Ad Alchemy — AI-Assisted Creative Reverse Engineering
Ad Alchemy — AI-Assisted Creative Reverse Engineering
TL;DR: A Claude skill that analyzes a competitor’s winning ad, extracts its structural “formula” (lighting, composition, copy skeleton), and outputs ready-to-run image generation prompts + native-language ad copy for your own product. Turns expensive creative reverse-engineering into a 15-minute workflow.
The Problem
Creative teams face a classic dilemma: you see a competitor running successful ads, and you want to learn from what’s working. But:
- Manual analysis is shallow — “it’s blue and has good vibes” doesn’t translate to executable creative briefs
- Wholesale copying is risky — legal issues + your brand looks like a knockoff
- Hiring consultants is expensive — art direction expertise costs $200+/hour
- Starting from scratch is slow — why reinvent when competitors already validated a formula?
The hypothesis: a multimodal AI can do art direction-level analysis if given a structured workflow and the right mental checklist.
The Solution: Formula vs. Skin
The core insight that makes this work:
| Component | What It Is | Transferable? |
|---|---|---|
| Formula | Structural choices: lighting direction, composition grid, focal hierarchy, palette weights, copy skeleton | ✅ Yes — these are reusable |
| Skin | Brand-specific elements: the product, colors, exact wording, model, setting | ❌ No — swap with your own |
A winning ad wins for structural reasons that are invisible to the naked eye. The skill’s job is to articulate those structural choices precisely enough that an image model can re-execute them with a different product.
Two Failure Modes to Avoid
-
Surface mimicry: “Photo of a bottle on a beach, like the reference” — copies the skin, not the formula. Output looks nothing like the source.
-
Wholesale cloning: Copying exact product silhouettes, headlines, or trademarked styling — legal risk and creative dead end.
The sweet spot: the formula transfers, the product is unmistakably yours.
How It Works
Input
The user provides:
- Reference ad — screenshot or image file of the competitor ad
- Their product — name, one-sentence description, what makes it distinct
- Brand colors — hex codes if possible
- Target language — for localized copy (defaults to English)
- Optional: target image model, number of variations, platform
The 6-Step Workflow
Step 1: Visual Deconstruction (10 layers)
The skill walks through a systematic checklist:
| Layer | What to Capture | Example |
|---|---|---|
| Composition grid | Where focal point sits, aspect ratio, eye travel path | ”Lower-third rule of thirds, 9:16 vertical, vertical scroll” |
| Focal hierarchy | Primary → secondary → tertiary visual weight | ”Headline > food circles > labels” |
| Lighting recipe | Key/fill/rim directions, temperature, contrast | ”Soft overhead key, ~5000K, low contrast, no drama” |
| Palette weights | Color % distribution and semantic roles | ”60% warm beige / 20% dark brown / 10% food colors” |
| Typography pattern | Type classes, sizes, placement zones | ”Bold condensed sans + handwritten script in headline” |
| Product framing | Archetype: hero, lifestyle, macro, flatlay, etc. | ”Checklist infographic with circular macro crops” |
| Environment/surface | What product sits in/on/against | ”Flat beige gradient canvas, no scene” |
| Supporting props | Category-level prop description | ”Curved hand-drawn arrows connecting circles” |
| Emotional promise | One-line feeling before reading copy | ”Organized warmth — a plan that feels homely, not clinical” |
| Copy pattern | Hook type, body structure, CTA verb class | ”Authority/list hook, declarative modules, soft-discovery CTA” |
Key discipline: Be concrete. “Golden-hour backlight from camera-right at ~30° elevation” is useful. “Warm lighting” is not.
Step 2: Extract Template
Compress the deconstruction into a reusable spec with competitor-specific bits abstracted out. This template is the most valuable artifact — everything flows from it.
