Hi, it's Peggy Burnett.
I asked four Claude models to make eight sales claims more specific, and gave them the full fact sheet to work from.
Between them they invented 10 details, and 8 of the 10 landed on the three claims the fact sheet barely supported.
So I tested the other side as well. Eight blind readers, all of them Claude sessions, saw every version of five claims and picked the one they believed. They chose the plain number 34 times out of 40. I had bet on the named example, and it won three times.
When four Claude models, working as writers, disagreed with those readers, they picked a more vivid version every single time.
This issue is about where a claim should stop climbing, and the prompt that takes it as high as your facts allow and no higher. Here is what the test showed.
Every Claim Has a Rung, and the Top One Is Rarely Right
In May I wrote about Caples's rule that "$50,000" beats "a fortune." The rule holds, and it leaves a question open. More specific than what, and how far up do you go?
"A fortune" to "$50,000" is one step. A claim can be made specific in at least four different ways, and each way is a rung.
I built a test product to climb them on: the Marlow No. 10, a $165 cast-iron skillet from a foundry that doesn't exist. The product is invented, and so is everyone named in this issue. Inside the test, every fact on its fact sheet was treated as true, which let me control exactly what a writer could and couldn't support.
Then I ran three tests, all on Claude models in clean sessions with nothing else loaded.
A reader test. Five claims, each written at every rung the facts supported, plus a decoy. Eight readers picked the most and least believable version of each claim.
A writer test. Four models recommended the rung each claim should sit on for a skeptical buyer.
A fabrication test. Eight claims and the full fact sheet, sent to four models with a plain "make each of these claims more specific," and separately with the ladder prompt printed below.
Five Rungs, One Claim, From "Holds Its Heat" to the Hiss
Here is one claim climbed all the way up, using only facts from the sheet.
Rung | What it adds | The heat claim at that rung |
|---|---|---|
1. Category | The claim as a general property | The Marlow No. 10 holds its heat. |
2. Instance | One concrete situation where it shows up | The Marlow No. 10 stays hot when a cold steak hits the pan. |
3. Number | A measured quantity | When a cold 12-ounce ribeye hits the pan, the surface drops about 40 degrees and is back to temperature in roughly 90 seconds. |
4. Named instance | A named person, place or case a reader could check | Dana Ruiz, chef-owner of Larkspur in Columbus, runs six of them on her line because the sear "doesn't quit when the steak goes down." |
5. Sensory | What the reader would see, hear or feel | Drop a cold steak into the Marlow No. 10 and the pan keeps hissing. |
Two things about the ladder are worth saying before the results.
Not every claim has every rung. "Customers love it" can climb to a survey and a named customer, and it has nothing sensory to offer. "It heats evenly" had no support above the category on our fact sheet at all. A missing rung is information about your brief.
And the rungs are not a ranking. Higher means more vivid, which is a different thing from more believable. That difference is what the reader test measured.
Readers Believed the Number, as Long as It Looked Measured
Each reader saw the five claims with the versions shuffled, and was told to read as a buyer who had been burned by marketing before. Each group also held a decoy, a version carrying invented precision or an unsourced comparison, such as "loses exactly 41.3 degrees" or "94.6% of owners say it's the best pan they've ever used."
The decoy was there because of Nova's issue last week, which showed how readily a model will believe whatever a question hands it. If my readers believed the decoy, the test was measuring nothing.
They didn't. Here is the full count across 40 judgements.
Category | Instance | Number | Named | Sensory | Decoy | |
|---|---|---|---|---|---|---|
Most believable | 0 | 0 | 34 | 3 | 3 | 0 |
Least believable | 1 | 0 | 0 | 0 | 0 | 39 |
Flagged as made up | 0 | 0 | 0 | 2 | 0 | 40 |
The number rung won every one of the five claims, with between six and eight readers out of eight picking it.
The part I find most useful is that the winner and the decoy were both numbers. "About 40 degrees" won and "exactly 41.3 degrees" was flagged as invented by every reader. "168 of 212 buyers" won and "94.6% of owners" lost in all eight readings.
The readers' own reasons sort into four rules, and they are the most tactical thing in this issue.
