The same brief, opened two ways. One pricing-page brief, Opus 5.5 both times, 30 October. Marks on the left are the habits W10V4 counts. Plain Opus 5.5 No system prompt, no pass, first 255 words W10V4 58 Before a buyer gives you thirty minutes, they want to know one thing : whether you are in their range. Not the exact figure, and not a signed quote. They want to know whether you cost roughly what a software subscription costs, roughly what a junior hire costs, or roughly what an office lease costs. Each of those is a different conversation with a different person in their company, and they can't tell which one they are about to start until they see a number. A "contact us for pricing" page answers that question with silence . It tells the buyer that a price exists, that you know it, and that you would rather not say . Then it asks them to book a call to find out . If you run a small B2B company with custom or usage-based pricing, you know how that call goes, because you have it most weeks. There are introductions, a loose description of the problem, and then, often before the demo, someone asks : "Roughly how much are we talking?" You give a range. Sometimes it lands. Sometimes it doesn't , and both of you spend twenty more minutes being polite about a deal that ended when you said the number. That question gets asked either way . Your only choice is whether the buyer gets the answer on your website, in their own time, or on a call you both scheduled to deliver it. This post argues for the website, and shows how to write a pricing page when every deal really is different. a kind for a case a thing with a mind you short beats Opus 5.5 with W10 self-review Our prompt, two passes, first 293 words W10V4 92 "Roughly how much?" sits under every demo request ever filled in. Our bet is that the buyer wanted that answer before typing a work email into the form, and the "contact us for pricing" page showed a calendar link instead. It is a strange trade to offer a stranger. We asked them to spend half an hour with us to learn a number we already knew, or at least knew the shape of. Selling has a long history of hiding the number, and buyers never liked it. When Sears, Roebuck mailed its catalog to farm towns at the end of the nineteenth century, the prices were printed next to the goods for people who would never meet a salesman, and the catalog became one of the most successful pieces of selling in American history. Car buying went the other way for most of the twentieth century. The sticker on the window was a number everyone agreed to ignore, and the buyer haggled with a dealer who knew far more about the margin. When General Motors launched Saturn around 1990, the brand leaned hard on a no-haggle price, and CarMax later built a used-car business on the same promise. Nobody claimed those cars were cheaper because the price was printed. The price stopped being a contest, and that turned out to be worth something on its own. B2B software and services kept the dealership model and gave it nicer fonts. The "contact us" button is a sticker price with the sticker peeled off. For a small company that sells on calls, it also means the founder answers the same question on most first calls, out loud, one buyer at a time, a little differently each time depending on mood and how the demo went. The left draft reads well, and the scorer still names eight habits in it. A kind of person stands for every case ("a buyer gives you"), a page answers, tells and asks, "you" comes 15 times in 255 words, and short beats pile up. The right one started at 40, and two passes of the model rewriting its own flagged sentences took it to 92. Our human reviewer, reading blind, put the right one ahead of the left one, by a few points.
Prompting the problem away
Our first fix for drafts that smelled of a machine was a longer prompt. We wrote twenty rules into the Writer's instructions, measured what they did, and wrote the result up in Twenty rules cut the stink in our drafts and showed us what they can't fix. The rules helped, up to a point. Past that point the next rule bought less than the one before, and the drafts still sounded like a model being careful.
Adding A reviewer
So for the second fix we left the prompt alone and added an AI review pass after the draft. The reviewer read the finished text and named its faults sentence by sentence. Then the writer fixed the sentences it had been pointed at and touched nothing else. Editors have worked like this for as long as there have been editors. We wanted to know whether one model could play both parts, and what a second pass would do to our style score and to a human ear. This post is one of four on the subject, and the series is summed up in A style score for AI drafts, and where it stops.
The score is W10V4. It takes 24 measures of a draft's habits, among them sentence length and how much it swings, how sentences open, who the subjects are, how often things are given minds, how many subjects stand for a whole category and how much the draft hedges. We colour each measure green, amber or red against a set of writers we chose, and roll the lot up into one number out of 100. W10V4 counts habits. On whether a draft is any good it has nothing to say, and the distance between those two things turned out to be the story.
Our Writer, the agent that drafts posts, runs on Anthropic's Claude Opus 5.5 with our system prompt. That prompt covers house style and what the Writer may and may not say about the company it writes for. Until 26 October it ran on Alibaba's Qwen 3.7 Plus, and the change of model behind it is its own post, Qwen tied Opus at a tenth of the cost, so we made it our Writer.
A reviewer wrote 149 notes on six drafts
Round four ran on 26 October. We gave the Writer a new prompt, a prose description of the style we wanted, with no list of rules in it anywhere. A reviewer, also on Opus 5.5, then read each of six drafts and wrote one note per fault, with the quote, a tag, the reason it counted as a fault and a suggested fix. The Writer applied the notes. By the end the reviewer had written 149 of them, between 23 and 29 a draft, which was a lot of red ink for prose we had thought was decent.
