Hey all, happy Friday,

Big news. James has been in school all week so far.* I am writing this on Thursday morning. I repeat: four days of childcare that do not involve Maddie and me taking turns asking, “What time is your call again?”

*UPDATE, five minutes after writing the above: daycare called. James has a fever. We are currently watching Sesame Street as I continue to write this.

And I will say, Elmo hasn’t aged a day since I was a kid (many eons ago… Hercules anyone?). That little red monster has gotta drop his skincare routine.

Let’s get into the important stuff:

Saturday night, Maddie and I went on a date night to Nitehawk Cinema. For those outside NYC, it’s a retro(ish) theater that serves booze, food, and popcorn. For those inside NYC, you already know. Their popcorn became an obsession of mine, and during the pandemic, given my deprivation, I reverse-engineered the recipe at home.

Truffle oil, butter, salt, citric acid. For those wondering.

Sunday was Grateful Dead for Kids at Brooklyn Bowl in Williamsburg. If you have not experienced a Grateful Dead for Kids show, it is a jam band concert with a parachute (and beer if you are 21+!). Hard to beat tbh as a parent tbh.

James was too funny: he scooted to the front of the crowd, leaned against the edge of the stage, stood up, and spent the rest of the “set” goggling up at the band. In jam band circles, this is called rail-riding (and probably other circles, too). You need to get there early to nab such a coveted spot, and typically, the people who do it consistently are a bit unusual. James was up there by song two. We will keep an eye on him.

(Fat) Linus update: We dropped off Linus with our friends in Williamsburg so he could romp with his two dog pals while we were at the show. I was hoping this exercise would address what multiple sources now describe as the “few extra pounds since James arrived” situation. Apparently, Linus did not romp; he sat on the couch, cuddled up, and stared at the other dogs while begging for food.

I took him to a very long park excursion later in the week, and he chased a singular squirrel for a solid hour. I mark that as progress.

Heads-up: There will likely not be a newsletter next week. I will be in Las Vegas at the Sphere for Phish.

A few months ago, I told you Nano Banana 2 changed the static ad game. It did. But this is the AI era we are living in, and changing the game lasts as long as James stays fever-free.

ChatGPT Image 2 landed. It is marketed as a competitor to Nano Banana. I did not believe the marketing because, well, marketing (and because OpenAI drops have been mid: often dropping an incredible model at first, then saturating it over time).

Good news for anyone sitting on a library of Nano Banana prompts + using my brand spec card system: they work with GPT Image 2.

TFG.

You need a few tweaks to get them perfect, but if you are okay with "great," you can probably just move to that model with your existing infrastructure.

I spent the week testing both on a real brand with editorial chops: Jones Road (that Motion bootcamp has me brainwashed into thinking there are only three brands).

15 formats. 30 generations. About six bucks.

Here is what I learned:

If you just want the prompts + MD files, and don’t care to read the rest, here you are:

This is the backbone of the newsletter. I ran 15 different formats through GPT Image 2 and Nano Banana 2, using the same prompt, spec cards, brand references, and hero product. Let me show you what GPT 2 produced:

Fifteen different vibes. One day of generation.

Six months ago, getting one of these right was a morning of work. Now you get fifteen distinct looks before lunch.

The bottleneck is increasingly no longer execution.

It is taste.

GPT won 15 of 15 on this brand.

The first three prompts I tried on GPT all put the headline inside the Instagram UI. y~100 of the frame.

Classic.

The standard safe-zone coordinates that Nano Banana respects by default? GPT treats them as a suggestion.

I tried three interventions in Round 1. Stronger language. An uploaded safe-zone template PNG. A rewritten block. All three failed. Same composition every time.

The crack landed in Round 2, and it is specific enough to matter.

I call it E+.

Three things stacked:

The full E+ block is in the Notion companion (linked below). Copy-paste verbatim. It locked my output across all 15 formats. Zero safe-zone failures after.

A. Social platform UI is rendered at near-pixel accuracy. The TikTok bottom nav, the right-rail interaction column with engagement counts, the spinning music disc, the @handle with timestamp. All there. Nano Banana produces “a TikTok-like thing” with stylized icons. GPT produces TikTok. Same for IG Story and IG Reels. If feed-native creative is part of your program, this is a capability gap, not a style gap.

B. Handwriting is the biggest single-format gap. GPT produces convincing ballpoint pen imperfection. Pressure variation, letter drift, imperfect heights. Looks like a person wrote it. Nano Banana defaults to a uniform handwriting font, which is the wrong answer. Post-It, letter, and whiteboard formats all go to GPT.

C. Dramatic lighting and monochromatic tonal baths. Dark mode with spotlight. Immersive rose tonal bath with fabric drape. Single-hue palette discipline. GPT commits to the brief. NB2 retreats to pale cream defaults when you ask for extreme palettes.

I wrote up the 15 comparisons into a Notion companion doc alongside every prompt (I gotta give you something!).

Every output side-by-side. Every format-specific calibration I found. You can plug in your own brand variables and ship.

For the Claude Code power users: I also built a GPT Prompt Converter markdown. Drop it into any Claude project. Fill in a brief block at the top (brand name, format, copy, scene language). Claude reads the file, assembles the full GPT-Image-2-ready prompt, and hands you something to paste into fal.ai or the OpenAI API.

No code. No scripts.

The markdown does the work.

(also inside the companion doc)

Pair all of this with the Brand Research Prompt from the Nano Banana playbook.

That part still works for both models!

The LinkedIn version of this post went up yesterday with the hook “GPT Image 2 cost me six months. But I do it for you.”

This is far more comprehensive.

Reply with which template you used and what it generated. I am building the dataset. Screenshots welcome. Failures especially welcome. I learn the most from the things that do not work.

Have a great weekend,

Will

Heads-up (x2): There will likely not be a newsletter next week. I will be in Las Vegas at the Sphere for Phish.