Round a rough measurement, and say so. "About" and "roughly" read as an honest kitchen test. A decimal place on a thermometer reading reads as a brochure.
Give the raw count and the sample size. "168 of 212" shows the method. "94.6%" hides it, and it is too flattering to be believed.
Describe the process when you can't describe the result. "Three coats of flaxseed oil, each baked on at 500 degrees" beat every outcome claim about seasoning, because nobody inflates a production step.
Never compare against an undefined average. "31.7% lighter than the average skillet" invites the question of whose average, and every reader flagged it as made up.
The named rung did worse than I expected in a second way, too. The only two non-decoy versions flagged as made up were both named quotes, one from the chef and one from a customer. A name is checkable in principle, and a reader with no way to check it treats it as a testimonial, which they have learned to discount.
The Writer in Every Model Reached One Rung Too High
The ladder prompt ends by asking the writer to recommend the rung each claim should sit on for a skeptical buyer. Here is how the four models' recommendations compared with what the readers believed.
Claim | Readers believed | Haiku | Sonnet | Opus | Fable |
|---|---|---|---|---|---|
Heat | Number | Number | Number | Number | Number |
Customers | Number | Number | Number | Number | Number |
Handling | Number | Named | Named | Named | Named |
Durability | Number | Named | Named | Named | Named |
Seasoning | Number | Number | Sensory | Sensory | Sensory |
Eleven of the twenty recommendations matched the readers. All nine that didn't went higher.
The writers' reasons were good writing advice. A 71-year-old customer saying it's "the first cast iron I can lift with my wrist" is "more persuasive than a weight figure, because buyers can't feel 5.2 lb through a screen." A sliding egg is something "a buyer can picture."
As readers, the same four models picked "The Marlow No. 10 weighs 5.2 pounds" seven times out of eight, because "a plain weight is easy to verify and has nothing to inflate."
I'd have made the writers' call on handling myself, which is why I think this finding is worth your attention.
Writing pulls you toward the version that is vivid and memorable. Reading as a skeptic pulls you toward the version that is checkable and plain.
The test can tell you where belief peaks. It can't tell you whether memorability matters more for a given line, and that judgement stays yours.
Asked for "More Specific," Claude Padded the Claims It Couldn't Support
In August, Nova ran the self-rating trick on a thin offer and watched "make it more specific" invent a dietitian, a testimonial and an 87% statistic. She ended by asking whether a fact-rich brief would tame that.
This test answers half of her question. Every model got the full fact sheet, and the plain "make each of these claims more specific" invented no statistics and no people. It used the real ones.
It still invented 10 details across the four models, and 8 of the 10 landed on the three claims the fact sheet barely supported.
Claim | What the fact sheet said | What the plain request added |
|---|---|---|
Easy to clean | Rinse with hot water, dry on the burner, wipe with oil | "No scrubbing, no soap" |
Heats evenly | Heats evenly | "Without hot spots," "works across all cooking methods" |
Versatile | Searing, cornbread, pizza | "Searing, baking, and roasting equally well" |
That is the shape of the problem with a rich brief. Where the facts run out, the model keeps going at the same level of confidence, so the padding sits next to real detail and reads exactly like it.
One answer did something subtler. It took a real fact from the heat test, that the pan keeps hissing after the steak goes in, and offered it as evidence that the pan "heats evenly," restated as "hissing steadily across the surface." Every detail in that sentence traces to the sheet, and the claim is still unsupported. A fact used for the wrong claim passes a line-by-line fact check.
The ladder prompt, on the same claims and the same fact sheet, invented 2 details across the four models. Both came from one model on the 1958 pan, one of them an age worked out from the cast date. It also did something the plain request never does, which is tell me what to go and get.
Climb Each Claim With a Prompt That Cites Its Source
Here is the prompt exactly as tested.
Below are claims from a sales page and the fact sheet for the offer. Take each claim up the specificity ladder using only what the fact sheet says.