Step 3: Cast Template onto User’s Product
Swap the skin while preserving the formula:
- Replace product with user’s product (same archetype)
- Replace environment/props with brand-appropriate equivalents
- Recompute palette around user’s brand colors (same weight distribution)
- Keep lighting recipe exact — highest-leverage transferable element
- Preserve composition grid and focal hierarchy exactly
Step 4: Construct Image-Generation Prompts
Different engines reward different prompt styles:
| Engine | Style | Best For |
|---|---|---|
| Midjourney | Comma-separated descriptors, --style raw | Photographic ads |
| Flux | Natural language prose, camera settings | Technical control |
| DALL-E / GPT-Image | 2-4 sentences, mood-focused | Quick iterations |
| Nano Banana / Gemini | Product reference + composition description | Product injection |
| Ideogram | Quoted text strings | Embedded typography |
Step 5: Write Native-Language Copy
Write copy directly in the target language — don’t translate from English. Respect:
- Character limits (Meta headline: 27 chars visible, body: 125 chars above fold)
- Formal/informal address (tu/vous, du/Sie)
- Cultural context (holidays, idioms, measurement units)
Step 6: Generate Variation Set
Five structured variations with testable hypotheses:
| Variation | What It Tests |
|---|---|
| Closest-to-reference | Safe A/B anchor — tightest formula execution |
| Hook swap | Same visual, different copy hook (curiosity → pain) |
| Framing swap | Same lighting/palette, different archetype (checklist → phone-hero) |
| Palette inversion | Accent becomes dominant, tests color psychology |
| Wild card | Deliberate departure — highest variance bet |
Output
A structured Markdown file containing:
- Full deconstruction (show your work)
- Extracted template (sanity-check the formula)
- Brand context used
- 5 variations, each with: image prompt + copy + testing hypothesis
- Review flags (language confidence, trademark risk, factual claims)
Real Example: fitme.lt × Tastier
Reference: A Tastier meal-plan infographic ad (9:16, checklist with circular food photos)
User product: fitme.lt — meal-planning and product-scanning app
Template extracted:
- Two-column editorial layout (type left, imagery right)
- Two-class headline typography (bold sans + handwritten script for brand name)
- Five circular food photos in dark rings, connected by curved arrows
- Warm beige canvas, 60/20/10/5 palette distribution
- “Organized warmth — a plan you can follow” emotional promise
Variations generated:
- Closest: Same checklist format, fitme green + orange palette
- Hook swap: “Nežinai, ką pavalgyt?” (pain hook instead of authority)
- Framing swap: Phone-hero with orbiting food circles
- Palette inversion: Orange-dominant canvas, green accents
- Wild card: 4:5 before/after split (“chaos → order”)
Review flags raised:
- Brand colors were inferred (should verify against actual brand)
- V1 structurally close to Tastier — consider differentiation if same market
- Language confidence high on headlines, moderate on body copy
Why This Works
The Advisor Strategy Pattern
This implements the automation/advisor-strategy — expensive expertise (art direction analysis) paired with cheap execution (image generation).
| Role | Who/What | Cost |
|---|---|---|
| Advisor | Claude analyzing the ad | ~$0.10 per analysis |
| Executor | Image model generating the creative | ~$0.05 per image |
| Verifier | Human reviewing outputs | 10 minutes |
Total: ~$0.50 and 15 minutes vs. $500+ and days for traditional creative reverse-engineering.
The Skill Pattern
This is a glossary/skill — a reusable instruction package that bundles:
- Domain expertise (art direction, copywriting, localization)
- Structured workflow (6 steps with quality checks)
- Reference materials (visual deconstruction checklist, copy frameworks, model-specific prompt guides)
- Output format (consistent, actionable deliverable)
See tools/claude-skills for more on building skills.
Limitations
- Static images only — video ads need keyframe extraction and motion analysis
- Prompts only — skill outputs prompts, doesn’t generate images
- Language fluency bounded — rarer languages flagged for human review
- No verification loop — can’t check if generated images match the formula
- Single-ad input — no batch mode for analyzing entire ad libraries
Key Takeaways
- Formula vs. Skin is the core insight — winning ads win for structural reasons that transfer
- Systematic analysis beats intuition — the 10-layer deconstruction forces concrete observations
- Variations should be testable — each has a hypothesis, not just visual noise
- Native copy, not translation — write directly in target language or flag for review
- The skill pattern scales — bundle expertise into reusable workflows
Possible Extensions
- Video variant — keyframe extraction + pacing/motion analysis
- Batch mode — analyze entire competitor ad libraries, cluster by structural similarity
- Image generation integration — actually generate the images, not just prompts
- Performance feedback loop — inform new variations from campaign performance data
Related
- experiments/ad-alchemy-competitor-piggyback — Experiment testing this approach on fitme.lt × Tastier
- automation/advisor-strategy — The cost structure that makes this viable
- tools/claude-skills — How to build reusable instruction packages
- competitor-analysis/overview — The broader competitive intelligence landscape
- cases/agenica-competitor-ads — Another approach to competitor ad monitoring
- marketing/ai-video-marketing — Video-first creative strategies
Sources
- Ad Alchemy skill experiment (Primores internal, April 2026)
- Reddit inspiration post — Web platform that reverse-engineers ad composition (no longer available)