When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning (constraints, edge cases, examples, tone) and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

Used by teams at Microsoft, Amazon, and Notion. Try Wispr Flow free on Mac, Windows, iPhone, and Android.

Start flowing free

Quick note on this one, because it lines up with the whole issue. The fix for Claude's copy below wasn't a smarter prompt. It was me explaining, in a full sentence, why each line was wrong. Most of those notes get dictated into Wispr Flow, because typed, they shrink to "make it better." House rule stands: I only run ads for tools I use every day, and I use this one all day.

This week covers the side of AI creative that never makes it into anyone's screenshots: the lines Claude botches, and the system that turns each of my corrections into a rule it reads before it writes again. The starter rulebook, the Claude-speak ban list and both skills are here:

Hey all, and happy Friday,

Last week we saw You Can See Everything, which was amazing. Obviously it broke the internet. No spoilers. Just go.

Friday night was the first book club, hopefully of many. We read Fleishman Is in Trouble. I organized it, which I'm pretty proud of. It's a bunch of my old people from the finance world, all doing an amalgam of different things now, and people I probably wouldn't see on a weekly basis otherwise. It's a really, really nice way to catch up.

Saturday was Maddie's 15-year high school reunion, and James came with us. Their mascot is a hornet, which entailed a full-size human in a hornet costume. James is fearless about most things. Not this. From a distance, he'd look at it and go "bee, bee!" When it came up to him, he lost it. So we retreated to the pumpkin bread, which was excellent, and James and I ate a ton.

Maddie showed me her old stomping grounds, including a seriously crazy athletic facility with an open gym for the kids: hula hoops, scooters, the works. James loved it. This kid is an athlete, if that wasn't abundantly clear from every previous newsletter. During the football game, he found a rogue football within minutes and started playing with it. I don't see a football career. Maybe flag football. Maybe baseball, although that's a spring sport. Fine, we're sticking with soccer.

Saturday night we went to Maddie's friend's house in Westchester, and the kid count got out of hand: James, the twins from his soccer practice, the two older boys who have appeared in this newsletter before, their little sister, and the host’s sister's two kids, with a third on the way. Ages zero to idk. A few watched Despicable Me, a few ran in circles, and a few built block towers purely to knock them down. Toys were flying. Then a big taco dinner with way too much corn elote. We've hung out with this exact group many times, and it was hilarious how different it looks now.

Before we left, Maddie's friend broke out Longford's ice cream. If you've never had it, get the espresso. I am still thinking about it. I genuinely could eat an entire pint. (The blue one was good too, just not a full pint good.) I'm a Ben and Jerry's guy through and through, and while I occasionally had Longford's growing up, I somehow forgot how good it is. It's fluffy, and the quality is excellent. Longford's, come to New York City. Until then, that's the perk of a Westchester visit.

Sunday we were wiped, so it was a chill day, apart from soccer practice at 9 am sharp. It was his first practice without rain. He started slow, and by the third drill (shooting on goal) it was "score, score, score, score." Drop a ball in front of this kid, and it ends up in the back of the net. At home he either kicks it or picks it up and line-drives it at you. If you've ever met me, or at this point have a parasocial relationship with me, you know I'm not a "my kid will play sports" dad. I don't really give a shit what he does, as long as it makes him happy (look how well-rounded I am!).

And Linus. Today, the day you're reading this, is his 6th birthday, and we still need to get him something. I took him to the park yesterday morning, and he spent the whole time ripping around with a Great Dane we had never met, having the time of his life. He's 55 to 60 pounds, solidly medium, and he only wants to play with the massive dogs. I don't know if it's a short king thing. Either way, beautiful fall morning in Prospect Park, though.

Last Thursday Claude delivered ten Black Friday headlines for a toothbrush brand. One slot was aimed at people with arthritis, and the line it wrote there was: "Every tooth at once, with the hand that still works."

I read it three times. The rhythm is there. It sounds like an ad. The meaning is nowhere, which is awkward, because I wrote the brief it came from.

My whole reply was one line: "this reads like Claude speak."

Once you know the sound, you hear it everywhere. Two clipped fragments with the verb gone. A pretty euphemism where a plain word belongs. A joke only the brief would get. It sounds exactly like copy and means close to nothing.

That batch also contained "The ADHD planner lasted a week. The habit app, three days." That one was also for the toothbrush. My reply was "I have NO idea what that means," caps included, which is how you can tell I was serious.

I wanted to know if it was just me, so I built a test. Three fresh AI readers each got 10 shuffled headlines from that batch (the 5 lines Claude drafted first plus the 5 that replaced them) with no image, no brand and no brief. Their job was to name what each line was selling and to whom, then score it from 1 to 5.

The first drafts averaged 2.7. The replacements averaged 3.6, and each one outscored the line it swapped out. All three readers said the ADHD line might be selling a planner or a habit app, and not one of them read "the hand that still works" as arthritis. All three said stroke.

The replacements aren't clever. One reader pointed out that "Why two minutes of brushing is so hard with ADHD" sounds like a blog title, and they're right. A blog title still wins against a line nobody can parse.

The real lesson is who does the grading. Our pipeline already had a cold reader (another model), and it waved this batch through. The model that wrote a line remembers what it meant, so it never confuses itself. The cold reader now restates every ad in a single sentence and answers yes or no: would an actual person say this out loud to a friend?

The second problem was sneakier. Every slot comes with a reference ad so Claude knows the layout: the line count, the placement, the length of each one. Claude took the words too.

A leather brand's "For The Man Who Has Everything" turned into "For The Dad Who Has Everything." A makeup brand's "7 Pieces. 1 Unbelievable Price." turned into "4 Pieces. 1 Black Friday Price." The sentence stayed and only the nouns moved (and that first one is the most tired gift cliché in existence, which Claude would know if it had ever been forced to buy a gift).

The rule is now written down: a reference dictates the format and never the wording. Each line also gets an echo test against its reference, and a matching sentence frame, or 3 identical words in a row, bounces it back.

This one is on me. For a PC gaming app, my idea was a gamer's arm in a cast with "I can only game for so long" scrawled across it. It got the best click rate in the batch.

Then the clicks hit the signup page, and almost none of them signed up. The ad with a baby monitor on the desk and the app's setting on screen got far fewer clicks, and signed people up for a fraction of the cost.

People were clicking the cast because they couldn't tell what it was. My note on it: "it has a crazy good CTR bc no one has any idea what it is."

That one went into the rules as well. The image has to explain the product, and the ad gets graded on cost per signup, not on click rate.

None of this lives in my head anymore. When I flag a bad line, Claude records a rule in the briefing skill the whole team runs on: my exact words, the date, the client, the offending line and the fix. It backs the file up before each edit, and the next batch for any client reads the whole thing before it writes anything.

There are 87 dated rules in the skill since July 13, plus 184 more feedback notes sitting in Claude's memory. Claude still writes Claude-speak on the first pass. (I am, functionally, a very expensive spell check for vibes.) What changed is that a mistake caught on Thursday is a rule by Friday, across every client at once.

Does it work? On the gaming app, the first ads we launched are signing people up for less than the rest of the account, and for less than the client's own new ads. It's early and the spend is small, so I'm saying it quietly.

You don't need our setup for this. You need a single file Claude reads on every run, plus the discipline to explain why a line failed instead of silently rewriting it yourself.

Have a great weekend,

Will

P.S. If you build a rulebook, hit reply with the single best rule inside it. I'll feature the good ones (with your permission) in a future issue.