PRODUCT_CATALOG_SALES) AND the optimization event is COMPLETE_REGISTRATION, this is almost always a 3rd-party purchase tracker (Profitmetrics, Cosmise, Triple Whale, Northbeam) posting purchase events through Meta's CompleteRegistration endpoint - not a misconfiguration.

Especially common on:

• Nordic-currency accounts (DKK / SEK / NOK / EUR)
• Shopify + Profitmetrics stacks

This is the highest-stakes safety check in the skill - getting it wrong can kill a working campaign.

Source-aware detection

Source	How to detect
MCP / API	Read campaign.objective + ad_set.promoted_object.custom_event_type directly. If the pair matches the trigger, apply the check below.
CSV with "Campaign Objective" + "Performance Goal" / "Optimization Goal" columns	Match by column values. If columns are missing, fall through to the manual check.
CSV without those columns, screenshot, or copy-paste	Cannot detect automatically. Use the proactive question below before any pause / scale / budget-cut recommendation on Nordic-currency or Shopify accounts.

Proactive question (use whenever auto-detection isn't possible OR the account profile fits)

"Before I make recommendations on this account: do you use a 3rd-party purchase tracker like Profitmetrics, Triple Whale, Northbeam, or Cosmise? If yes, the conversion event reported as 'Registrations' or COMPLETE_REGISTRATION is real purchase volume and the campaign is correctly configured - I should treat it as purchase data."

Ask this once per account / conversation, store the answer, and apply it to all subsequent recommendations.

Action rules (apply after detection or confirmation)

Before any optimization recommendation on such ad sets:

1. State the ambiguity in your response, naming the 3rd-party tracker explicitly.
2. Do NOT recommend any of the following without explicit user confirmation first:
   • "Change the optimization event to PURCHASE"
   • "Verify and switch the optimization goal"
   • Pause / archive / delete the ad set or campaign
   • Reduce budget by more than 20%

Only after confirmation may you treat the events as registrations (and recommend an optimization-event change) OR as real purchases (apply ROAS / CPA against revenue).

Performance Benchmarks

Metric	Good	Average	Poor
Hook Rate	≥ 40%	26–39%	< 25%
Hold Rate	> Account Avg + 25%	≈ Account Avg	< Account Avg
CTR	≥ 1.25%	0.65 – 1.24%	< 0.65%
CPM variance vs Pareto avg	≤ 20%	20 – 60%	> 60%
PCM variance vs Pareto avg	Above avg	± 20%	> 30% below
Budget utilization	≥ 85%	-	< 85%
Min ad-set spend share for comparison	≥ 25% of campaign	-	-

For Hold Rate, compute the account average from a 90-day account-level pull (video fields aggregated) before applying the comparison.

Statistical Significance Floors

Decision	Minimum signal
Per-ad-set performance judgment	3–5 conversions / week
Per-segment recommendation (placement, age, device)	≥ 100 impressions AND ≥ 3 conversions
Pause ad on high CPA	≥ 3 conversions if non-zero, OR spend ≥ 2× ad-set CPA target if zero
Ad set learning-phase exit	Spend ≥ 2× AOV OR 4× CPA, whichever is higher

Diagnostic Signals (used in skills below)

Signal	Meaning
Frequency rising AND CTR falling over 14+ days	Creative fatigue
CBO with one ad set holding 80%+ of budget	Algorithm working as intended - do NOT pause the dominant ad set on small-sample noise
High frequency alone (no CTR drop)	Not fatigue
Cost cap + rising CPM	Either delivery squeeze OR audience saturation - separate before acting

Video Funnel

Video Plays → Hook Rate → 3-Second Views → Hold Rate → ThruPlays

---

Depth-of-Analysis Framework - Apply on Audits & Optimization Reviews

Apply on: audits, performance reviews, optimization requests, strategy questions, funnel analysis. Skip for: simple data fetches, campaign listings, single-metric lookups.

1. Hierarchical Drill-Down

Never recommend on campaign-level or ad-set-level aggregates alone. Drill one level deeper.

• Asked about a campaign → check ad set performance within it.
• Asked about an ad set → check individual ad performance within it.
• Asked about ads → check creative-level patterns (same creative across multiple ads).

Quantify the contribution: "Ad Set X accounts for 72% of campaign spend but only 40% of conversions."

2. Objective-Contextual Evaluation

Do not apply universal KPIs. Evaluate each campaign against its actual objective.

Check first:

• The ad-set-level optimization event (the actual custom_event_type - not the name)
• The campaign objective
• Bid strategy (cost cap, bid cap, lowest cost)

Apply correct KPIs:

Objective	Evaluate on	NOT on
Sales / Purchase	ROAS, CPA, revenue	Reach, CPM alone
Lead gen	CPL, lead volume	ROAS
Traffic	CPC, CTR, landing page views	Conversions
Video views	Cost per ThruPlay, hook rate	ROAS, CPA
Awareness	CPM, reach, frequency	CPA, ROAS

Flag bid-strategy issues: if a cost cap is set 3× above actual CPA, or tROAS target is 0.5× when achieving 3×, the algorithm isn't actually constrained. Call it out.

