Hey all, happy Friday,

Last weekend, we drove out to Long Island to see my dad and stepmom. The big news from our house is that James now has two words on heavy rotation: “Dada” and “Buh-bye.” “Dada” means roughly forty different things depending on the hour. “Buh-bye” means bye, only bye, and he deploys it with the finality of a man sailing off to sea in search of whale blubber (too niche?).

We hit the beach both days and learned something important: James is a beach kid. He sat in the same orange bucket I played with as a kid for the better part of an hour, and he only tried to eat the sand once, which tells me Suffolk County sand simply cannot compete with the Rhode Island sand he sampled a few weeks ago (they must put calamari in it or something). We dipped his toes in a 60-degree ocean, and it splashed him to the belly, and he loved it. This kid is water through and through.

Sunday, we went to the Upper East Side for Father’s Day dinner at Maddie’s parents': fried chicken, mac and cheese, cucumber salad, the works. James and his cousin Anu went hard on the chicken and then even harder on the chocolate cake, which produced the Sugar Rush of the ages: the two of them up on the bed, a full pillow fight, and, at some point, James losing all his clothes and running the operation in a diaper. I cannot tell you how. Meanwhile, Linus and my in-laws’ dog Scout raided the apartment for every squeaky ball they could find. It was pure chaos, but also pure joy. Writing this out is genuinely making me happy.

Tonight, Maddie and I are seeing the new Spielberg film, Disclosure Day. And tomorrow, the actual reason I’m writing this on a Wednesday, I’m going to MetLife with Maddie’s brother for Germany vs. Ecuador. I have never been to a World Cup game. I will be MIA tomorrow afternoon, and may God have mercy on me at Penn Station.

Every time a new model drops, my feed fills with the same question: how does it handle creative strategy?

I keep landing in the same place: the question is aimed at the wrong half of the job.

Here is the cleanest way I can say it after months of building this for real.

An AI is the best historian your ad account has ever had. Point it at six months of spend, and it will tell you, to the dollar, every persona that worked, every angle that scaled, every emotion that converted, and which of your products is quietly carrying the whole thing.

That is enormous, and it is real ROI. I’ll show you exactly how I set it up.

But a historian has never once predicted the future (this is probably not true, but work with me here). The next angle, the one nobody has run yet, the one with no row in your data because it has never existed, is still a human job. That is not a temporary gap; the next model closes. That is the actual division of labor, and once you see it, you can stop fighting it.

I’m consulting with a skincare brand right now (only 3 more spots open for Q3!), and last week we spent an entire session on the least glamorous, highest-leverage thing in the stack. No generations, no pretty outputs. Just me narrating Meta’s developer portal over a screen-share for an hour, which is the riveting content you all subscribed to.

But it’s the part that makes everything else work: wiring their Meta account into Claude through the Marketing API.

The setup is more boring than it sounds, and I’ll spare you the click-by-click (it’s in the kit, where boredom belongs).

Short version:

Once the token is live, here is what you need to tell the machine to do:

Pull every ad the account has ever run, tell Claude to have Gemini review the statics and videos, and tag them: which product, which persona, which angle, which emotion. And then correlates each tag with ROAS and CPA against the target you set, sorting the proven from the wasteful. A few hours later, you have a graded markdown ledger of your entire creative history.

This is the part people mean when they say AI changed their ROI. The cost is almost nothing, it never gets tired, and it spots patterns across hundreds of ads that no human is holding in their head. If you do one thing from this email, build this. (I keep waiting for the moment it whiffs so I can feel useful again. Still waiting.) On thin accounts, it comes back thin, and I’ll get to that. But when there’s real spending history, it is genuinely good at this.

A couple of weeks ago, my head of growth, Janet, and I sat down to brief some statics for a protein brand. We had the dashboard open. It told us what it always tells us: the proven buyer was the gut-health persona, the winners were mostly partner ads we couldn’t easily clone, and one cohort flagged as “fitness enthusiast,” which, again, means nothing.

