每月創意素材封存

INPUT

• Last 90 days of ad performance data
• Top 20% performers → visual element analysis

STEP 1 — Classify & Flag

For each creative active in the past 90 days, calculate:

• Lifetime CPA (total spend / total conversions)
• Total spend

Tag each creative by tier:

• Top 20%: CPA in top 20% (lowest cost per result) → TIER_TOP, document winning attributes
• Middle 60%: all others → TIER_MIDDLE
• Bottom 20%: CPA in bottom 20% OR zero conversions with significant spend → TIER_BOTTOM, flag for archival

Also flag any duplicate creatives: same image/video + same copy + same CTA across multiple placements.

STEP 2 — Winning Creative Analysis

For each top performer, retrieve creative assets (image URL or video thumbnail via ad ID). Then analyze visual elements:

• Hook type (first 3s of video or first frame of image)
• Visual style (color palette, format, composition)
• CTA type
• Copy angle (headline and description themes)

Analyze top performers only — skip middle and bottom tier. Limit to 5 creatives per execution (prioritize highest performers).

STEP 3 — Compile Creative Library Index

Build the full index as a markdown table (see OUTPUT below). This serves as the creative library.

STEP 4 — Archive Log

Produce a log of all assets flagged for archival this cycle: asset ID, reason archived, date.

OUTPUT

📁 CREATIVE LIBRARY INDEX (90-day window)

Asset ID	Spend (90d)	CPA	Tier	Hook Type	Visual Style	CTA

🗑️ ARCHIVE LOG

Asset ID	Reason Archived	Date

🎯 CREATIVE DIRECTION SUMMARY (Top 20% patterns)

Attribute	Pattern
Hook type	...
Visual style	...
CTA	...
Copy angle	...

GUARD

• Fewer than 10 creatives in 90-day window → skip tier classification, output raw performance list only
• An asset with no conversion data → classify based on spend and frequency only, tag as insufficient_data
• Archival means pausing the ad that carries this creative — ad and creative are 1:1 bound, creative cannot be replaced or deleted independently
• Archive list is for human review — requires confirmation before any pause action is taken
• Duplicate detection: approximate — same image/video + same headline + same CTA
• Never analyze more than 5 creatives per execution