Quick life update for those of you here for the parasocial relationship, not the AI workflows (I see you, and I respect it 🤝).

I finally started “A Knight of the Seven Kingdoms.” Yes, I know. I'm late. I read all the Song of Ice and Fire books, was appropriately devastated by the Game of Thrones ending like everyone else, and have been dragging my feet on anything HBO touches in that universe since. But I started it last night while doing Claude Code because my evenings are now "HBO and Claude," which I'm told is not what people mean when they say "Netflix and chill." It's so good (I just finished episode 3… holy guacamole!).

James update: the kid is fully mobile. Pulling himself up on everything, pushing himself around the apartment, standing on his own, and looking at us like "what are you going to do about it?" His cousin Anu came over on Sunday and blew bubbles in his face for 20 minutes straight, and I have never heard a human being laugh that hard. He also ate raw broccoli by choice the other night. So did Linus. Only one of them had bad gas.

We're taking him to Gun Hill Brewery in Industry City on Saturday because my guitar teacher is performing in an Allman Brothers cover band, and James is a genuine music fanatic (born a Ramblin’ Man?).

Full Linus report: the park has been a mud pit all week. It was pouring rain this morning, and I was one of five people there because I feel guilty if I don't take him. He does not feel guilty about tracking mud through the apartment afterward.

Anyway.

I need to be upfront about something. I almost didn't write this newsletter.

Last week, I showed you Nano Banana 2 and made 14 ads with Seed Probiotic. That was the fun "look what this new toy can do" issue.

Well, now I’ve built an actual workflow that takes one winning ad and turns it into 15 visually distinct on-brand formats, pulling from structured research, with every prompt templated so you can do it for any brand in an afternoon.

I wanted to keep it for client work. This is the kind of thing you charge for, not give away in a newsletter.

BUT my team told me to share it. Janet's exact words were "Will, you literally built this newsletter to give people stuff they can't find anywhere else, stop being weird about it."

She was right. She's usually right. Don't tell her I said that.

So here it is. The whole system. Every prompt. Nothing held back.

But first, let me tell you why this matters and what we're building around it.

Here's the honest truth: the research half of this workflow is a very simplified version of what Skipper (getskipper.ai) does.

Step 1 of what I'm about to show you is a big research prompt. It uses Claude's web search to dig into a brand's identity, visual system, product details, pain points, competitive positioning, and pre-built hooks. It's good. It works. It will make your ads dramatically better than winging it.

But it's not even close to what Skipper captures.

Skipper doesn't just search the web. We built custom scrapers over a year that read Reddit threads, Facebook comments, Amazon reviews, TrustPilot, your own site reviews, really anywhere on the internet where your products are named. Every edge case. Every platform's weird formatting. A year of engineering so we could capture what real customers actually say about you in places a web search will never find.

From that data, Skipper identifies real buyer personas from real language, not demographics. It extracts pain points, emotional triggers, and proof points from comment sections and review threads, not from your marketing site. Then it generates net new ad concepts targeted to those actual human segments.

I'm sharing it because (a) even the simplified research prompt makes this workflow useful, (b) it gives you a taste of what the full Skipper research looks like if you want it done for you.

Skipper is in beta: getskipper.ai

Now. Let me give you the whole thing.

Last week: throw brand reference images at Nano Banana 2, hope they match, iterate when they don't.

This week: build a structured brand research brief and two visual brand reference cards FIRST, then use those as the foundation for every single ad you generate.

The difference is MASSIVE.

Last week, I was getting maybe 60% brand accuracy on a first generation. With the brand cards, I'm getting 90%+. The model isn't guessing your brand anymore. You're telling it exactly what the brand looks like, and it listens.

This is the unlock. Not Nano Banana 2 itself (that was last week's unlock). The unlock is the system that makes Nano Banana 2 actually usable for real brands with real guidelines.

Here's the system:

Step 1: Research the brand and generate a structured brief (one-time per brand)

Step 2: Create two visual brand reference cards from that research (one per brand)

Step 3: Run 15 format-specific prompts that iterate on your winning ad, pulling all brand details from the brief

(Quick note since I got a bunch of DMs about this last week: I've fully moved to Claude for all workflows. I, of course, am still using Gemini models for image generation)

Specifically, their Miracle Balm "Tired of guessing what shade is right for your skin?" campaign. It's a genuinely great ad. Clean message, strong CTA ("Take the Quiz"), proven performer.

