From one ad to ten variations: the template-casting workflow

TL;DR: Template casting is the step that turns a deconstructed reference into a structured variation set. The workflow has six steps: (1) extract the template from the deconstruction, (2) cast the template onto your product, (3) generate the image prompt, (4) write native-language copy, (5) produce the 5-variation structure (closest-to-reference / hook swap / framing swap / palette inversion / wild card), (6) attach a testing hypothesis per variation. Output: 5 ready-to-test ads with stated performance hypotheses for each, in 25–35 minutes of operator time.

Where casting sits in the overall workflow

The full reverse-engineering workflow has six phases:

Pick a reference (from Meta Ad Library winner identification)
Deconstruct it using the 10-layer framework
Extract the template ← template casting starts here
Cast the template onto your product ← the heart of this cluster
Generate image prompts + copy per variation ← finishing this cluster
Test, measure, scale winners

Casting is where abstract analysis becomes executable production. It’s the step most teams get wrong because it sits between the “creative” and “operational” parts of the workflow — easy to under-invest in.

Step 1 — Extract the template from the deconstruction

The deconstruction produced 10 paragraphs of structural observations. The template compresses those into an abstracted specification with competitor-specific details stripped out.

A complete template has:

Composition grid: where the focal point sits, what the secondary and tertiary elements are, negative-space distribution, aspect ratio
Lighting recipe: key-light direction, elevation angle, quality, color temperature, fill ratio, rim-light presence
Palette weights: dominant %, secondary %, accent %, with semantic role
Product framing archetype: hero-on-pedestal, lifestyle-in-use, macro-texture, levitation, flatlay, hand-held, before/after, founder-selfie
Environment / surface: what the product sits in / on / against
Supporting props: category-level description (not item-level)
Typography pattern: class, weight, case, placement zone
Emotional promise: the one-line pre-read emotional trigger
Copy skeleton: hook type, body structure, CTA verb class

The template should read like a production specification — abstract enough that the competitor’s specific product is irrelevant, specific enough that an image model can execute it reliably.

Show the template verbatim in the final output so the user can sanity-check the formula you extracted. If the template feels thin (“warm lighting, product centered, clean copy”), go back and make it more specific. Vague templates produce vague output.

Step 2 — Cast the template onto your product

Casting means applying the template to your product while preserving what’s structural:

Start from the template fields as constants.
Replace the product framing with your product — same archetype, different object.
Replace environment and props with ones that serve the same structural role but feel native to your brand.
Recompute the palette around your brand colors while preserving the weight distribution and the “accent = CTA magnet” logic.
Keep the lighting recipe exact. Lighting is the highest-leverage transferable element and the one most teams are tempted to vary. Don’t.
Preserve composition grid and focal hierarchy exactly. Same placements, same sizes, same hierarchy order.

The rule: hold structure constant, swap skin. If you find yourself wanting to change the composition or lighting because “it would be more on-brand,” stop. You’re surface-swapping, not structurally casting.

Step 3 — Generate the image prompt

Image prompts come from the casting output, formatted per the target image model. The image prompt pillar has per-model guidance; the summary for casting purposes:

Lead with subject and framing — product + archetype + key placement
Then lighting — direction, elevation, quality, color temperature
Then environment and palette — specific hex codes if the model honors them
Then lens/camera language — focal length, depth of field, perspective compression (separates “AI-looking” from “shot-looking”)
Then medium/finish — photography style, retouching level, grain
Then negative prompts / avoid list — what the reference deliberately omits
End with aspect ratio and model-specific parameters

For each variation, produce one fully-formed prompt, not a comma-separated list of fragments (unless targeting Midjourney). If your target model supports reference-image input (Nano Banana / Gemini, GPT-Image, Flux Kontext, Midjourney with --iw), include a line like [Attach product reference: <filename>] so the operator knows to upload the product photo alongside the text.

Step 4 — Write native-language copy

Copy is written natively in the target language using the copy skeleton from the template.