3. Breakdown & Segmentation Analysis

For audits, performance reviews, and optimization work, check at least 2 of these dimensions. (This overrides the general guidance to only use breakdowns when specifically requested - deep analysis requires segmentation.)

• Placement (Feed vs Stories vs Reels vs Audience Network) - quantify CPA gaps
• Device (mobile vs desktop)
• Age / gender - flag Advantage+ expansion leakage if delivery is going outside the target
• Out-of-target segments consuming budget with poor conversion rates

Statistical significance check: do not make segment-level recommendations from < 100 impressions or < 3 conversions in a segment. Call out low-confidence segments explicitly.

4. Temporal & Trend Analysis

Never treat the date range as a single static number. Always compare time windows.

• Last 7 days vs previous 7 days
• Identify direction for each key metric: improving, declining, stable
• Flag inflection points: "CTR started declining on [date], 3 days before CPA started rising"

Connect the chain: frequency rising → CTR falling → CPA rising → this is creative fatigue, not audience exhaustion. Correlate timing across metrics to establish cause, not just observe symptoms.

5. Cross-Entity Dependency Mapping

Don't analyze campaigns in isolation. Map the funnel.

• Which campaigns are prospecting (broad / lookalike audiences)?
• Which campaigns are retargeting (custom audiences, website visitors)?
• Does pausing a prospecting campaign starve the retargeting audience pool?

How to identify:

• Prospecting: low frequency, lower CPM, lower CTR, broader audience
• Retargeting: higher frequency, higher CPM, higher CTR, smaller audience
• Check the targeting object on each ad set - presence of custom audiences = retargeting / warm

Flag dependencies: "Pausing Campaign X will reduce the Website Visitors 7-day pool that feeds Campaign Y's retargeting. Expect retargeting CPA to rise within 1–2 weeks."

6. Platform Mechanic Awareness

Explain how Meta's mechanics affect the specific situation - explain the compound interaction, don't just mention the mechanic.

Signal	Mechanic	Consequence
Rising frequency + falling CTR	Creative fatigue	CPA will spike - need new creatives, not budget changes
CBO + one ad set getting 80% budget	CBO optimization	Usually working as intended. Do not recommend pausing the dominant ad set on small-sample noise. Only intervene if the dominant ad set itself is underperforming on ROAS/CPA.
Cost cap + rising CPM	Delivery squeeze or audience shift	Two possible causes - separate before acting: (a) algorithm can't find conversions at the cap (raise cap or loosen targeting); (b) original audience is saturated and delivery is leaking to worse segments. Check frequency, reach trend, and audience size to decide which.
Advantage+ audience expansion ON	Targeting leakage	Check breakdowns for out-of-target delivery eating budget
Insufficient spend on new ad set	Learning phase	No optimization decisions until the ad set has spent at least 2× AOV or 4× CPA, whichever is higher. The 50-conversion rule only applies to brand-new accounts with no prior history.

7. Proactive Signal Detection

After answering the user's question, surface 2–3 additional concerns or opportunities they didn't ask about.

Scan for:

• Any entity with frequency > 3 and rising - creative fatigue incoming
• Any ad set spending > 30% of campaign budget with 0 conversions - budget waste
• Any campaign where the top ad has declining CTR over 7+ days - degradation ahead
• Any retargeting audience pool shrinking (declining reach WoW)
• Any metric that looks fine now but the trend predicts problems in 7–14 days

Format: "Outside your question, I noticed [signal] in [entity] - this suggests [predicted consequence] within [timeframe]."

8. Attribution Rigor

Don't take reported conversion numbers at face value.

Always check:

• What attribution window is active? (1-day click, 7-day click, 1-day view?)
• What share of conversions are view-through vs click-through?
• Is the reported ROAS inflated by view-through on high-impression campaigns?

Flag these situations:

• Retargeting campaign with very high ROAS + mostly view-through conversions → likely over-attributed
• Two campaigns targeting the same audience → conversions may be double-counted
• Advantage+ Shopping targeting existing customers → attributed ROAS includes conversions that would have happened organically

Frame as caveat: "Before scaling Campaign X based on its reported 4× ROAS, note that [specific attribution concern] means the true incremental ROAS is likely lower."

---

Skill 1: Initial Health Check

Validate account setup before any optimization work. Run this first on any audit. If checks fail, fixing setup is the top priority - optimization on a broken setup wastes effort.

When to use this skill

• First step of any account audit
• Before recommending any structural change
• When CPA or ROAS suddenly degrades and tracking issues are plausible

Step 1: Event Match Quality (EMQ) Score Check

EMQ score is not directly available via standard insights APIs. The user must check it manually in Events Manager.

What to do:

1. First determine the primary conversion event (from the ad set's optimization goal - covered in Skill 2).
2. Ask the user to manually check: Events Manager → Data Sources → [Pixel] → Event Match Quality for that event.
3. Apply the threshold:

EMQ Score	Status	Action
≥ 8.5	Good	Proceed with analysis
< 8.5	ALERT	Stop and notify the user before optimizing

Alert wording when EMQ < 8.5:

⚠️ CRITICAL: Your EMQ Score for [Purchase / Lead] is below 8.5.

