Hey all, and happy Friday,

We kept the weekend local, which is always so nice.

Maddie was out Friday, so I closed out Widow's Bay. I went in skeptical, and I ended up right where I started. A show with that many Emmys made me assume I was the problem. My verdict: it's Lost and The Office sharing one trench coat. I enjoyed it. Probably.

Our friends ran a beer mile in Williamsburg on Saturday. We could not get organized in time (a likely excuse), so we showed up for the beer half and none of the mile. Anyone who’s had an 18-month-old knows that being at an establishment that serves food and drink with a toddler makes the experience a contact sport. Thankfully this one was outdoors, which meant James could run laps instead of running the place (or a beer mile).

Then we had a playdate with the British family down the block who just moved here. Their son was in James's daycare class before graduating to a bigger school. They are the first friends we have made through James rather than on our own, which tells you plenty about our social skills (#nonewfriends, or whatever the hashtag is). Side note: are hashtags still a thing? Remember when every wedding had one? Or maybe they are still a thing, and I'm just bad at my job and no longer in the wedding rotation.

Anyway, it was hysterical. The two toddlers kept mirroring each other flawlessly. One lay down, so the other went horizontal. One hit downward dog, so the other struck the same pose.

Then Big James arrived.

Big James is the other James, who also graduated from our son's daycare, and he is why ours still answers to Baby James (Sweet Baby James, anyone?). Within 30 minutes, the three of them had ditched the playground for an empty soccer field and kept making breaks for the street, following the big kid's lead.

James was cooked after all of that, but Sunday was his first soccer practice, so he had to rally. There were no excuses. "Soccer practice" is a loose term. He wore a cone as a hat. He also placed the red bean bag on the red cone when instructed (two essentials for any real scrimmage), which is stronger instruction-following than I can honestly handle.

He kicked the ball too, so he might turn out more athletic than either of his parents. (The bar is low. I lasted one year of JV soccer before fleeing to cross country, where nobody passes you anything… That's football, for my metric readers).

We had dinner Sunday night with my stepbrother and his wife, and James was shockingly well-behaved.

The solve?

A steady stream of French fries. He put away so many that it became the night's main entertainment. I would love to file that under parenting strategy.

It was a bribe.

Claude Opus 5.5 dropped on Tuesday. Anthropic priced it 20 percent under Opus 5, made it faster, and devoted a surprising chunk of the launch page to writing. The model "follows the writing rules you give it." It is "less likely to use jargon." According to early testers, the writing is "clearer and easier to follow."

So I ran my usual routine. I put it in a room with Fable 5.1 and GPT-6 Astra, handed all three the same beer brand, and set up a blind copy test. The result is waiting at the bottom of this issue, and it is not the one I was rooting for.

The more interesting story happened while I was busy scoring headlines. Two of my briefing sessions this week ran on Opus 5.5, and both did the thing I have spent a year begging AI tools to do.

They quit handing me finished files.

And I got them to GIVE ME THE RAW FILES WITH LAYERS!

Every AI ad tool I have touched ends the same way. It hands you a PNG or an MP4, and that's the end of the road. The output looks great. Then the client asks "can we change one word," and suddenly you are back inside the model, begging it to redraw the whole ad and praying the hand holding the phone has five fingers this round.

This is why editors and designers quietly resent AI output. The quality isn't the issue. The issue is that a flat file is a dead end. Nobody on the team can adjust it without starting from zero.

(I know Canva has a solution for this… but tbh it is a bit mid).

The same client note, handled two ways. A flat file sends you back to the image model. A layered file or a timeline makes it a one-word fix.

So this week I changed the ask. I didn't want the ad. I wanted the ad in pieces, inside the software my team already lives in.

The first build started on a client call. The chief science officer at one of our supplement clients went on a popular health podcast. The episode runs 85 minutes, and the client shared two facts on a recent call: podcast-style content featuring their scientist outperforms a lecture, and this specific episode drove new customers the week it went live.

I pointed Opus 5.5 at the episode and the account's data. Here's what it did, step by step:

All of that is impressive. It is also the part every AI video tool now advertises. The part after it is the part I actually care about.

For every clip, it also assembled a Premiere package. Each element lives as its own file: the podcast window, every diagram, the headline, the speaker label, every caption exported as its own image, the end card, the voice track and the music. Sitting beside them is an XML file. You import it into Premiere, and a timeline opens with each piece already placed on its own track, exactly where it belongs. (The first version of the script tripped over one of its own variable names and crashed. The second one didn't.)

