Skip to content
Run this on autopilot — free
Workflow · Creatives

Turn your best Meta ad into 50 new creative variations to test

Time to first output: 10 minutes to run once connected, plus time spent generating the 50 images in ChatGPT

What it does?

The prompt finds your best-performing Meta ad, extracts its brand kit (colors, fonts, visual style) via GoMarble MCP, and writes 50 ready-to-paste ChatGPT image prompts across 7 variation dimensions for testing.

What you need

  • Meta Ad Account ID (act_XXXXXXXXXX)
  • Date range (default Last 30 Days)
  • Winning metric (default ROAS; options CTR, CPA)

First, connect Claude to your ad accounts

Add GoMarble AI as a custom connector in Claude so it can pull your Meta Ads performance data directly inside the chat.

1

Sign up

Sign up on GoMarble using your work email.

2

Add a custom connector

Go to Claude Integrations and click "Add Custom Connector".

3

Add the following details

Enter the Name and URL below.

Name

GoMarble AI

URL

https://apps.gomarble.ai/mcp-api/sse

4

Finish setup

Click "Add" and then "Connect".

Make sure your Meta ad account is connected on GoMarble first — this prompt pulls the winning ad from that account.

The prompt

Paste this into Claude at claude.ai/new and update the Meta Ad Account ID, date range, and winning metric before running. The output is 50 GPT Image prompts — for the best results, paste those into ChatGPT (with Images 2.0 enabled) to generate the actual visuals.

Lead-magnet prompt · free

## ✏️ FILL THESE IN BEFORE RUNNING

1. Meta Ad Account ID: ""  ← paste your account ID here (format: act_XXXXXXXXXX)
2. Date Range: "Last 30 Days"  ← change if needed (e.g. "Last 7 Days", "Last 60 Days")
3. Winning Metric: "ROAS"  ← change if needed (options: CTR, CPA)

---

You are a performance creative strategist. Your job is to:
1. Connect to GoMarble MCP and find the highest-performing ad in the Meta account provided above
2. Visually analyze the winning creative and extract a brand kit from it
3. Output 50 ready-to-use ChatGPT image generation prompts across 7 variation dimensions

---

## STEP 1 — PULL WINNING AD FROM GOMARBLE MCP

- Connect to GoMarble MCP (endpoint: https://apps.gomarble.ai/mcp-api/sse)
- Call the ad-level creative performance tool for the Meta account ID provided, over the selected date range
- Sort by the winning metric (ROAS by default)
- Surface the TOP 3 performing ads with these columns:
  - Ad Name / ID
  - ROAS (or CTR / CPA)
  - Spend
  - Impressions
  - Clicks
  - Frequency
  - Creative image URL (retrieve from GoMarble MCP response — this will be used in Step 2)

Present the table to the user and say:
"Here are your top 3 performing ads. I'll use #1 as the winning creative — confirm or tell me if you'd like to use a different one."

After confirmation, proceed to Step 3.

---

## STEP 2 — BUILD THE BRAND KIT BY ANALYZING THE WINNING CREATIVE

Do NOT ask the user to describe the ad. Instead:

1. Take the creative image URL returned by GoMarble MCP for the winning ad
2. Fetch and visually analyze the image directly
3. Extract the following brand kit elements from what you see:

**BRAND KIT — [Ad Name/ID]**
- Product name (infer from visual or ad name)
- Product appearance (shape, color, packaging, size)
- Hero visual (person, product-only, or lifestyle scene)
- Setting / background
- Dominant brand colors (extract from image)
- Headline text visible on the ad
- Subheadline or CTA text (if visible)
- Brand logo — position and color
- Overall mood/tone (clean, bold, luxury, playful, clinical, energetic)
- Inferred target audience (age, gender, lifestyle — from visual cues)
- Key compositional rules to lock across all variants (e.g. product always center-frame, logo always bottom-right)

Present the extracted brand kit to the user for a quick sanity check:
"Here's what I extracted from the winning ad. Let me know if anything needs correcting before I generate the 50 prompts."

Wait for confirmation or corrections, then proceed to Step 3.

---

## STEP 3 — GENERATE 50 CHATGPT IMAGE PROMPTS

Once you have the brand kit, generate the following 50 prompts.

**STRUCTURE OF EACH PROMPT:**
Every prompt must open with a "MASTER LOCK" block that keeps brand, product, and composition consistent. Then apply the variation axis. Format each prompt so it can be pasted directly into ChatGPT's image generation interface.

Use this master lock template at the start of every prompt:

---
MASTER LOCK:
Product: [product name and exact appearance from brand kit]
Brand colors: [from brand kit]
Logo: [position and color from brand kit]
Composition rule: [e.g. product always center-frame, clean negative space on left for text]
Mood: [from brand kit]
Output: photorealistic advertising image, studio-quality, 4K, no watermarks
---

Then apply the variation below it.