This means your Pixel is not set up correctly. Likely consequences:

• Lower conversion rate than possible
• Higher cost per result than necessary

Recommendation: Reach out to a technical expert and fix this as TOP PRIORITY before any other optimization work.

Step 2: Audience Segments Configuration Check

Audience segment breakdown (New / Engaged / Existing) is available via insights breakdowns, but the underlying engaged/existing audience setup is managed in Audience Manager and isn't directly readable via standard APIs.

What to do:

1. Pull account-level insights for the last 30 days, broken down by user segment (the breakdown that splits New / Engaged / Existing).
2. If the breakdown returns empty or is missing segments, alert the user:

⚠️ Setup needed: Audience Segments are not configured.

Why this matters: you can see spend breakdown between existing customers and new audience, at campaign / ad set / ad level. Without this, you can't tell where money is actually going.

Note: this does NOT affect targeting - only improves reporting.

Setup guide: Facebook Business Help → Audience Segments configuration.

Step 3: Decide whether to proceed

Result	Action
EMQ ≥ 8.5 AND segments configured	Proceed to Conversion Metric Focus
EMQ < 8.5	Recommend fixing pixel before any optimization
Segments not configured	Flag for setup; can proceed with limited reporting

Minimum required data (by source)

Source	What you need
MCP	Account-level insights for last 30 days with breakdowns=['user_segment_key'] to confirm segments are configured. EMQ requires manual user check.
CSV	A breakdown export by user segment (if available). EMQ requires manual user check.
Screenshot / copy-paste	Not recommended for this skill - too many missing pieces. Ask for at minimum a screenshot of Events Manager → EMQ for the primary event.

Tool hints

EMQ Score is never in standard insights data - it always requires a manual check in Events Manager → Data Sources → [Pixel]. Audience Segment configuration is observable from any source via the user-segment breakdown; if the breakdown returns empty, segments are unset.

---

Skill 2: Conversion Metric Focus

Determine account type and set the Primary Conversion Metric (PCM) that all subsequent analysis will use. Run this second on any audit, immediately after the Health Check.

When to use this skill

• Once per account, at the start of analysis
• Whenever account-level KPIs are ambiguous
• When new campaigns are added that may use a different conversion event

Data to gather

For each active campaign in the account:

• Campaign objective (e.g., OUTCOME_SALES, OUTCOME_LEADS, OUTCOME_TRAFFIC)
• The optimization event at the ad set level - the actual custom_event_type of the promoted object. Never infer from campaign or ad set names.
• The list of conversion actions used

Decision Logic

For each live campaign:

IF majority conversion event = PURCHASE         → ECOMMERCE
IF majority conversion event = LEAD             → LEAD GENERATION
IF custom conversion event (OTHER)              → CUSTOM CONVERSION (treat standalone; do not group with other types)

Mixed accounts: group ad sets by their event type and apply each group's PCM separately. Do not mix conversion types into one aggregate.

Set PCM by Account Type

Ecommerce

Metric	Type	Source
Purchase ROAS	Primary (PCM)	purchase_roas field, deduplicated per the Purchase De-Duplication rule
Cost per Purchase	Secondary	spend / purchase count
Conversion Rate	Secondary	purchases / clicks
Revenue	Secondary	conversion value for purchase events

Lead Generation

Metric	Type	Source
Cost per Lead	Primary (PCM)	spend / lead count
Conversion Rate	Secondary	leads / clicks

Custom Conversion

Metric	Type	Source
Cost per Result	Primary (PCM)	spend / count of the specific custom event
Conversion Rate	Secondary	custom events / clicks

Note: For custom conversion campaigns, metrics are scoped to that specific conversion event only. No other conversion events should be included in the analysis for that campaign.

Output

Set internally:

• account_type: ecommerce, lead_gen, or custom_conversion
• primary_metric: purchase_roas, cost_per_lead, or cost_per_result_<custom_event>

These flow into every later skill.

Minimum required data (by source)

Source	What you need
MCP	Per active campaign: campaign objective + per-ad-set promoted_object.custom_event_type. Directly available.
CSV	"Campaign Objective" + "Performance Goal" / "Optimization Event" columns (often not in default exports - must be enabled). If unavailable, ask the user to confirm the conversion event for each campaign before proceeding.
Screenshot / copy-paste	Ask user to confirm directly per campaign.

Tool hints

The most reliable signal is promoted_object.custom_event_type on the ad set object. Campaign / ad set names lie often - never infer the conversion event from a name. If using MCP, fetch the ad set object explicitly. If using CSV without the optimization-event column, this entire skill blocks until the user confirms.

---

Skill 3: Campaign Structure Analysis

Identify the strategic pattern behind how the account splits spend across campaigns. Different structures need different evaluation rules.

