Winning Creative Variations
This skill should be used when the user wants to find top-performing ad creatives from their Meta ad accounts via Glued MCP, analyze why they work, and generate variation prompts for Gemini image generation. Triggers on requests like "find winning creatives", "generate creative variations", "analyze top ads", or "make variations of best performing ads".
Winning Creative Variations
Find your top-performing ad creatives, analyze why they work, and generate variation prompts for Gemini image generation.
Your best ads already exist — this skill finds them, downloads them, visually analyzes each one using Claude's vision, and produces a variation matrix with ready-to-run Gemini image generation prompts. Stop guessing which concepts to iterate on.
What it does
- Scout Agent — Discovers your workspace and picks 2-3 active ad accounts with non-zero spend via Glued MCP.
- Analyst Agent — Pulls the top 30 creatives by purchase count, filters to static images only (excludes video and carousel), takes the top 10, and downloads each image locally.
- Strategist Agent — Reads each downloaded image using Claude's vision. Analyzes why each creative works across visual, copy, and format dimensions. Generates 5 variation angles per creative with detailed Gemini generation prompts.
- Generator (optional) — Runs
generate-variations.mjsto produce variation images via Gemini 3.1 Flash API. Requires explicit user confirmation before executing.
Output
output/<YYYYMMDD_HHmmss>/
├── images/ # Downloaded winning creatives
│ ├── creative_1.jpg
│ └── ...creative_10.jpg
├── variation_matrix.csv # Full matrix with Gemini prompts
├── analysis_report.md # Human-readable creative analysis
└── variations/ # Generated variations (after Phase 4)
├── creative_1_v1.png
└── ...
What the analysis actually looks like
Each winning creative gets analyzed across:
- Visual — Composition, color palette, focal point, imagery style
- Copy/Text — Headline hook, CTA, emotional trigger, value proposition
- Format — Aspect ratio, text-to-image ratio, branding placement
Each variation prompt includes style/mood/composition directives and references the original image for style transfer.
What this replaces
| Before | After |
|---|---|
| Manually checking Ads Manager for top performers | Automated top-10 by purchase count across accounts |
| "This ad is doing well" without knowing why | Structured analysis across visual, copy, and format dimensions |
| Briefing designers with vague "make more like this" | 5 specific variation angles per creative with rationale |
| Writing Gemini/Midjourney prompts from scratch | Ready-to-run generation prompts in CSV format |
| Iterating on mediocre creatives | Data-driven selection of proven winners to iterate on |
| Hours of creative analysis meetings | Automated analysis report with actionable variation matrix |
Who this is for
- Performance marketers who want to systematically iterate on what's already working
- Creative teams who need specific variation briefs, not vague "make more like this"
- Media buyers looking to scale winning concepts across new angles
- Growth teams building creative testing pipelines grounded in performance data
- Solo founders who want agency-level creative analysis without the agency
Requirements
| Requirement | Details |
|---|---|
| Glued MCP | Connected Meta ad accounts with active spend |
| Claude Code | Runtime environment |
| Gemini API Key (optional) | GEMINI_API_KEY in .env — only needed for Phase 4 image generation. Get one at aistudio.google.com/apikey |
| Node.js (optional) | For running the generation script in Phase 4 |
| curl | For downloading creative images (standard on macOS/Linux) |
Skill structure
winning-creative-variations/
├── SKILL.md # Skill definition with 4-phase workflow
├── references/
│ └── subagents.md # Full prompts for Scout, Analyst, and Strategist agents
├── scripts/
│ └── generate-variations.mjs # Gemini 3.1 Flash variation generation script
├── .env.example # Optional environment variables
└── README.md # This file
How it works under the hood
| Agent | Model | Job |
|---|---|---|
| Scout | Haiku | Discover workspace, pick 2-3 active ad accounts |
| Analyst | Sonnet | Pull top 30 creatives, filter to images, download top 10 |
| Strategist | Sonnet | Visual analysis via Claude vision, variation angles, Gemini prompts |
| Generator | Node.js script | Gemini 3.1 Flash image generation (optional, requires confirmation) |
Key technical details:
- Analyst filters by creative ID prefix (
image::) to exclude video and carousel - Top 10 selection uses purchase_count as primary sort, ROAS as tiebreaker
- Strategist uses Claude's native
Readtool on images — zero-cost visual analysis - Generation script processes CSV rows sequentially with rate limiting
- Phase 4 always requires explicit user confirmation before running
For performance marketers who want to scale what's already winning, not start from scratch.