So we closed the dashboard and did the actual work: as we wanted to find:

NET NEW ANGLES. This is where human research comes into play*

Two people, Reddit and TikTok open, reading customer reviews out loud, riffing, no brief yet. Just looking forward.

In thirty minutes we got to a coffee occasion nobody on the account had run, off the plain observation that people already add protein to the coffee they drink every single morning.

That led to the line that made us both stop, which was Janet’s: you already have a coffee habit, so why not a protein habit. We got to an identity angle for a very specific, very online cohort that deals with gut sensitivity, pulled from how they actually talk about themselves and not from any persona label.

We got to a GLP-1 adjacency, people stacking protein into their coffee to hold muscle, which is culturally live right now and has exactly zero history in this account. And we got to overnight oats for parents whose kids are home for the summer, because summer wrecks the morning routine and that’s a moment, not a demographic. We also generated about six angles we threw out immediately, which is the part nobody writes down but is most of how you get to the good ones.

At one point Janet said it out loud: look how much research we just did in thirty minutes, and we don’t even have the brief yet. That sentence is the whole point. (She has a habit of saying the thing I’ve been circling for ten minutes in one clean sentence. I’ve made peace with it.) None of those angles existed in the ledger. The machine could not have handed us a single one, because every one was a bet on something that had never been tested.

*I am working on something here that could greatly catalyze this!

This isn’t a knock on the models, it’s how they work. They pattern-match on what exists. Point ten different tools at the same backward-looking export, and they converge on the same five concepts, because they’re all reading the same past. I put this to the test in the last issue: two AI models on one real account, and the math converged almost perfectly while the strategy split. The genuinely new idea came from stepping away from the data, not delving deeper into it.

So here’s the good news, and it’s better than it sounds. If everyone in your category is feeding AI the same proven angles, everyone in your category is about to ship the same ads. The forward bet, the angle that isn’t in anyone’s data because nobody has run it, is the last durable edge left. That’s not the part of the job AI is taking. That’s the part that’s quietly making more valuable.

The ROI math I actually believe: the machine gives you volume, speed, and coverage on everything that’s proven. You give the next bet. Stop asking the historian to do the forecasting and the whole thing gets easier.

The kit up top has one file for each half.

Persona-tagging is the backward pass: the full Meta API setup plus the prompt that pulls and tags your entire history and grades it against your ROAS target. It hands back two maps. One of what’s proven. One of the white spaces, the personas, angles, and formats you have literally never tested. The white space is the gift. It’s the machine telling you exactly where the unexplored edges are, so you can go stand on them.

Briefing-to-brief is the forward capture. You record a messy session like the one above, drop in the transcript, and it pulls every angle, hook, persona, and quote you stumbled into, dedupes them, and scales them into your week’s static count. The humans do the thinking. The machine makes sure nothing gets lost and turns five good riffs into fifteen briefs.

Map the known with the machine. Bet the unknown as a human. Then let the machine scale the bet. In that order.

I’m not going to oversell it. The map is only as good as your spend history, so a small account hands you a thin, slightly sad little ledger. The tagger will label a vague cohort (“fitness enthusiast,” my beloved nemesis) as if it were a real, actionable persona, with full confidence, and you have to be the adult who says no. The forward-capture prompt has zero taste. It organizes your ideas; it does not have them, so a thin strategy session gets you thin briefs, garbage in and tidy garbage out.

And on setup: the Meta token can stall for an hour waiting on an admin, and if you run image gen through Higgsfield instead of fal.ai you’ll need to feed the system the Higgsfield credentials and model docs. It works. It’s just a half-step, and I won’t pretend it isn’t.

Want the kit? The Meta API tagger and the briefing extractor are both paste-ready, and the full walkthrough is in the companion doc here.

Reply and tell me one angle you’ve never tested but have a gut feeling about. Not one your dashboard suggested. One you’d actually bet on. I’m collecting these, and the best ones are always the ones with no data behind them yet.

Have a great weekend,

Will