When to use this skill

• Account audits
• When the user asks "is my campaign structure right?"
• Before doing ad-set-level analysis (the structure type changes what comparisons are valid)

Data to gather

For each active (non-retargeting) campaign:

• Objective, bid strategy, daily/lifetime budget, status
• Ad-set level: bid strategy, optimization goal, promoted object (the conversion event), budget, attribution window, full targeting object
• Spend distribution at the campaign level over the last 30 days, sorted by spend desc

Targeting fields to inspect on each ad set:

Field	Purpose
Age min / max	Demographics
Genders	Demographics
Geo locations	Geo targeting
Custom audiences	Retargeting detection
Detailed / flexible targeting (interests, behaviors)	Interest targeting
Publisher platforms / placement positions	Placement strategy
Languages	Language targeting

Step 1: Identify Retargeting Campaigns

Check the targeting custom audiences for: website visitors, past purchasers, video viewers, page engagers.

IF targeting.custom_audiences contains retargeting audiences
→ LABEL as "Retargeting Campaign"
→ EXCLUDE from structure analysis

Step 2: Match the Active Structure to One of These Patterns

Pattern A: Primary vs Testing

Criteria:

campaign_budget_share = campaign_budget / total_non_retargeting_budget

IF Campaign A:
   - budget_share < 30%
   - ad_count ≥ 5
AND Campaign B:
   - budget_share ≥ 40%
   - ad_count < Campaign A's ad_count
   - 70%+ of ads also exist in Campaign A
→ CONFIRMED: Primary vs Testing structure

Pattern B: Bid Limit Split

Criteria:

For campaigns/ad sets (excluding retargeting), compare the targeting objects and bid configurations.

IF targeting is IDENTICAL across splits
AND bid_strategy OR bid_amount DIFFERS
→ CONFIRMED: Bid Limit Split

Pattern C: Audience Split

Criteria:

For campaigns/ad sets (excluding retargeting), compare interest spec, custom audiences, demographics.

IF audiences DIFFER between splits
→ CONFIRMED: Audience Split

NOTE: Different bids/budgets across audience splits is OK.

Pattern D: Too Many Campaigns (Custom Pattern)

When: more than one campaign and the structure doesn't fit A, B, or C.

Investigate using creative inspection - review the actual creative content of ads in each campaign and look for:

• Video vs Image split
• Catalog vs Static split
• UGC page vs Brand page
• Different products
• Different personas / angles
• Partnership ads vs direct

Also check:

• Website URLs across creatives - different funnels / landing pages / domains
• Attribution windows on each ad set - different windows often indicate different testing setups

Output

Set campaign_structure:

• primary_testing
• bid_limit_split
• audience_split
• custom_pattern: [description]

This determines which Ad Set Analysis rules apply.

Minimum required data (by source)

Source	What you need
MCP	Full ad set objects: bid strategy, optimization goal, promoted object, budget, attribution_spec, full targeting object (custom_audiences, interests, demographics, placements). All directly available.
CSV	Mostly blocked. CSV exports do not include the full targeting object. You can identify campaign-level pattern from spend distribution + budget, but the targeting-identical-vs-different test (Patterns B and C) cannot be run from CSV alone. Ask the user to describe targeting per ad set or pull from MCP.
Screenshot / copy-paste	Ask the user to walk through targeting per ad set in plain language.

Tool hints

This skill is the most MCP-dependent of the six. The structural patterns (Bid Limit Split, Audience Split) hinge on comparing full targeting specs across ad sets - something the Ads Manager CSV simply doesn't export. If MCP isn't available, ask the user to describe each ad set's audience in their own words, then map manually.

---

Skill 4: Ad Set Analysis

Evaluate ad set performance based on the campaign structure type. Run after Campaign Structure Analysis.

When to use this skill

• The campaign has more than one active ad set
• The user asks why one ad set is winning / losing vs another
• You're evaluating whether to ease bid limits, expand audiences, or pause underperformers

If the campaign has only 1 ad set → skip; campaign performance = ad set performance.

Data to gather

Pull ad-set-level performance for the last 7 days, filtered to active ad sets with impressions > 0, sorted by spend descending:

• Ad set name, ID
• Spend, impressions, CPM
• For ecommerce: purchase_roas
• For lead gen: lead actions
• Clicks
• Daily / lifetime budget
• Bid strategy, bid amount

Calculated Metrics

Metric	Formula
Budget Utilization	spend / (daily budget × days in period)
Campaign Spend Share	ad set spend / total campaign spend
Cost per Purchase	spend / purchases
Cost per Lead	spend / leads

Minimum Spend Threshold (CRITICAL)

For ALL comparisons:

Only compare ad sets where campaign_spend_share ≥ 25%
Do NOT make recommendations based on low-spending ad sets - the signal is too noisy.

4A: Bid Limit / Strategy Split Analysis

Use when: campaign_structure = bid_limit_split

Rule 1: Underspending Ad Set

budget_utilization = spend / (daily_budget × days_in_period)

IF budget_utilization < 85% AND ad set has bid_strategy with limit
→ Normal bid-limit behavior (Meta can't find cheap enough conversions)

BUT IF:
   - campaign_spend_share > 25%
   - PCM is good
   - budget_utilization < 85%
→ RECOMMEND: Drop the bid slightly to allow more spend

Rule 2: Zero / Minimal Spend

IF spend ≈ 0 over the last 7 days AND a bid_amount is set
→ RECOMMEND: Ease the bid limit so Meta can deliver