---

### DIMENSION 1 — HOOK ANGLE (8 variants)
Generate 8 prompts. Same product, same visual setup. Only the psychological framing / implied message changes via scene staging and facial expression.

Variant 1.1 — FEAR / LOSS AVERSION
[MASTER LOCK]
Scene staging: Show the problem-state — a frustrated or stressed [target audience] surrounded by signs of the problem that [product] solves. Lighting: harsh, slightly desaturated. Headline overlay (bottom third): "[write a fear-based headline from brand kit copy]"

Variant 1.2 — SOCIAL PROOF / BELONGING
[MASTER LOCK]
Scene staging: A group of [target audience] happily using [product] together. One person holds it up as if recommending it. Warm social lighting. Headline overlay: "Join [X] people who already [benefit]"

Variant 1.3 — CURIOSITY / OPEN LOOP
[MASTER LOCK]
Scene staging: Close-up of [product] partially hidden, being revealed by a hand. Dramatic side lighting. Headline overlay: "Why are [target audience] switching to this?"

Variant 1.4 — AUTHORITY / CREDIBILITY
[MASTER LOCK]
Scene staging: [product] displayed on a clean surface next to a professional credential, award badge, or clinical-looking tools. Clean white/grey background. Headline overlay: "The #1 [category] trusted by [expert group]"

Variant 1.5 — ASPIRATION / IDENTITY
[MASTER LOCK]
Scene staging: Aspirational lifestyle shot — [target audience] in an elevated, successful moment using [product]. Golden hour or luxury interior lighting. Headline overlay: "This is what [desired identity] looks like."

Variant 1.6 — URGENCY / SCARCITY
[MASTER LOCK]
Scene staging: [product] with a subtle "low stock" or "limited" visual cue — minimal units displayed, bold accent color. Headline overlay: "Only [X] left — [benefit] before it's gone."

Variant 1.7 — BEFORE/AFTER CONTRAST
[MASTER LOCK]
Scene staging: Split composition — left side shows the "before" problem state (dull, frustrated), right side shows the "after" result state (bright, confident) with [product] bridging both. Headline overlay: "[Before state] → [After state]. Meet [product]."

Variant 1.8 — SIMPLICITY / ANTI-OVERWHELM
[MASTER LOCK]
Scene staging: Ultra-minimal — [product] alone on white or single-color background. Lots of negative space. Zero clutter. Headline overlay: "Finally. [Benefit] in [timeframe/steps]."

---

### DIMENSION 2 — BACKGROUND / SCENE SWAP (8 variants)
Same hero shot of [product/person], background only changes.

Variant 2.1 — Clean studio white: [MASTER LOCK] Background: seamless white studio. Soft shadows underneath product. Professional product photography lighting.
Variant 2.2 — Outdoor lifestyle: [MASTER LOCK] Background: natural outdoor setting relevant to [target audience lifestyle]. Golden hour. Organic, unposed feel.
Variant 2.3 — Kitchen/Home: [MASTER LOCK] Background: warm modern kitchen or living space. Lifestyle context. Natural light from window.
Variant 2.4 — Urban city: [MASTER LOCK] Background: blurred urban street or city skyline. Neon or warm street lighting. Dynamic energy.
Variant 2.5 — Dark luxury: [MASTER LOCK] Background: deep charcoal or black background. Dramatic single-source spotlight on product. Luxury editorial feel.
Variant 2.6 — Nature/Wellness: [MASTER LOCK] Background: lush greenery or serene natural landscape. Soft, airy, wellness aesthetic. Pastel sky.
Variant 2.7 — Office/Workspace: [MASTER LOCK] Background: clean modern desk setup. Productivity context. Neutral tones. Laptop visible.
Variant 2.8 — Abstract color wash: [MASTER LOCK] Background: bold single-color gradient wash in [brand color]. Product in sharp focus against it. No environmental context.

---

### DIMENSION 3 — PERSONA ROTATION (8 variants)
Same product, same scene structure. Only the featured person changes.

Variant 3.1: [MASTER LOCK] Featured persona: Woman, 25–34, urban professional. Confident, direct gaze.
Variant 3.2: [MASTER LOCK] Featured persona: Man, 35–50, active lifestyle. Outdoors, energetic.
Variant 3.3: [MASTER LOCK] Featured persona: Woman, 45–60, sophisticated. Home or wellness setting.
Variant 3.4: [MASTER LOCK] Featured persona: Man, 22–30, fitness-oriented. Gym or sports context.
Variant 3.5: [MASTER LOCK] Featured persona: Woman, South Asian / Indian, 28–40. Professional or family context.
Variant 3.6: [MASTER LOCK] Featured persona: No person — product-only hero shot with strong compositional framing.
Variant 3.7: [MASTER LOCK] Featured persona: Older couple, 55–65. Warm, relatable lifestyle moment.
Variant 3.8: [MASTER LOCK] Featured persona: Young adult, gender-neutral presentation, 20–28. Casual, authentic, slightly editorial.

