Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

This week's sponsor is Wispr Flow (shocker). I talk into Flow all day, and it's how most of my notes to Claude get written. They also make a notetaker for calls that's built to get the hard words right, like product names and slang. That's useful for this issue, because a lot of our best ad lines come from things customers actually said on calls, and you can only use them if the transcript got them right. (I only run ads for stuff I actually use.)

Last week was the rulebook Claude reads before it writes. This week is what goes in it: the 10% a person still has to fix in the words, the reference ad and the picture. The Claudisms field guide, the checker skill and the picture checklist are here:

Hey all, and happy Friday,

Last Friday Maddie had plans, so I had a THC drink, ordered Shake Shack and watched Vision Quest on Disney+. It's actually very good. I've written off Marvel for a lot of things, and I'm guessing a lot of you have too, but I'd recommend this one. My bevvy was a pretty low dose, and I imagine it can get really trippy at a higher one.

Saturday morning was a Halloween show with Mr. Simon, a musician who plays at James's school every week. He's going for the Australian Raffi, and you can find him on Spotify. We tried to put James in his Elmo costume. He was not having it, so we brought the Elmo hat as a backup. We ran into two of his friends from school there, including his best friend, and they played the whole time. They kept walking up to the poster of Mr. Simon and pointing at it, with no idea he was playing right next to them.

After that we got pizza with James's friends and their parents. We drove, because we had Linus with us. It was his birthday weekend, and we wanted to give him a great day. James fell asleep in the car for five minutes on the way home and then refused to nap for the rest of the day, which was fun.

Saturday night we celebrated Linus with a squeaky toy, a big bone and every scrap left over from dinner. He was very happy. James even kepy saying "Li Li birthday," which was very cute.

Sunday morning was soccer practice, of course. Maddie and I suspect the coach was either hungover or had been hitting an AM pen. The word of the day is usually "go." This week it was "stop." The big game was putting beanbags in cones to make ice cream cones. It all felt a little mid effort.

After that we went to the farmers market for Sunday dinner stuff, grabbed bagels and did a long playground session. In the afternoon Maddie's aunt came over and we went to the Prospect Park Fall Festival, which was honestly a bit of a bummer. The haunted house was legit: they build it on the ice skating rink, with real actors. Everything else was mid. You could buy a pumpkin to paint, but there was no paint. James got obsessed with a pumpkin that weighed maybe 50 pounds anyway, so I carried it all the way home while he tried to help hold it. It was a whole to-do.

We also bought a second Elmo costume, plus T-shirts for Maddie and me so we can join in on Halloween and not be bummers. We'll see if this costume works. It came with a little stuffed baby Elmo, and James hasn't put it down. He's been saying "baby Elmo, baby Elmo" nonstop. He already has a big Elmo, the same one I had as a kid, but baby Elmo is the new thing. More on Halloween next week.

On Monday a strategist on our team briefed a jewelry brand, and I fed the brief into a skill that takes a brief all the way to finished statics. I made zero edits along the way. It produced 18 ads.

Our art director looked them over and said they were 90% there. I wrote that down, because it's the most honest review of AI ad work I've heard: the skill gets you 90% of the way, and taste gets you the last 10%.

Last week I showed you how every correction becomes a rule Claude reads first. This week I went back through those corrections to see what they had in common. (One of the worst lines turned out to be mine. We'll get there.)

I asked Claude to pull every line it drafted for us over the past couple of months and put each one next to the version that shipped. That came to about 110 lines, from about 70 ads across 12 clients.

The persona made it through in most of them. The angle did too. The data read, the person we picked and the problem we went after held up through review and client notes.

What changed was the words, the reference ad or the picture. Usually two of the three.

When we went through them, I said: "AI is good at putting us in the direction of what we should be focused on, what we should be testing, but it's not good at finding good references. It's not good at writing copy. That's where taste and opinion come in."

So the 10% is the words, the reference ad and the picture. Here's each one.

A Claudism is a line that sounds like copy and that no person would ever say out loud. I showed you a few last week. With 110 pairs to look at, they sort into 12 tells in three families: too tidy, too clever, and it explains.

Too tidy covers the shape of the line. The flip shows up most often: "Snacking isn't a willpower problem. It's a plate problem." The line manufactures a belief nobody held just to tear it down. The shipped version read "Protein bar at 3. Starving by 4."

Next is the payoff tag, a last line that explains the joke. One gaming ad ended with "Works even in games with no easy mode." The version that ran ends with "We won't tell." That answers the real worry, which is feeling a little embarrassed about turning the difficulty down.

Too clever is about word choice. That means puns you have to decode ("Shift's not over, knees already are?" got the review note "nobody says knees are over"), and labels we made up internally that ended up on the art. "Fourth Playthrough? You're in the Boring Zone" became "BEAT YOUR FAVORITE GAME 4 TIMES? MAKE IT HARDER."

The last family covers the content, and we call it "it explains." Claude keeps reaching for numbers the reader would have to do math on. "Died 16 Times To The Same Boss" was rewritten as "Died And Lost All Your Souls Again?", and only people who play that game know that exact pain.

Claude also writes customers who talk like the brand's marketing team. A polished three-step how-to for motion sickness lost to one sentence from a real player: "I went from only being able to play for 20/30min periods to 2/3 hour periods." We kept the typos.

This one's on me. Not every bad line in the pile came from Claude. For a toothbrush ad aimed at people caring for an aging parent, I pasted in a line myself: "Two minutes of a brush going in and out is a lot to ask of a tired parent." The client's note was "the 'going in and out' phrasing is really odd though and i dont really understand what thats supposed to mean?" Fair. It shipped as "Two minutes is a long time to hold a toothbrush in your mom's mouth."

My hand edits on a gaming batch were worse. One had a typo and ended with "You'll be eating that mf now." Someone on the team quietly turned it into "Now you're the scary one." They were right.

I didn't expect this one. When Claude digs through an ad library for a reference, it matches against the copy inside each ad rather than the layout. The results say something similar to our concept while looking nothing like what we plan to build.

We audited the reference notes on one account, and Claude had the format wrong on 6 of 19. Every static in that batch shipped with a different reference than the one Claude picked.

There was a pattern in the swaps. Claude defaults to a product photo topped with a headline, while our people grab formats the audience sees in their feed every single day: a missing poster zip-tied to a fence, a meme they already send each other, a plain text block written by an actual mom.

The last part of the 10% is the picture, and it's what clients notice first. A plain background swap sails through approval with zero comments. A crowded AI scene with several people in it keeps getting flagged.

Our art director keeps a running list of giveaways: a highlight sitting on a surface with no light source pointed at it, one bathroom shelf holding props from three different rooms, five text treatments stacked with random spacing. During one rebuild, four of fourteen product shots came in 25 to 45% oversized, a truck tailgate measured roughly 60% of its actual height, and the staircase narrowed toward the back like something out of a dream. I posted five basics for the team later that week.

Taste isn't the same as polish, either. A fuzzy AI creature video we made for a client got great clicks and comments like "I was going to buy this but won't now." So their media buyer asked for a dragon drawn in MS Paint. I made it the same day (I graduated 4th grade computer class, what can I say), and it's now carrying most of their reach campaign at a lower cost than their last brand push.

You can skip our whole setup. Our hand rewrites boil down to six moves, and they apply to any prompt or any review pass.

Have a great weekend,

Will

P.S. Point the checker at the last thing AI wrote for you and reply with the ugliest flag it finds. The best ones go in next week's issue (with your permission).