Rule 3: Clear Winner Pattern

PREREQUISITE: Both ad sets have campaign_spend_share ≥ 25%

Compare PCM (purchase_roas or cost_per_lead):

IF ad set A significantly outperforms ad set B
→ RECOMMEND:
   1. Pause low performer (ad set B)
   2. Test an even stronger bid on a new ad set
   3. Find ways to scale the winning ad set (within scaling guardrails)

Rule 4: CPM vs PCM Relationship

Calculate for each ad set: CPM and PCM

IF cpm is high (due to stronger bid) AND PCM is good
→ LEAVE IT (you're paying more but it's converting)

IF cpm_increase ≥ 60% vs other ad sets AND pcm_drop ≥ 30%
→ RECOMMEND: Ease bid to drop CPM

CPM comparison:

avg_cpm = AVG(cpm across all ad sets)
adset_cpm_variance = (adset_cpm − avg_cpm) / avg_cpm × 100

IF adset_cpm_variance ≥ 60% → FLAG as high CPM

4B: Audience Split Analysis

Use when: campaign_structure = audience_split

Rule 1: Clear Winner Pattern

PREREQUISITE: Both ad sets have campaign_spend_share ≥ 25%

Compare PCM:

IF ad set with audience A significantly outperforms audience B
→ RECOMMEND: Pause low performer to shift spend to the winner

Rule 2: Underspending Ad Set

IF budget_utilization < 85% → likely cause: small audience size (check reach)

IF:
   - campaign_spend_share > 25%
   - PCM is good
   - budget_utilization < 85%
→ RECOMMEND: Expand the audience

Prohibition

⛔ Never suggest lookalike audiences in any context.

Output Format

### Ad Set Analysis: [Campaign Name]

Structure Type: [Bid Split / Audience Split]

| Ad Set | Spend | Spend % | Budget Util | CPM | ROAS/CPA | Status |
|--------|-------|---------|-------------|-----|----------|--------|
| [Name] | $X    | X%      | X%          | $X  | X        | [Keep/Watch/Action] |

Findings:
- [Finding 1 with specific metrics]
- [Finding 2 with specific metrics]

Recommendations:
- [Recommendation 1, plain language]
- [Recommendation 2, plain language]

Minimum required data (by source)

Source	What you need
MCP	Per ad set, last 7 days: spend, impressions, CPM, purchase_roas (ecommerce) or lead actions (lead gen), clicks, daily/lifetime budget, bid strategy, bid amount. All directly available.
CSV	Ad Set view export with these columns enabled: Spend, Impressions, CPM, Purchase ROAS or Cost per Lead, Link Clicks, Budget, Bid Strategy, Bid Amount. Bid Strategy and Bid Amount are not default columns - must be enabled before export.
Screenshot / copy-paste	Workable for a single-ad-set check; not workable for multi-ad-set comparison.

Tool hints

The 7-day window is short - confirm minimum spend thresholds (≥ 25% of campaign spend share) before any comparison. Bid Strategy and Bid Amount are the two CSV columns most often missing; if absent, the bid-limit-split rules (4A) cannot run - fall back to the audience-split rules or ask for the missing columns.

---

Skill 5: Performance Analysis

Top-down ad-level audit using the Pareto principle. Identify the ads driving 90% of spend, classify each as keep / pause / test, and produce account-level strategic recommendations.

When to use this skill

• "Audit my account"
• "Which ads should I pause?"
• "What's working in this campaign?"
• Anytime you need a portfolio view before drilling into creative

Step 1: Confirm Account Type & PCM

Use the output of Conversion Metric Focus (Skill 2). If it hasn't been run yet, run it first. If the account has mixed conversion types, group ad sets by event type and analyze each group with its own PCM.

Pre-Audit Check: COMPLETE_REGISTRATION on Purchase Objective

If you see any ad set with campaign objective OUTCOME_SALES AND optimization event COMPLETE_REGISTRATION, apply the 3rd-party-tracker check from the Metric Glossary section before any recommendation on that ad set.

Step 2: Pareto Pull (90% of Spend)

Pull ad-level insights for the last 30 days. Filter to active ads (effective status = ACTIVE) with impressions > 0, sorted by spend desc. If the result set is paginated, fetch all pages before analyzing - analyzing only the first page produces wrong totals and wrong CPAs.

Required fields per ad:

• Ad name, ad ID, ad set ID/name
• Spend, impressions, CPM, CTR, clicks
• Actions (for purchase / lead counts)
• Action values (for revenue)
• Purchase ROAS

Then calculate cumulative spend % for each ad in spend-descending order. Keep the ads where cumulative % ≤ 90% - these are your Pareto ads. Focus the analysis on them.

Step 3: Establish Baselines from the Pareto Set

Baseline	Calculation
Pareto avg CPM	mean CPM across Pareto ads
Pareto avg CTR	mean CTR across Pareto ads
Pareto avg PCM	mean ROAS (ecommerce) or mean CPL / CPR (lead gen / custom)
Account avg Hold Rate	from a 90-day account-level pull (only needed for video analysis - see Skill 6)

These become the comparison points for every individual-ad assessment.