---

### DIMENSION 4 — ASPECT RATIO REFORMATS (6 variants)
Same creative concept. Recompose layout for each format.

Variant 4.1 — Square 1:1 (1080×1080): [MASTER LOCK] Canvas: 1080×1080px. Product centered. Headline text centered below or overlaid. Balanced composition.
Variant 4.2 — Portrait 4:5 (1080×1350): [MASTER LOCK] Canvas: 1080×1350px. Product upper-center. More vertical breathing room. Text block in lower third.
Variant 4.3 — Story/Reel 9:16 (1080×1920): [MASTER LOCK] Canvas: 1080×1920px. Full-bleed vertical. Product in upper 60%. Bold headline in lower 30%. CTA strip at very bottom.
Variant 4.4 — Landscape 16:9 (1920×1080): [MASTER LOCK] Canvas: 1920×1080px. Product on right two-thirds. Headline on left one-third with white space. Banner ad composition.
Variant 4.5 — Display 300×250: [MASTER LOCK] Canvas: 300×250px. Highly simplified. Product dominant. 5-word headline max. Logo bottom-right. No clutter.
Variant 4.6 — Landscape 4:3 (1440×1080): [MASTER LOCK] Canvas: 1440×1080px. Horizontal lifestyle composition. Product foregrounded, scene behind.

---

### DIMENSION 5 — COPY OVERLAY CLASSES (8 variants)
Same image. Change only the headline text style overlaid. Tell ChatGPT to keep all visual elements identical and only swap the text block.

Variant 5.1 — DIRECT BENEFIT headline: [MASTER LOCK] Overlay text: "[Specific, measurable outcome] in [timeframe]." Bold white sans-serif, center-bottom.
Variant 5.2 — QUESTION hook: [MASTER LOCK] Overlay text: "Why does [target audience] keep buying [product]?" Italic or mixed weight. Top-third position.
Variant 5.3 — SOCIAL PROOF number: [MASTER LOCK] Overlay text: "[X,000]+ [target audience] can't be wrong." Large numeral, secondary brand color accent.
Variant 5.4 — ONE-WORD impact: [MASTER LOCK] Overlay text: Single word in massive display type — "[RESULT]." Minimal, editorial. Full-width text.
Variant 5.5 — BEFORE/AFTER micro-copy: [MASTER LOCK] Overlay text: Two lines — "Before: [problem state]" struck through, "After: [result state]" below in bold.
Variant 5.6 — URGENCY CTA: [MASTER LOCK] Overlay text: "Limited — [benefit] before [deadline/stock runs out]." Red or accent color pill badge.
Variant 5.7 — TESTIMONIAL pull-quote: [MASTER LOCK] Overlay text: Short quote in quotation marks — "[1-sentence customer result]" with a first name below. Handwriting-style or italic font.
Variant 5.8 — NO COPY (visual-only): [MASTER LOCK] No text overlay. Pure product/lifestyle image. Let the visual carry the message. Highest-quality render.

---

### DIMENSION 6 — VISUAL STYLE (6 variants)
Same composition and subject. Render style changes completely.

Variant 6.1 — Photorealistic: [MASTER LOCK] Render style: Ultra-photorealistic. DSLR camera look. 85mm lens, shallow depth of field. Shot as if by a professional commercial photographer.
Variant 6.2 — Illustrated/Editorial: [MASTER LOCK] Render style: Bold editorial illustration. Flat vector-adjacent with expressive linework. Think high-end magazine insert.
Variant 6.3 — Minimalist/Swiss: [MASTER LOCK] Render style: Swiss modernist minimalism. Clean geometry, primary colors, generous whitespace, Helvetica-inspired type layout.
Variant 6.4 — Vintage/Retro: [MASTER LOCK] Render style: 1970s retro advertising aesthetic. Warm muted palette, aged paper texture, slightly faded. Serif display type.
Variant 6.5 — Color Block / Graphic: [MASTER LOCK] Render style: Bold color-blocking. High contrast solid color fields. Graphic design poster aesthetic. Pop-art energy.
Variant 6.6 — Soft Lifestyle / Film: [MASTER LOCK] Render style: Film photography aesthetic. Slight grain, warm tones, candid moment energy. 35mm film look.

---

### DIMENSION 7 — LOCALISATION / MARKET VARIANTS (7 variants)
Same creative. Adapt for 7 markets. Native-script headline. Keep brand kit locked.

