Ecom AI Image Ads (Plan → Prompts → Variants)

This skill produces a daily creative plan and prompt packs you can run in your image tool of choice (Gemini/MJ/SDXL), built from your angle matrix.

Required inputs

• •Workspace with foundation docs + ads/ads.csv (from ecom-foundation-docs)
• •Competitor ad examples (URLs, screenshots, or notes on what’s working)

Required outputs (create these files)

In outputs/ecom/<brand_slug>/<run_id>/creative/:

• •plan.yaml (what to make; format/aspect/angle mapping)
• •prompts.md (copy/paste prompt pack with variants)
• •winner_variations.md (prompt pack for “20 variants of winner”)

Daily creative plan rules (DTC)

For each angle, generate a balanced set:

• •UGC look (authentic, imperfect, believable)
• •Product hero (clean, DTC aesthetic, benefits visible)
• •Proof-first (reviews, demo frames, ingredient/mechanism visual)
• •Comparison/how-to (simple “why it works” visuals; avoid policy landmines)

Aspect ratios to cover:

• •1:1, 4:5, 9:16

Prompt structure (recommended)

Each prompt should include:

• •subject + setting (UGC/home, bathroom, gym, kitchen, etc.)
• •mechanism/proof (what’s visually different)
• •shot type (selfie, handheld, product close-up, before/after only if true/allowed)
• •lighting + lens (phone look vs studio)
• •negative constraints (no text overlay, no logos, no medical claims)

Winner → variants loop

When you identify a winner, generate 10–30 variants that preserve the “why it worked”:

• •change background, wardrobe, framing, angle, props
• •keep mechanism + core visual cue consistent
• •keep “UGC authenticity” (avoid over-polish)

Output these as winner_variations.md with explicit instructions for subtle vs moderate changes.

Registry tagging (for learning loop)

Tag each creative with:

• •brand:<slug>
• •product:<slug>
• •angle:<angle_id>
• •format:<ugclook|product|proof|comparison|howto>
• •aspect:<1:1|4:5|9:16>

Use ecom-learning-loop to store prompt → performance so you can generate variants from data.