Headline obeys the platform limit (Meta: 27 chars visible on mobile feed, hard cap 40; TikTok: 40 chars as the no-truncation line).
Primary text/body — put the hook in the first 125 chars (Meta’s “above-the-fold” cut-off).
CTA — short action phrase matching the reference’s CTA verb class.
Optional description — 30 chars, only if the template uses one.

Don’t draft in English and translate. Translation kills rhythm, idioms, and character count. If you’re not confident at native level, flag the specific lines for human review — explicitly, in the output.

Full guidance in the ad copy pillar. The short version: respect the template’s hook type and body structure, write natively, respect character limits, flag localization traps.

Step 5 — Produce the 5-variation structure

The default variation set is five, structured as:

Closest-to-reference — the tightest execution of the formula, minimal deviation. The safest A/B starting point.
Hook swap — same visual template, a different hook type. If the reference uses curiosity, try contrast or social proof.
Framing swap — same lighting and palette, a different product framing archetype. Hero → lifestyle, or lifestyle → macro.
Palette inversion — same composition and framing, but the accent color becomes dominant and vice versa. Tests whether the formula works at a different color weight.
Wild card — one deliberate departure from the template along whichever single axis seems most likely to yield a usable-but-different output. Flag it as the higher-variance bet.

If the user wants a different count, scale accordingly but always include #1 (closest-to-reference) and #5 (wild card) as the extremes. They represent the safe floor and the ambitious ceiling of what the template can produce.

Step 6 — Attach a testing hypothesis per variation

Every variation needs a stated hypothesis. Without one, variations become arbitrary — indistinguishable from random creative variance.

Example hypotheses:

Closest-to-reference: “Formula holds; expect baseline performance matching category average.”
Hook swap: “Curiosity hook converts better for our product than the reference’s pain hook, because our audience is earlier in the buying cycle.”
Framing swap: “Lifestyle framing outperforms hero on our audience because they value social proof over product-purity signaling.”
Palette inversion: “Our accent color works as dominant because our brand is already associated with it, providing instant recognition.”
Wild card: “Adding motion blur to the product creates urgency/action feeling appropriate for our category.”

The hypothesis serves two purposes: it tells the media buyer which variations to expect to win under which audience conditions, and it provides the retrospective lens when results come in — you’re not just ranking ads, you’re validating or invalidating specific hypotheses about the audience-formula fit.

A realistic timing breakdown

Steps 1–6 timing for an experienced operator using ad-alchemy or equivalent:

Step	Time
1. Extract template	5 min
2. Cast to your product	5 min
3. Image prompts (×5)	10 min
4. Copy (×5, native language)	10 min
5. 5-variation structure	(baked into 3+4)
6. Hypotheses (×5)	5 min
Total casting phase	~35 min

Plus the upstream deconstruction (~10 min) and downstream QA (~5 min), the full per-ad workflow is 45–50 minutes from reference to five ready-to-test variations.

Key takeaways

Casting is six steps: extract template → cast to your product → image prompt → copy → 5-variation structure → hypothesis per variation.
Hold structure constant, swap skin. This is the single most important discipline in the workflow.
The 5-variation structure (closest / hook swap / framing swap / palette inversion / wild card) isn’t arbitrary — each variation tests a specific axis.
Every variation needs a hypothesis. Without it, you’re producing variance, not signal.
~35 min casting, ~45 min full workflow once practiced.

seo/ai-creative-reverse-engineering-complete-methodology — the pillar
surface-vs-structural-mimicry — why casting must be structural
glossary/ai-creative-reverse-engineering — canonical definition
glossary/visual-deconstruction — the deconstruction that feeds casting
glossary/framing-archetype — the archetypes referenced in Step 2
how-long-reverse-engineer-ad — full workflow timing
ad-copy-for-reverse-engineered-creative — Pillar 6 for copy craft
ai-image-prompts-for-performance-creative — Pillar 5 for image prompts

Sources

Primores ad-alchemy skill specification — the 5-variation structure and casting discipline.