Hey all, happy Fridizzle,

We got 20 inches of snow in New York City on Sunday. James's daycare closed for the first time in 10 years, which is apparently unprecedented. So I did the only reasonable thing: I took my cross-country skis and Linus (the dog, for anyone new here) and skied from my in-laws' apartment across several avenues into Central Park while all the roads were closed.

Did a lap around the reservoir. It was a total whiteout. My goggles fogged up completely, and I genuinely got lost on the perimeter, which is embarrassing because it's a loop. Linus had the best day of his life. Top five for him, easily.

Monday was still technically a work day, though, so Maddie and I did what every modern parent does when they need 45 minutes of uninterrupted focus: we tried to put James in front of Sesame Street. He's 13 months old. He looked at the TV, looked back at us, picked up a ball, and threw it at my face. He wanted to play catch. For an hour. Which is cute but also slightly concerning, given that neither Maddie nor I has what you'd call good hand-eye coordination. The athletic gene may have skipped a generation. Or two.

So yeah. Very productive snow day. Really got a lot done.

Anyway. Nano Banana 2 came out TODAY. That's right, I am writing this on Thursday night, because I love you guys (and understand it’s a one-way street).

For anyone new: Nano Banana is the nickname for Google's Gemini image model. Nano Banana 2 is Gemini 3.1 Flash Image. I stopped doing everything I was supposed to do to test one question: can this thing actually make ads?

Not concept mood boards. Not "cool AI art." Real performance creative you could put in a Meta campaign without apologizing first.

I grabbed 3 existing Seed Probiotic ads as reference images and tried to generate every ad format a performance team actually runs. 14 formats. 3 hours of testing. Several skipped meals.

Here's what I found. And more importantly, here's what went wrong first.

1. Text rendering is solved. Headlines, body copy, fine print, star ratings, even a full Amazon review card with a "Verified Purchase" badge. Zero typos across 14 formats. The era of "YOTUR GUT, OPTMIZED" is over.

2. Brand matching from reference images works. I didn't give it hex codes or font names. I attached 3 existing Seed ads and said "match this." It pulled the exact Seed green, the exact type weight, the exact vibe. It reads your brand guide through the pixels.

3. Camera specs are the biggest quality lever nobody talks about. This is the single most important hack I found. More on this below.

"Professional photo" gives you stock. "85mm f/1.8, medium format sensor" gives you editorial. Massive shoutout to Linus Ekenstam who demonstrated this with portrait tests and saved me an hour of figuring it out myself.

4. The UGC output broke my brain. I asked for a TikTok screenshot with full native UI. Heart icon. Comment count. Spinning music disc. I showed it to my team and nobody flagged it as AI. That's either impressive or terrifying and I've decided not to think about which one.

5. The model will clone your references if you let it. Without explicit anti-cloning instructions, it just remakes your existing ads with slightly different vibes. Always prepend: "Use the attached images ONLY as brand reference. Do NOT replicate the layouts or compositions shown."

Step 1: Create a Claude Project for your brand.

Step 2: Run the Brand Research Prompt below. Claude extracts your actual fonts, colors, photography style, and visual identity.

Step 3: Pull 3-5 reference images: hero product shot, website screenshot, 1-2 existing Meta ads, and a lifestyle shot.

Step 4: Paste any template from the below ad types into the Claude Project. Claude adapts the [BRACKETS] using the Brand DNA it already researched.

Step 5: Copy the adapted prompt + reference images into Nano Banana 2. Generate.

As a reminder, I always use fal.ai to generate multiple versions. This is the more expensive route, but it is far more efficient than doing it directly in Gemini.

Start here. No text to trip up the model.

Tests text rendering. If this comes back clean, you're golden.

This is the one.

[IMAGE: Social proof — C0c7KypmFH1eBO87ymwEl_WTXj5h11.png]

This one failed until I showed Claude actual examples of the format. See above.

Has to look like a real person posted it.

Same story as Us vs Them. Had to show Claude examples before it understood the format.

This is the single biggest takeaway from the entire test.

Some ad formats are so industry-specific that neither Claude nor Nano Banana 2 knows what they look like out of the box. "Negative marketing" to us means a fake 1-star Amazon review that's actually a rave.

To Claude, it meant a confrontational attack ad.

The first prompt Claude wrote me came back from Nano Banana with: "Your probiotic is probably dead before it reaches your gut."

Black background. Aggressive copy. Completely wrong.

Same with Us vs Them.

Claude wrote me a prompt for a "comparison ad" and Nano Banana generated an artsy photographic diptych. Beautiful.

Not how a Us vs Them ad works in performance creative. Those are structured grids with checkmarks and X marks.

Here's what fixed it: I uploaded actual examples of these ad formats into our Claude conversation. Screenshots from my ad creative swipe file. I showed Claude what negative marketing actually looks like in our world, what a Us vs Them grid actually looks like, and said "write me a prompt that describes THIS."

References I Used For Negative Marketing:

Reference I Used For Us vs. Them:

Claude then wrote prompts specific enough that Nano Banana 2 generated the format correctly without ever seeing those reference images itself. The model only got brand references and the prompt. Claude was the translator between my visual references and Nano Banana's generation.

The Us vs. Them marketing went from this to this:

The negative marketing went from attack ad to this:

The workflow: show Claude what you want → Claude writes the prompt → give Nano Banana 2 the prompt + brand refs only. Claude handles format knowledge and prompt engineering. Nano Banana handles pixels and brand matching.

Product description specificity has a ceiling. Once you hit "dark green translucent glass jar with visible nested capsule," more detail doesn't help. But below that threshold, it hurts a LOT.

Generate 3-4 variations. The first output is rarely the best. But Nano Banana 2 is more consistent than anything I've tested: your third output refines the first instead of being completely different.

Build your ad format swipe file now. When Claude doesn't know what a format looks like, show it. Upload 3-5 examples of every ad format you run regularly. Claude translates the visual into a prompt. Nano Banana generates it. Your swipe file is now training data, not just inspiration.

I want to be clear: this isn't "AI replaces your creative team." That take is lazy and wrong.

This is: your team can concept 50 ideas before lunch instead of 5. The brief-to-concept phase just went from days to minutes. You still need humans to judge what's good. But the brainstorming bottleneck is gone, and the brands that figure this out first are getting 10x more concepts before they spend a dollar on production.

That creative velocity gap is about to become the whole game.

This workflow is somewhat of a manual version of what we're building at Skipper. Skipper scrapes your reviews, discovers who's buying and why, and generates ad creative targeted to real buyer personas.

The beta is open: getskipper.ai

Okay. I'm going to go make dinner (or more realistically, pick up dinner) before Maddie realizes I've been writing about AI prompts for the last four hours instead of being a functional adult.

If you try any of these prompts, reply and tell me which format worked best. I'm genuinely curious.

Have a great weekend,

Wil