Variant 7.1 — US English: [MASTER LOCK] Language: English (US). Headline: [US-localized version of headline]. Cultural context: American lifestyle imagery cues.
Variant 7.2 — UK English: [MASTER LOCK] Language: English (UK). Headline: [UK spelling/tone]. Cultural context: British lifestyle cues, slightly understated.
Variant 7.3 — India (Hindi): [MASTER LOCK] Language: Hindi. Headline in Devanagari script: "[translate headline to Hindi]". Cultural context: Indian family/aspirational lifestyle cues.
Variant 7.4 — Spain / LATAM (Spanish): [MASTER LOCK] Language: Spanish. Headline: "[translate headline]". Cultural context: warm, family-oriented, vibrant color palette.
Variant 7.5 — Germany (German): [MASTER LOCK] Language: German. Headline: "[translate headline]". Cultural context: precise, quality-focused framing. Clean, functional aesthetic.
Variant 7.6 — UAE / MENA (Arabic): [MASTER LOCK] Language: Arabic. Headline in Arabic script (right-to-left): "[translate headline]". Composition mirrored for RTL. Warm luxury aesthetic.
Variant 7.7 — Brazil (Portuguese): [MASTER LOCK] Language: Brazilian Portuguese. Headline: "[translate headline]". Cultural context: energetic, social, sun-forward lifestyle.

---

## STEP 5 — OUTPUT FORMAT

After generating all 50 prompts, present them in the following format:

1. A summary table:
   | # | Dimension | Variant | Key Variable |
   |---|-----------|---------|--------------|
   List all 50 rows.

2. Then output all 50 prompts in clearly numbered sections, each ready to copy-paste directly into ChatGPT's image generation interface.

3. End with a **3-Stage Test Plan**:

---
### 3-STAGE AD TESTING PLAN

**Stage 1 — Hook & Format Screen (Week 1)**
Run 8 Hook Angle variants (Dimension 1) + 3 Aspect Ratios (1:1, 4:5, 9:16) simultaneously.
Budget: Equal split. Min 1,000 impressions per variant.
Kill threshold: Remove any variant below 50% of winner's CTR after 1,000 impressions.
Promote: Top 2 hook angles advance to Stage 2.

**Stage 2 — Style & Persona Refinement (Week 2)**
Take the 2 winning hooks. Apply 3 Style variants (6.1, 6.5, 6.6) × 3 Persona variants (3.1, 3.2, 3.6).
Budget: 70% on top-performing hook from Stage 1.
Signal: ROAS and CTR. Find the Style × Persona combo that indexes highest.

**Stage 3 — Copy & Localisation Scale (Week 3+)**
Take the Stage 2 winner. Apply all 8 Copy Overlay variants to find the text layer that lifts CTR.
Then localise the Stage 3 winner across 3 priority markets from Dimension 7.
This is your scaled creative pack.
---

Begin now. Start with STEP 1 and ask the user for their inputs.

What you get back

  • Format: Three parts, in this order:
    • A summary table of all 50 image prompts (# / Dimension / Variant / Key Variable)
    • All 50 prompts written out in full, ready to paste into ChatGPT's image interface
    • A 3-Stage Test Plan
Upgrade

Want this automated?

Run the prompt above every Monday morning automatically. GoMarble Agents don't just deliver the report (Slack / email / in-app) — they can execute the recommended changes directly in your ad account, too. Offload the whole recurring task and stop doing it manually.

Try GoMarble Agents free →

FAQ

Do I need to build anything myself?
No — paste the prompt into Claude at claude.ai/new after connecting the GoMarble AI custom connector; Claude pulls your top-performing Meta ad and generates the prompts for you.
Where do I actually generate the images?
The prompt outputs 50 ChatGPT image-generation prompts; you paste them one by one into ChatGPT Images (chatgpt.com/images) with Images 2.0 enabled.
Can I change the ranking metric used to find the winning ad?
Yes — edit the 'Winning Metric' field at the top of the prompt (default ROAS; options are CTR or CPA).
How many total image variants does this produce?
Running each of the 50 prompts at n=5 in ChatGPT Images produces 250 image variants (5 visual variants per angle).
Does this work for platforms other than Meta?
The prompt as written pulls the winning ad from a Meta ad account only.
What if I want vertical (9:16) creatives?
Change 'A 1:1 ad creative' to 'A 9:16 vertical ad creative' in the relevant prompt block and regenerate.
What's the 3-stage test plan for?
It's the testing sequence included at the end of the output — Week 1 screens 8 hook angles across 3 aspect ratios, Week 2 refines style/persona combos, and Week 3+ tests copy overlays and localises the winner.

Skip the prompt — let GoMarble do this for you.

Sign up, connect your ad accounts, and winning ad variations generator runs on every account, every week, automatically.