Experiment: Piggybacking Competitor Ad Concepts with AI

Experiment: Piggybacking Competitor Ad Concepts with AI

TL;DR: Used the Ad Alchemy skill to extract a creative formula from a Tastier meal-plan ad and generate 5 variations for fitme.lt. The formula transferred cleanly — variations are production-ready and distinct enough from the source to be legally safe and brand-appropriate.

The Business Problem

You see competitors running successful ads. You want to learn from what’s working. But:

• Copying is risky — legal issues, looks derivative, damages brand
• Hiring consultants is expensive — $200+/hour for art direction
• Starting from scratch wastes intelligence — the market has already validated certain approaches
• “Inspiration” is vague — “I like the vibe” doesn’t translate to executable creative briefs

The question: Can AI extract the structural formula from a winning ad and apply it to your product — giving you the benefit of creative validation without copying?

Hypothesis

A multimodal AI with a structured analysis workflow can:

1. Identify the formula (lighting, composition, copy pattern) that makes an ad work
2. Separate formula from skin (brand-specific elements)
3. Apply the formula to a different product with different brand colors
4. Produce variations that are production-ready and legally distinct

Method

Setup

Role	Who
Competitor with proven ads	Tastier — Lithuanian meal-planning service, established social presence
Brand with no ads running	fitme.lt — meal-planning app, same category, needs creative
Tool	Ad Alchemy skill for Claude
Reference ad	Tastier 9:16 infographic — 5 meals per day, checklist format

Process

1. Visual Deconstruction — Analyze reference ad across 10 layers (composition, lighting, palette, typography, etc.)
2. Template Extraction — Compress findings into reusable spec with competitor-specific bits abstracted
3. Formula Application — Apply template to fitme.lt with their brand colors and product
4. Variation Generation — Create 5 variations with distinct testing hypotheses
5. Quality Assessment — Evaluate if outputs are production-ready and distinct

Success Criteria

Criterion	Threshold
Formula identified	8+ of 10 layers concretely articulated
Variations generated	5 distinct variations with testable hypotheses
Production readiness	Prompts executable by image models, copy fits character limits
Legal distinctness	No trademarked elements, no direct copying
Brand fit	Colors, voice, product clearly fitme.lt, not Tastier

Results

Formula Extracted

The 10-layer deconstruction successfully identified the structural choices:

Layer	Finding
Composition	9:16 vertical, two-column editorial (type left, imagery right)
Focal hierarchy	Headline (two-class type) → food circles → labels
Lighting	Soft overhead key, ~5000K neutral, low contrast, no drama
Palette	60% warm beige / 20% dark brown / 10% food colors / 5% near-black
Typography	Bold condensed sans + handwritten script (two-class headline)
Framing archetype	Checklist infographic with circular macro crops
Environment	Flat beige gradient canvas, no scene
Props	Curved hand-drawn arrows connecting circles (flow device)
Emotional promise	”Organized warmth — a plan that feels homely, not clinical”
Copy pattern	Authority/list hook, declarative modules, soft-discovery CTA

Key insight: The two-class headline typography (bold sans + handwritten script for brand name) and the circular-frame + arrow connector system are the most distinctive transferable elements.

Template Produced

The compressed template abstracts competitor-specific elements:

Aspect ratio: 9:16 vertical
Canvas: soft cream gradient, no scene
Composition: two-column editorial — left type, right imagery
Header: two-class type (bold sans + handwritten script for brand word)
Module (5×): all-caps label + one-line description + circular ring-framed photo
Circle frames: heavy dark ring in brand's structural color
Connector: curved hand-drawn arrows in accent color
Lighting: soft overhead, neutral temp, seamless white inside circles
Palette weights: 60% canvas / 20% structural / 10% accent / 5% anchor
Emotional promise: organized warmth — "a plan you can follow"

Variations Generated

#	Variation	What It Tests	Key Change
1	Closest-to-reference	Safe A/B anchor	Tightest formula execution
2	Hook swap	Pain vs. authority hook	”Nežinai, ką pavalgyt?” instead of list
3	Framing swap	Checklist vs. phone-hero	Same palette, different archetype
4	Palette inversion	Color psychology	Orange dominant, green accent
5	Wild card	4:5 before/after split	Different ratio, transformation frame

Each variation includes:

• Full image-generation prompt (Nano Banana / Gemini format)
• Native Lithuanian copy (headline, primary text, CTA)
• Testing hypothesis explaining what it’s meant to validate

Sample Output — Variation 1 (Closest-to-Reference)

Image prompt:

9:16 vertical social ad infographic. Background: soft warm cream gradient,
#F5F0E6 at top fading to #E8E0D0 at bottom, no scene, flat graphic canvas.

Headline band across top 15%: two-class type on one line. Bold condensed
black sans reading "Dienos meniu su" in near-black #1A1A1A, followed by
handwritten italic script reading "fitme" in deep green #2E7D32, followed
by "– sveikas ir greitas" in the same bold black sans.

