Hi, it's Peggy Burnett.

The usual complaint about Claude's copy is that every sentence comes out the same length, so the page reads flat. I measured it against three of the best-known ads of the 1920s, with a script that counted every sentence the same way.

Claude's sentence lengths varied as much as Caples's did. Its average variation score was 0.59, and the three 1920s ads averaged 0.58.

The difference sat in two other places: where the long sentences go, and what the short ones travel with. "Vary your sentence length," the instruction most of us reach for, made both of them worse.

Then I asked Claude to judge the edited drafts blind, and it preferred the choppy originals 35 times out of 44. Here is what the counting showed, and why I would keep Claude out of the judging.

The claim I set out to test comes from our own planning notes, which said models "produce sentences of remarkably even length." It is a claim you can check by counting, so I counted. Mark's Sugarman issue in October worked on whether each sentence earns the next one, and this one measures how long they run.

For the human side I wanted long copy that was proven in the mail and free to reprint in full, which pointed to the 1920s. US works published before 1931 are in the public domain, so I could measure these and quote them without asking anyone.

The Caples ad is the one I quoted in May as the most cited example of his specificity rule. I cleaned the scanned text by hand and kept the body copy only, with no headline, subheads, offer panel or coupon.

Then I wrote a brief for each ad from its own facts. Each one carried the product, price, proof, offer and reader, the creative concept in a paragraph (the party and the piano, the list of common mistakes, the young woman and the French menu), and the length of the original. Four Claude models (Haiku, Sonnet, Opus and Fable) wrote each ad twice, each in a clean session with nothing else loaded.

A script split every text into sentences with the same rules and counted the words in each. I tracked four numbers.

One complication turned up early. About a third of one Fable draft on the English brief was Sackheim's original, recalled word for word, so I dropped both of its plain drafts on that brief. That left 22 drafts from the plain brief.

I expected the variation score to be where Claude fell short. The data put it level with the human ads.

On the Piano brief Claude matched Caples on every measure, including the share of long sentences, and on the Etiquette brief it matched Eichler. The variation score never separated Claude from the 1920s writers. The one clear gap in the table sits in Sackheim's row.

Sackheim wrote one sentence in five at 25 words or more. Claude's six drafts on his brief averaged one in twenty.

This is the kind of sentence that went missing.

Four parallel clauses build to a single payoff, and the reader takes them in one breath.

Opus covered the same ground in a run of separate sentences: "Every time you talk and every time you write a letter, you show people what kind of person you are.

They judge you by your English. They judge your education, your ability and your background."

Splitting each ad into its first and second half showed where they belong.

Across the three 1920s ads, long sentences rose from 10% of sentences in the first half to 16% in the second, where the story ends and the selling starts.

Caples is the clearest case, at 2% in the party scene and 12% in the pitch, while Sackheim, who had no story to tell, held at one in five throughout. Claude's Piano drafts held at 7% in both halves.

The human writers saved their long sentences for the pitch. A reader who has already bought the story can take a sentence that stacks three reasons and a request, and that is where Caples and Eichler put theirs. Claude wrote its pitch at the same pace as its story.

It is the same part of the page Mark examined last week, measured from a different angle.

Short sentences themselves turned out to be fine. Caples used slightly more of them than Claude did on the same brief, 24% of his sentences against 22% of Claude's. The difference is what sits next to them.

In the 1920s ads, a short sentence nearly always stands between longer ones. "The crowd laughed." comes after "'Jack is up to his old tricks,' somebody chuckled" and before "They were all certain that I couldn't play a single note." Claude tends to put its short sentences side by side.

Fifteen of the 22 Claude drafts came in above the highest human figure. The pairs have a recognisable shape.

The first two set something up and knock it down, the third is two fragments, and the fourth is the same sentence twice. My read, which the test can't prove, is that this pairing is the sound people mean when they call Claude's copy flat.

Next I gave the same models the same briefs with one line added, "Vary your sentence length," which is the instruction most of us type when a draft sounds monotonous.

Claude heard "vary" as "add short sentences." I compared each model's first plain draft with its varied draft on the same brief, and the average sentence got shorter in all ten pairs. The share of long sentences fell by half overall, and all eleven varied drafts came in above the highest human figure for short pairs.

Fable's varied version of the English ad gives the flavour, with "Those are not promises. They are records."

Here is the editing pass I wrote before I had seen any of the data, printed as tested with a paste field added. I ran it on one plain draft from each model on each brief, eleven in all, with the same model editing. Swap "magazine ad" for whatever you are editing.

Four parts of it carry the weight.

Here is what the pass did to the same eleven drafts.

It moved every measure in the table toward the 1920s ads. It did least on Sackheim's brief, where the share of long sentences rose from about 4% to about 9%, still well short of his 20%.

After the counting, I wrote a second pass aimed squarely at the finding. It told Claude to merge related points into long sentences and to put at least one sentence of 25 words or more in every section that argues or explains.

It overshot. The average sentence went from 12.4 words to 20.5, past Sackheim's 17.3, and one sentence in three ran to 25 words or more. The longest ran to 108 words. It opened with "If you have ever been in that chair, you will recognize the feeling, and perhaps it was not a menu at all but a dinner party where you sat for ten minutes wondering which of three forks was yours," and it kept going for another seventy words.

A floor in the prompt became a quota in the copy. Claude met the minimum by merging nearly everything it could, and Opus turned the list of common mistakes into one 65-word sentence held together by three semicolons. I've left that prompt out of this issue, because the first pass came closer to the human numbers and the lesson from the second fits in one line: name the move and leave the number out.

Last, I wanted to know whether the edited drafts read better. Sonnet and Opus each read every plain draft beside its edited version, blind and in both orders, and said which read more like the work of an experienced human copywriter and which a reader would be more likely to finish.

For the first pass, they picked the edited version 9 times out of 44 on both questions. For the overcorrected pass, they picked the edit 3 times as more human and once as more likely to be finished. Each pair was shown in both orders, so position can't explain it.

Their reasons describe exactly the sound the counting flagged. Sonnet preferred one of Opus's plain drafts because it "breaks up long sentences into short, punchy ones ('They judge you by your English. They judge your education...')," which it called "the varied, speakable rhythm of a seasoned copywriter." In the nine cases where a reader picked the edit, the reason ran the other way, against a draft that "keeps breaking its sentences into short, staccato fragments."

There are two ways to read that, and this test can't separate them. Claude may be right that today's readers want short, punchy copy, which would make the 1920s ads an old yardstick. Or Claude's ear may simply match Claude's habits, in which case asking it to grade rhythm will keep handing back the rhythm you were trying to fix.

Either way, keep Claude out of the judging. Use the count to find the pairs and the short-winded pitch, then read the passage aloud and make the call yourself.

Claude's sentence lengths already vary as much as a 1920s classic. What its copy lacks is the long, building sentence in the part of the page that sells, and what it adds is the short pair. Ask it to combine before you ask it to vary, and do the listening yourself.

More opens, clicks, and conversions,
— Peggy Burnett