The rungs:
1. Category: the claim as a general property ("holds heat").
2. Instance: one concrete situation where the property shows up ("stays hot when a cold steak hits the pan").
3. Number: a measured quantity taken from the fact sheet.
4. Named instance: a named person, place or case from the fact sheet that a reader could in principle check.
5. Sensory: what the reader would see, hear or feel, used only where the fact sheet describes it.
For each claim:
- Write one sentence at every rung the fact sheet supports, in rung order.
- For any rung the fact sheet cannot support, write "NEEDS:" and name the exact fact you would need. Do not fill the rung.
- Never invent a number, name, place, date, quote, comparison or result. If it is not on the fact sheet, it does not go in the copy.
- Then recommend the rung this claim should sit on in the final copy, with one line on why. Higher is not automatically better. Pick the rung a skeptical buyer is most likely to believe.
Return each claim as a short table with three columns: rung, sentence (or NEEDS), and the fact-sheet line it came from. End with a list of every NEEDS item, so I know what to go and get.
CLAIMS:
[paste the claims, one per line]
FACT SHEET:
[paste every fact you have about the offer, including numbers, names, quotes and what people notice when they use it]Four parts of it carry the weight.
The rungs are defined with an example. "Make it more specific" leaves the model to choose the kind of specificity, and it chooses vivid. Naming five kinds turns one vague instruction into five checkable ones.
An empty rung gets a NEEDS line. This is the instruction that stops the padding. The model has somewhere to put the gap other than the copy. On "heats evenly," Opus wrote "NEEDS: a measured temperature variation across the cooking surface," which is exactly the test the client should run.
Every sentence cites its fact-sheet line. This is what catches a real fact used for the wrong claim. When the source column for "heats evenly" points at the heat-recovery test, you can see the mismatch at a glance.
The shopping list comes last. Opus listed thirteen missing facts. Hand that list to the client before you write, and the next draft can climb further honestly.
The fact sheet field is the whole input. If you paste in only the brief, the prompt can only climb as high as the brief goes, and it will tell you so in the NEEDS list. That is the point of it.
Read the Recommendation Column One Rung Lower
The recommendation step is the weakest part of the prompt, and the writer test showed why. In nine cases out of twenty, it pointed higher than the readers believed.
So I treat the recommendation as a starting point and apply one rule to it. When the prompt recommends the named or sensory rung, check whether that claim has a number. If it does, lead with the number and put the name or the picture directly behind it.
Here is how that sorts the rungs into jobs in the copy. The first row is what the test measured. The other rows are my judgement, because the test measured belief and nothing else.
Rung | The job it does best | Where it goes wrong |
|---|---|---|
Number | Carries belief in the claim line itself | False precision, and comparisons to an undefined average |
Named instance | Supports the number as proof, in a testimonial or case block | Reads as an unverifiable testimonial when it stands alone |
Sensory | Lets the reader picture using it, which builds desire | Asked to carry belief, it reads as marketing |
Instance | Holds a claim that has nothing measured behind it | Fine on its own, and better than a vague category |
Category | Only for a claim you can't support yet | Usually a sign the claim needs a fact, or needs cutting |
For the handling claim, that produces "At 5.2 pounds, it's light enough to lift with one hand to pour off the fat," followed by Margaret Okafor's line in the reviews block. The number does the believing and the name does the persuading.
Where This Test Is Thin, and What It Can't Tell You About Your Buyers
One invented product and eight claims. A different category, especially an intangible one like software or coaching, may have a different shape of ladder.
One run per model per test. The patterns were consistent enough to report, and the counts are directional.
Every reader was Claude. The decoy check shows the readers punish false precision. It doesn't show that human buyers rank the rungs the same way, and I would want a live split test before treating the number rung's lead as settled.
Belief is one job among several. A named customer or a vivid detail may lose on belief and still win on memorability or desire, which this test did not measure.
The scoring was strict. The coder counted "open campfire" against a fact sheet that said "campfire." I've reported its counts as they came.
The Burnett Matrix
The number rung won on belief because it is the rung a reader can check without trusting you.
Climb as high as your facts allow, then stop one rung lower than your writing instinct wants, and put the name and the picture behind the number rather than in front of it.
More opens, clicks, and conversions,
— Peggy Burnett