There are 30 captions in the first clip. With a flat file, a typo in caption 17 forces a full re-render. With the timeline, my editor selects caption 17 and types the fix. If the music runs too loud, the fix is a slider. If the client hates one of the diagrams, you delete a single clip.

The AI handles the first cut. An editor still makes the final one. That division of labor is the whole point.

Build number two is for a prenup service.

For these statics we made one rule: the image model never writes the ad copy. Image models excel at photographs and remain shaky on words, and the words are exactly what the client fights over. So we build the ad in deliberate layers:

The result is a Photoshop document where the photo sits on one layer, every shadow and shade sits on its own layer, and each line of copy is live type in Work Sans. This is not a picture of type. It is type you can click into and edit.

I verified this the boring way. Comparing Photoshop's own export of the file against the final ad shows a difference under 1 on a 0 to 255 color scale. The only pixels that disagree are the soft edges of the letters.

The ad survived several rounds of my notes. One earlier note was that the phone screens looked pasted on. They did, which is why step 2 exists. In the final round I asked for the phone copy centered, the ring moved to her ring finger, and a drive that read as the Hamptons rather than a parking lot. Those notes touched the photo, so Claude re-shot the photo. The copy never needed redrawing, because it was never baked into the photo in the first place. When the client wants one word swapped on a Friday afternoon, that becomes a ten-second Photoshop job instead of a regeneration.

Some honest caveats are due here, because none of this is magic.

The Premiere file doesn't include every effect. The brief dip to cream ahead of the end card, plus the 4:5 crop, still get added manually. My editor will also need to relink the media once if their drive mounts in a different spot.

In the prenup ad, the phone screen is the single place an image model handled words, since it re-photographed the screen into the scene. We checked it word by word, but it is not live type. If the client changes the app, that screen gets rebuilt from scratch.

Then there's the obvious caveat: nothing here replaces the editor or the designer. It moves their starting line. Rather than opening a blank timeline or a flat PNG, they open something that is most of the way done and completely editable.

Now back to the copy test, since the launch page leaned so hard on writing.

The first run played out the way I hoped it would. All three models got the same beer account, and each wrote 23 ad briefs from a blank page. Opus 5.5 took 13 of the 23 formats, and even Fable 5.1 rated Opus 5.5's set above its own.

Then I dug in and saw the test was rigged, by accident. Every model had picked its own persona, angle and product for each ad. Opus 5.5 wasn't beating anyone on words. It was beating them on strategy. I had built a test that proved Opus 5.5 was the better strategist, then congratulated it on its writing.

So I reran everything with the variables locked down. One persona, one angle, one product, one visual, identical text slots. The models wrote words and nothing else. Four blind judges then chose the strongest set of words for each of the 23 ads.

Fable 5.1 collected 45 of the 92 picks. Opus 5.5 landed 32. Astra took 15. Opus 5.5 also served as a judge on its own test, and it preferred Fable 5.1's words to its own more often than not. I was rooting for the new model, and the new model voted for the other guy.

Reading them side by side made the pattern obvious. When the reader needs fast orientation, Opus 5.5 takes it: "Not into bitter? Start here." When one small human detail has to carry the line, Fable 5.1 takes it: "Top shelf is for the guests skipping alcohol. Everyone else, I counted."

That settled our current setup. Opus 5.5 constructs the ad and decides who it targets. Fable 5.1 handles the words. And a human still calls which version ships.

The kit is both builders, rewritten so they work on anyone's files, not just ours.

The Premiere builder takes a short list of your clips, images and audio, with the second each one starts and ends, and writes the XML. Import it and the timeline opens with every piece on its own track. It comes with a demo you can run in a minute to see it work before you touch real footage.

The Photoshop builder takes an ad built as a web page, splits it into layers, and has Photoshop save a layered file where every line of copy is live type in your font. It includes a template ad to start from.

There is a README written for people who have never opened a terminal, and a prompts file. Open the kit folder in Claude Code, paste the prompt, and it runs the whole thing for you.

Opus 5.5 launched Tuesday. Judged on words alone, Fable 5.1 remains the better ad copywriter, taking 45 picks to 32 in a blind test. Opus 5.5 is stronger at figuring out who an ad should target. The most useful thing I did with it this week, though, was ask for editable files for my team instead of finished ones. One podcast episode turned into four clips plus four Premiere timelines with each caption on its own track. One static ad arrived as a Photoshop file with live type set in the brand's font. The AI writes the first draft. The editor and designer still own the final pass, and for once they actually can.

Have a great weekend,

Will

P.S. If you adapt the kit to run in Final Cut or DaVinci instead of Premiere, hit reply and tell me about it. I will steal your version and give you credit.