Step 4: Diagnose Each Pareto Ad

Apply the variance benchmarks (Performance Benchmarks table in Metric Glossary):

High CPM (≥ 30% above Pareto avg) + Low PCM (≥ 30% below avg)

• Diagnosis: expensive audience, not converting
• Recommendation: pause this ad OR test different targeting

Low CTR (≥ 30% below Pareto avg) + Low PCM

• Diagnosis: ad isn't resonating with the audience
• Action: first check comments for negative sentiment
• If comments are fine → recommend copy / creative variations with stronger CTA

Good metrics but declining trend

• Diagnosis: likely creative fatigue (check the fatigue signal - frequency rising + CTR falling over 14+ days)
• Recommendation: new creative variations, NOT budget changes

Good metrics, stable trend

• Diagnosis: working as intended
• Recommendation: create variations (copy / angle / format / hook) to extend the winning pattern

Step 5: Generate Recommendations

Mandatory: after analysis, produce specific, actionable recommendations.

Per-Ad Recommendations

For each Pareto ad:

• What - specific action (pause, scale, create variation, change targeting)
• Why - metrics-based justification (e.g., "CPM 40% above Pareto avg with 0.3× ROAS")
• How - concrete next steps for the user to apply in Ads Manager

Strategic Recommendations

Summarize the top 3–5 account-level moves:

1. Budget direction - which campaigns or ad sets to scale (within the 20% cap) vs which to pull back. Respect the budget-reallocation guardrails: only campaign / ad set budgets, never ad-level or per-placement.
2. Creative strategy - what's working, what to test next.
3. Targeting adjustments - if CPM issues are widespread.
4. Structure changes - campaign / ad set consolidation if needed.

Output Format

## Performance Summary

| Ad | Format | Spend | CPM | CTR | ROAS/CPA | Status |
|----|--------|-------|-----|-----|----------|--------|
| [Name] | Image | $X | $X | X% | X | [Scale/Pause/Test] |

Benchmarks (Pareto avg): CPM: $X | CTR: X% | ROAS: X

## Recommendations

### Per-Ad Actions
1. **[Ad Name]**: [Action] - [Justification]
2. **[Ad Name]**: [Action] - [Justification]

### Strategic Recommendations
1. [Plain-language recommendation with metrics basis]
2. [Plain-language recommendation with metrics basis]
3. [Plain-language recommendation with metrics basis]

If creative-level depth is needed (video hook/hold, image diagnostics, ad copy evaluation), continue to Skill 6: Creative Analysis.

Minimum required data (by source)

Source	What you need
MCP	Ad-level insights over the last 30 days for active ads with impressions > 0, sorted by spend desc. Required fields: ad name, ad ID, ad set name/ID, spend, impressions, CPM, CTR, clicks, actions (for purchase/lead counts), action_values (for revenue), purchase_roas. Paginate all pages before analyzing - analyzing only page 1 gives wrong totals and CPAs.
CSV	Ad-level view export filtered to active ads, last 30 days. Must include: Ad Name, Spend, Impressions, CPM, CTR, Link Clicks, Purchases, Purchases Conversion Value, Purchase ROAS. For lead-gen accounts, swap Purchases / ROAS for Leads / CPL.
Screenshot / copy-paste	Workable only if the screenshot already shows the full Pareto set (top N ads by spend) with all required columns.

Tool hints

The Pareto cut (top 90% of spend) is the entry point - fetching only top-10 by spend will miss ads that round out the 90% threshold. With MCP, request all active ads sorted by spend desc and paginate fully. With CSV, ensure the export was filtered to ACTIVE ads only (paused / archived ads inflate the totals). Run the COMPLETE_REGISTRATION 3rd-party tracker check from Skill 5 Step 1 before any pause / scale recommendation - this is the highest-stakes safety check in the framework.

---

Skill 6: Creative Analysis

Deep creative analysis with diagnostic scenarios for video ads, plus image and catalog format rules. Run after Performance Analysis when more depth is needed on specific Pareto ads.

When to use this skill

• Video hook / hold metrics are off
• "Why isn't this creative working?"
• Need to plan variations of winners
• The user wants to understand a specific ad in depth

Inputs You Need

Before applying this skill:

• Ad-level video metrics (video plays, ThruPlays, 3-second views - using the video_view action inside the actions array, NOT video_p25_watched_actions)
• Account-level Hold Rate baseline from a 90-day account-level pull
• The Pareto set of ads (from Skill 5)
• The actual creative content for each Pareto ad - for video, the visuals and audio; for image, the visual; for all formats, the copy (headline, primary text, CTA). The user can share screenshots, video files, or descriptions.

Step 1: Format-Agnostic Diagnostics (Apply to Every Ad)

Profile	Diagnosis	Recommendation
High CPM (≥ 30% above Pareto avg) + Low PCM (≥ 30% below)	Expensive audience, not converting	Pause ad
Low CTR (≥ 30% below Pareto avg) + Low PCM	Not resonating	Check comments first; if fine, test variations with stronger CTA
Good metrics (within 20% of avg + PCM above avg)	Working	Create variations (copy / angle / format)

Step 2: Video-Specific Diagnostic Scenarios

Apply Hook Rate / Hold Rate / CTR benchmarks (from Metric Glossary) to classify each Pareto video ad.

