All 15 head-to-head ad comparisons from this week’s bake-off. Every prompt. Every output. Side by side.
Brand tested: Jones Road Beauty: Miracle Balm (shades rotated across formats) Scope: 15 ad formats × 2 models = 30 generations Cost of the run: under $3.20 total Prompt system used: E+ safe zone + brand spec card v2 with “Type in Context” examples (locked after 3 failed interventions in Round 1)
For Claude Code power users and anyone running local markdown workflows: drop this into any Claude project, fill in a short brief block at the top, and Claude assembles a full GPT Image 2 prompt you paste into fal.ai, ChatGPT, or the OpenAI API.
Covers all 15 formats in this doc plus a few extras. Bakes in the E+ safe zone, brand typography block, and format-specific exceptions (Post-It allows paper, UGC allows platform UI, etc.). No code. No scripts. Paste-and-go.
Each entry has:
[BRACKETS] to swap in your own brand)Every prompt below prepends the same E+ safe zone block and brand typography block. The converter markdown handles that automatically.
The scroll-stopper. A casual handwritten note stuck to the product lid. Tests whether the model can produce ballpoint-pen imperfection or defaults to a handwriting font.
[PASTE E+ SAFE ZONE BLOCK + BRAND TYPOGRAPHY BLOCK]
FORMAT: Post-It Note on Product (Handwriting Style A).
SCENE: A real photograph of [YOUR PRODUCT] on [SURFACE, e.g. "cool white marble bathroom counter"]. Full-bleed continuous surface. Shallow depth of field. Soft overhead natural light.
POST-IT: YELLOW Post-It note stuck to the top of the closed product lid. Casual black ballpoint pen handwriting, all lowercase, with natural ink pressure variation and imperfect letter spacing. NEVER a handwriting font with uniform spacing — that fails the format.
EXCEPTION: The Post-It IS the intended physical subject. The surface it sits on continues past it to all image edges.
COPY:
Handwritten note (lowercase): "[YOUR NOTE, under 20 words, signed]"
CTA button: [YOUR CTA, ALL CAPS]
Kicker: [CONTEXT, small mono]

GPT Image 2

Nano Banana 2
This was the widest single-format gap in the entire test. GPT’s handwriting has genuine ballpoint pressure variation, imperfect letter heights, natural line drift. NB2 defaults to a uniform handwriting font AND doubled the kicker line (a rare but real NB2 failure mode). GPT also executes the cool white marble bathroom counter with shallow depth of field; NB2 ignored the surface spec.