Below, two-column layout. Left column: five vertically stacked modules...
[continues with full specification]

Copy (Lithuanian):

• Headline: “Dienos meniu su fitme” (22 chars — clears 27)
• Primary: “Skenuoji produktus, gauni visą dieną ant lėkštės. Be streso, be skaičiavimų – tiesiog sveika mityba.” (108 chars — under 125)
• CTA: “Pabandyk nemokamai” (18 chars)

Quality Assessment

Criterion	Result	Notes
Formula identified	✅ 10/10 layers	All concretely articulated
Variations generated	✅ 5 variations	Each with distinct hypothesis
Production readiness	✅ Ready	Prompts executable, copy fits limits
Legal distinctness	✅ Distinct	No Tastier branding, different product
Brand fit	⚠️ Mostly	Brand colors inferred (need to verify hexes)

Review Flags Raised

The skill automatically flagged these concerns:

1. Brand colors inferred — #2E7D32 and #FF6B35 are reasonable defaults but should be verified against fitme.lt’s actual brand
2. Trademark adjacency on V1 — Structurally close to Tastier; consider more differentiation if competing in same market
3. Language confidence — High on headlines, moderate on body copy; recommend native speaker review
4. “fitme” as handwritten script — The reference renders brand name in script; fitme’s actual wordmark is typeset sans — decide if this is ad-only stylistic choice

Analysis

What Worked Well

• ✅ Formula extraction was precise — Lighting recipe, palette weights, typography pattern all captured concretely enough to re-execute
• ✅ Variations are genuinely distinct — Each tests a different axis (hook, framing, palette, format)
• ✅ Copy is native, not translated — Lithuanian reads naturally, not like English-first
• ✅ Review flags are honest — Skill doesn’t hide its limitations

What Could Be Improved

• ⚠️ No image generation — Skill produces prompts but doesn’t generate images; need separate step
• ⚠️ No A/B test execution — Can’t know which variation wins without running ads
• ⚠️ Single reference — Analyzing multiple competitor ads would reveal category-level patterns

Surprises

• 🤔 The two-class typography pattern was the most distinctive transferable element — more than lighting or palette
• 🤔 The arrow connector system is a flow device that works across brands — turns static infographic into “journey”
• 🤔 Palette inversion (V4) is higher-risk than expected — orange-dominant canvas fundamentally changes the emotional promise

Conclusions

1. AI can extract actionable creative formulas — The 10-layer deconstruction produces specifications concrete enough for image models to execute

2. Formula vs. Skin separation works — The output is clearly fitme.lt, not Tastier, while preserving the structural choices that made the reference effective

3. Structured variations beat random exploration — Each variation has a hypothesis; you’re running tests, not generating noise

4. The skill finds real issues — Trademark adjacency, language confidence, brand color uncertainty all flagged proactively

5. Time/cost compression is real — 15 minutes + ~$0.50 vs. days + $500+ for traditional reverse-engineering

Practical Applications

For Brands with No Ads Running

1. Find 3-5 competitors with proven ads
2. Run Ad Alchemy on each
3. Build a “formula library” of what works in your category
4. Generate variations for your product
5. Test with small budget to find winners

For Creative Teams

1. Use extracted templates as creative briefs
2. Designers execute the formula with brand-specific interpretation
3. Reduces “blank canvas” problem
4. Speeds up concepting phase

For Agencies/Consultants

1. Offer “competitive creative audit” as service
2. Produce documented formulas + variations
3. Show clients exactly what competitors are doing (structurally)
4. Deliver ready-to-test concepts

Limitations

• Single ad analyzed — Would need multiple to establish category patterns
• No performance data — Don’t know if variations actually convert
• Lithuanian-specific — Results may differ in other markets
• Static images only — Video ads need different analysis

Next Steps

• Verify fitme.lt brand colors (currently inferred)
• Generate actual images using Nano Banana / Midjourney
• Run A/B test with small budget (€50-100)
• Document which variation wins and why
• Analyze 2-3 more competitor ads to find category patterns

Key Takeaways

• AI can extract reusable creative formulas from competitor ads
• The “formula vs. skin” framework prevents both surface mimicry and wholesale copying
• 10-layer visual deconstruction forces concrete, executable observations
• Structured variations with hypotheses turn creative production into testing
• Time-to-concept drops from days to minutes

Related

• cases/ad-alchemy-creative-reverse-engineering — The skill and approach documented
• competitor-analysis/overview — The broader competitive intelligence landscape
• automation/advisor-strategy — The cost structure that makes this viable
• marketing/ai-video-marketing — Video-first creative strategies
• tools/claude-skills — How to build reusable instruction packages

Sources

• Ad Alchemy skill experiment (Primores internal, April 2026)
• fitme.lt × Tastier output file: ad-alchemy-fitme-tastier-20260422.md

---

Experiment conducted: 2026-04-22