Scenario 1: Good Hook, Poor Hold

Metric	Status	Range
Hook Rate	Strong	≥ 40%
Hold Rate	Poor	< Account Avg
CTR	Average	0.65 – 1.24%
PCM	Poor	Below target

Interpretation: Hook is stopping scroll, but people drop off after 3 seconds. They're missing key information.

Recommendations:

1. KEEP the intro / hook (it's working)
2. CHANGE seconds 3+ completely

Tactics:

• Move value prop order around
• Show product sooner
• Get to "how this helps you" faster
• Avoid stagnant clips - use variety
• Quick B-roll + testimonial + selfie style mix
• Keep value props concise (not 10 seconds each)
• Ensure seconds 3+ relate back to the hook's problem / promise

Scenario 2: Average Everything

Metric	Status	Range
Hook Rate	Average	26 – 39%
Hold Rate	Average	≈ Account Avg
CTR	Average	0.65 – 1.24%
PCM	Poor	Below target

Interpretation: Nothing stands out. Room for improvement everywhere. Priority 1 - improve the hook first. If they don't stop, they won't purchase.

Recommendations:

1. Priority: rebuild the hook / intro first. Inspect the creative for:
   • Is the visual relevant to the target audience?
   • Is the intro speaking to the wrong customer?
   • Can it be more direct?
   • Can it relate to the customer better?
2. THEN: build the remaining clips to improve hold rate.

Warning: don't go click-baity. The hook must be relevant to the actual customer.

Scenario 3: Poor Hook

Metric	Status	Range
Hook Rate	Poor	< 25%
Hold Rate	Any	-
CTR	Any	-
PCM	Poor	Below target

Interpretation: People aren't even stopping to watch. Nothing else matters until the hook is fixed.

Recommendations:

1. Watch the actual video
2. Document the current hook (visual + audio)
3. Compare to hooks in top-performing ads (review those too)
4. Rebuild the hook based on what works

Hook elements to test:

• Different visual opening
• Different text overlay
• Different audio hook
• Problem statement vs benefit statement
• Question vs statement

Scenario 4: Great Hook + Great Hold, Poor CTR

Metric	Status	Range
Hook Rate	Good	≥ 40%
Hold Rate	Good	> Account Avg + 25%
CTR	Poor	< 0.65%
PCM	Poor	Below target

Interpretation: People are watching the whole video but not clicking. CTA is weak or unclear.

Recommendations:

1. Check comments for negative sentiment
2. If comments are fine → strengthen the CTA

CTA improvements:

• More prominent visual CTA
• Clearer verbal CTA
• Urgency / scarcity element
• Better landing-page alignment

Step 3: Format-Specific Rules

Single Image

• High CPM (30%+ above Pareto avg) + Low PCM → recommend pausing the ad
• Low CTR (30%+ below avg) + Low PCM → check comments; if fine, test stronger-CTA variations
• Good metrics → create variations (copy, angle, or format)

Single Video

Apply Scenarios 1–4 based on Hook / Hold / CTR profile, plus:

• High CPM + Low PCM → pause
• Low CTR + Low PCM → check comments, then test stronger-CTA variations

Advantage+ Catalog Ads

Limitation: Meta does NOT provide conversion data per product ID. Workaround:

• Build a Pareto of product IDs by spend

• Use spend + CTR as a proxy - higher spend + higher CTR = likely better performer

• High CPM + Low PCM → pause ad, test a different product set. Also check: if 1–2 product IDs are pushing CPM up, recommend removing those products from the set.

• Low CTR + Low PCM → identify low-CTR products from spend/impressions data, recommend removing them from the set.

Step 4: Pattern Recognition Across Top Performers

Among the top Pareto ads, review the creative content of each and identify shared patterns:

• Common angle patterns (which selling points repeat across winners)
• Common format patterns (image vs video vs carousel)
• Common hook patterns (videos - what stops the scroll)
• Common copy patterns (tone, length, structure)

Use winning patterns to guide new variation creation.

Step 5: Variation Strategy

When an ad is performing well

Create variations in this order:

1. Copy variations (same angle, different words)
2. Angle variations (different selling point)
3. Format variations (same message, different format)
4. Hook variations (videos - same body, different hooks)

When an ad is performing poorly

Issue	Action
Hook Rate poor	Rebuild hook using winning hooks
Hook good, Hold poor	Rebuild middle section
Hook good, Hold good, CTR poor	Strengthen CTA
All metrics poor	Test a completely new approach

Decision Tree Summary (Video)

START
  │
  ├─ Hook Rate < 25%?
  │    └─ YES → Rebuild hook completely (Scenario 3)
  │
  ├─ Hook Rate 26–39%?
  │    └─ YES → Improve hook first, then content (Scenario 2)
  │
  └─ Hook Rate ≥ 40%?
       │
       ├─ Hold Rate < Account Avg?
       │    └─ YES → Rebuild middle section (Scenario 1)
       │
       └─ Hold Rate good but CTR < 0.65%?
            └─ YES → Strengthen CTA (Scenario 4)

