Fast flow

0) Hypothesize → 1) Shoot one photo → 2) Generate visuals (Nano-Banana) → 3) Generate angles (GPT-5 mini) → 4) Build lean test matrix → 5) Launch → 6) Analyze → 7) Scale.

0) A/B testing hypothesis (do this first with GPT-5 mini)

Use GPT-5 mini to create a hypothesis matrix of image angles you’ll test. Keep one variable per test (visual or copy).

Prompt (copy-paste)

You are a senior experiments strategist. Build an A/B test hypothesis matrix for [PRODUCT] to [AUDIENCE] on [CHANNEL: Meta/Pinterest/Email].

Constraints:
• Start with 6 high-contrast hypotheses.
• Each hypothesis must isolate ONE variable (Visual-only OR Copy-only), name the “control”, and define the “variant”.
• Map each to a visual archetype: Studio hero, Lifestyle scene, UGC phone shot, Color-block flat-lay, Seasonal backdrop, Pattern/texture.
• Angle buckets for copy: Benefit-led, Proof/Review, Objection-buster, Comparison, Offer/Promo, Feature highlight.

For every hypothesis output a row with:
(H1 ID) | Variable isolated | Visual archetype | Copy angle | Expected effect (+CTR/+CVR/-CPA and why) | Primary KPI | Success threshold | Min sample-size rule (per variant) | Run time cap | Kill/scale rule.

Then propose ONE “phase-1 plan” of ≤8 variants that covers the widest contrast while respecting isolation.

Tip: If you’ll mainly test images, tell GPT-5 to make 4–6 visual-only hypotheses and 1–2 copy-only for email.

1) Capture a couple of good base photos

• Lighting: bright window, soft shadows.
• Background: neutral or sweep.
• Name Files: sku_base.jpg (45° or front), sku_label_close.jpg (sharp label).

2) Generate visuals in Nano-Banana (per your hypotheses)

Always attach the base photo and label close-up. Add guardrails to prevent drift.

Studio hero

Create a clean studio hero at 1200×1200 (JPEG, sRGB).
Use attached product + label close-up. Product centered ~88% height.
Pure white #FFFFFF, soft contact shadow.

Guardrails: preserve label artwork/text exactly; keep geometry/proportions; no color shifts.