This is the foundation. You run it once per brand in a Claude Project with web search enabled, and it outputs a structured brief that every ad pulls from.

The prompt runs through 7 phases: brand identity and visual system (fonts, colors, hex codes, design rules), product intelligence (ingredients, pricing, offers, hero claims), proof points and social proof (stats, press mentions, testimonials, review metrics), transformation and pain points (customer problems in their own words), competitive positioning (differentiators, us vs. them angles), founder story (origin, credibility, quotable moments), and ad-ready hooks pre-built for all 15 formats.

It outputs everything as structured JSON. The JSON becomes your single source of truth. No more guessing what hex code to use. No more hoping the model figures out your typography. Every format prompt in Step 3 pulls from this brief.

The full research prompt is in the doc. It's long (7 phases, ~200 lines) which is another reason I moved it out of this email. Gmail would have clipped this newsletter somewhere around Phase 3 and we both know you wouldn't have clicked "view entire message."

Save the output. Put it in the Claude Project library. This is your brand bible. Every prompt in Step 3 pulls from it.

This is the big unlock compared to last week.

Last week, I uploaded existing ads as reference images, hoping Nano Banana 2 would reverse-engineer the brand. Sometimes it worked. Sometimes I got "close enough green" instead of "Seed green." The text rendering was solved, but brand consistency was a gamble.

Brand reference cards fix that. Instead of making the model guess, you're giving it an explicit visual spec sheet. The difference in consistency is night and day.

And you don't have to build these manually. You prompt Claude to generate them straight from the research brief.

Card 1: Brand Spec Card covers logo, typography system with sample text in each font role, color palette with hex swatches, design do's and don'ts, and CTA button style. Claude generates it as an HTML page from the research data. You screenshot it.

Claude generates the HTML. Open it in your browser, screenshot it as a PNG. That's your Ref 01.

Card 2: Visual Style Card covers brand essence adjectives, founder quote, photography direction (product, skin, lifestyle, backgrounds), always/never rules, product styling notes, and mood spectrum. Same deal: Claude generates the HTML, you screenshot it.

The prompts for both cards are in the doc. Claude already has all the research data from Step 1, so it builds these cards with the actual brand details filled in. No copy-pasting hex codes from a PDF. No hunting for font names. The research feeds the cards, the cards feed Nano Banana 2.

For every ad you generate, upload:

These three stay the same across all 15 formats. The brand cards are doing the heavy lifting that the reference images failed at last week.

(Yes, I'm telling you to make a mood board for a robot. But at least now the robot is also making the mood board. Welcome to 2026.)

Every prompt is a template. The [BRACKETS] get filled by the research brief from Step 1.

You don't fill them manually. You paste the template into your Claude chat (where the research brief already lives), and Claude fills the brackets using the brand data it already has. That's the whole point of using a Claude Project.

Everything about the brand's visual identity, colors, typography, CTA style, photography direction, all of it, comes from the research. None of it is hardcoded. If Jones Road's CTA is a black rectangle with white text, that comes from the brief. If your brand uses a rounded coral button, Claude adapts accordingly.

Every prompt is a template. The [BRACKETS] get filled by the research brief from Step 1. You paste the template into your Claude Project, Claude fills the brackets using the brand data it already has, and you copy the adapted prompt into Nano Banana 2 alongside your brand cards.

Everything about the brand's visual identity, colors, typography, CTA style, photography direction, comes from the research. Nothing is hardcoded. Claude adapts each template to your brand automatically.

All 15 prompts are in the doc. What I'll show you here is what each format looks like and why it works, so you know which ones to prioritize for your brand.

One winning ad. 15 iterations. Same proven message, 15 different visual containers. Everything pulls from one research brief. That's the system.

If you try this for your brand, reply and tell me which format performed best. I'm building a dataset of which formats work for which categories, and your data would genuinely help.

Have a great weekend,

Will