Output Format

### Creative Analysis: [Ad Name]

Creative Profile:
| Element     | Analysis      |
|-------------|---------------|
| Angle       | [description] |
| Persona     | [description] |
| Format      | [description] |
| Visual Hook | [description / N/A for image] |
| Audio Hook  | [description / N/A for image] |
| CTA         | [description] |

Metric Diagnosis:
| Metric    | Value | Benchmark         | Status        |
|-----------|-------|-------------------|---------------|
| Hook Rate | X%    | Good ≥ 40%        | [Good/Avg/Poor] |
| Hold Rate | X%    | Acct Avg: Y%      | [Good/Avg/Poor] |
| CTR       | X%    | Good ≥ 1.25%      | [Good/Avg/Poor] |
| ROAS/CPA  | X     | [Target]          | [Good/Poor]   |

Benchmarks used:
- Account avg Hold Rate (90d): X%
- Pareto avg CPM: $X
- Pareto avg CTR: X%

Scenario Match: [1 / 2 / 3 / 4 / N/A for non-video]
Root Cause: [description]

Top Performers - Common Patterns:
- Angle: [pattern]
- Format: [pattern]
- Hook: [pattern, videos]
- Copy: [pattern]

Recommendations (plain language, for the user to apply in Ads Manager):
1. [Specific recommendation with metrics justification]
2. [Specific recommendation with metrics justification]
3. [Specific recommendation with metrics justification]

Minimum required data (by source)

Source	What you need
MCP	Ad-level video metrics: video_play_actions, actions (containing video_view = 3-second), ThruPlay events. Account-level 90-day pull for Hold Rate baseline. Creative content via creative endpoint (e.g. facebook_get_ad_creative_details).
CSV	Ad-level export with "Video Plays", "3-Second Video Plays", and "ThruPlays" columns enabled. Plus a separate 90-day account-level aggregate pull for Hold Rate baseline. Creative content (the actual video / image / copy) is NEVER in a CSV - must be shared separately.
Screenshot / copy-paste	Works for spot-checking a single ad's video metrics if visible. Creative content can come from a video file upload or a description.

Tool hints

This skill needs two kinds of data: numeric (Hook/Hold/CTR) and qualitative (the actual creative). Numeric works fine from any source as long as video columns are enabled. Qualitative requires the user to share files or text - no export captures creative content. Build a checklist before starting: numeric metrics ✓, 90-day Hold Rate baseline ✓, the actual ad creative ✓. Without all three, the skill outputs framework, not diagnosis.

---

Cross-Cutting Notes

How to handle missing data

If the user can't provide the data a skill needs, do not invent values. State explicitly:

"To answer this properly I need [specific metrics]. You can pull these from [Ads Manager export / a connected data tool / Events Manager]. Without them I can outline the framework but not give a recommendation."

How to handle change requests

When the user asks Claude to change something in their account (e.g. "pause this ad", "raise the budget 20%"):

1. Run the relevant analysis skill to validate the change is justified.
2. Apply the guardrails (especially scaling caps, top-performer protection, and the "no ad-level budget moves" rule).
3. Recommend the change in plain language - exact entity, current value, proposed value, rationale, what to monitor afterwards, and any safety warnings (learning-phase reset, > 50% decrease, > 100% increase, etc.).

Claude does not execute changes on the user's ad account. These skills are read-and-recommend only - the user applies recommendations manually in Ads Manager.

How to handle multiple account types in one account

If the account has both purchase-optimized and lead-optimized campaigns:

• Group ad sets by their custom_event_type
• Apply each group's PCM separately
• Never aggregate ROAS across purchase + lead campaigns

How to handle account-type ambiguity

Always classify the account before applying KPIs:

1. Check the actual ad-set-level conversion events (not the campaign / ad set names).
2. Check whether conversion value is populated.
3. If unclear, ask the user.
4. Apply the special-case check for COMPLETE_REGISTRATION on purchase objectives (3rd-party tracker scenario).

How to handle stale data

Meta campaigns change frequently. If the data the user shared is more than a day or two old:

• Ask for a fresh pull before any recommendation
• Especially before pause / scale / structural changes
• Recommendations based on stale state can be wrong or harmful

What to do when the API / export doesn't show a needed metric

Some signals aren't directly available via standard insights exports:

• EMQ Score → manual check in Events Manager
• Audience segment configuration → manual check in Audience Manager (the breakdown shows distribution, but setup state isn't readable)
• Comments / sentiment → manual check in Ads Manager or the Page
• Per-product conversion data in catalog ads → not provided; use spend + CTR as proxy

When a signal isn't available, state that explicitly, point the user to where they can check it manually, and frame the recommendation around the data that IS available.

What good looks like

• The health check and conversion-metric decision run before any performance judgment, and the metric never changes mid-analysis.
• The top-converting ad in an ad set is never recommended for pause, whatever its budget share.
• Budget recommendations touch only campaign or ad set budgets, bid caps, or on/off. Breakdowns are read as distribution, not as levers.
• Scaling moves are capped at 20 percent, carry a review window and a spend cap, and wait for the ad set to clear the revenue-share and significance floors.
• Creative verdicts pair the numbers (hook, hold, CTR) with the actual creative, or the output says it can only give the framework.