The tags were the interesting part. The biggest by far was the "not X, but Y" shape, with 31 notes. It showed up in every one of the six drafts, subheads included, about five times a post, and the prompt had never mentioned it. Nobody asked the model for that sentence. It reached for it the way a nervous best man reaches for the stag weekend story, out of habit and against everyone's interest. A warm description of the style we wanted did nothing to stop it.
After that came intensifiers with 11 notes, facts about the company that nobody had supplied with 10, colons bolting an explanation onto a claim with 9, and causes stated as settled with no source behind them with 8. Invented scenes, decorative numbers and echoes of the brief got 7 or 8 each. A decorative number is the "three weeks" or the round percentage a draft adds because a figure sounds like evidence. It has nothing behind it, and it is the first thing a sharp reader stops trusting.
Round four also settled a question about the prompt. The reviewer caught the mechanical faults, intensifiers, bolted colons and announcements of what the post was about to say, every single time. Those never need room in the Writer's prompt again, since a second pass cleans them up for the price of a call. The "not X, but Y" shape is a different animal. It turned up so often and so stubbornly that it earned one line in the prompt, written plainly, with the review there to catch whatever slips past.
The Writer invented ten facts about the company
We worried about the ten invented facts more than the rest put together. Ten times across six drafts, the Writer said something about the company it was writing for that appeared nowhere in what we had given it. For a product whose whole case is that a company should publish only what it can stand behind, that is the fault that matters. The reviewer caught them because it had the brief open beside the draft. The writer, busy being fluent, never made that comparison about its own work.
The score got worse that round
Then we scored the drafts, and the number went the wrong way. Round four used 18 measures on each of the six drafts, 108 cells in all. The legacy prompt had left 41 of them red. The new prompt took that to 50, and the new prompt plus review took it to 53. Green cells fell in step, from 43 to 37 to 34. We had added a careful second reader and made the drafts worse by our own measure.
The cause was a trade the reviewer made on its own initiative. Where a draft ran several sentences of "you", the reviewer's fixes swapped in "a founder" or "the team". For anyone who has been told the reader should not be addressed in every line, that looks like a sensible edit. But W10V4 counts a category subject as a fault of its own. Across the drafts, category subjects rose from 45.6% to 51.5% of main clauses. The reviewer had fixed one habit by committing another, which is the most human thing any of our models has done.
Still, we came out of round four convinced the notes were the better fault list. W10V4 can count how many category subjects a draft carries. It names no sentence. A note quotes the sentence, says what is wrong with it and offers a fix, and a writer, human or model, can act on that in seconds. W10V4 showed us the pattern, the reviewer's notes quoted each case, and we made every repair at the case.
So in round seven we handed the finding to the scorer. W10 flags the sentences that sit far from our writers, such as a thing given a mind, a kind standing in for one case, or a staged scene with a "the" subject. Opus 5.5 then rewrites only those, as find-and-replace edits on exact sentences. With that change we took the call on what was wrong away from the reviewer, and the model's job shrank to rewriting the sentences W10 had flagged.
Four drafts went into a blind test
On 30 October we ran the pass on a fresh brief, "Put your prices on your website, even when the answer is 'it depends'". It was a brief with an opinion in it and plenty of room for a model to pad. We had four drafts of about 2,500 words each written from it and scored all four on W10V4.
W10V4 scores of the four blind-test drafts, from start to finish A. "Who acts" prompt, two self-review passes 40 → 92 D. Production Writer, one self-review pass 67 → 88 C. Production Writer, no pass 67 B. Plain Opus 5.5, no prompt 58 W10V4 scores of the four blind-test drafts, from where each started to where it ended. Reading blind, our human reviewer put A ahead of B, called C on par with A, and could not tell D from C.
Draft B came from Opus 5.5 with no system prompt at all and scored 58, with eight measures in the red. Draft C came from our Writer, the same model with our prompt, and scored 67. Draft D was C after one round-seven pass by the same Opus 5.5. The pass proposed 16 edits, all 16 were applied, and D scored 88. Draft A came from a prompt we are still testing. On its own it scored 40, the worst of the four, and after two round-seven passes and 32 edits in all it finished on 92.
The edits were surgical. All of them were find-and-replace changes that landed on the exact sentence the scorer had named, and D kept every fact and the whole structure of C. Nothing was reorganised, no section was added, and the argument stayed where the Writer had put it, so the 21 points between C and D are about as clean a measurement as we know how to take. Same model, same draft, sixteen sentences changed.
Judged on the scores alone, the decision looked easy. A second pass cost one call and bought twenty-odd points, and you would be mad to skip it. Then a human reviewer from the Propaganda team read the drafts.
It's the Pepsi Challenge all over again
In 1975 Pepsi set up tables in shopping malls with two unmarked cups, one of Pepsi and one of Coca-Cola, and asked shoppers which they preferred. More of them picked Pepsi, and the Pepsi Challenge went into Pepsi's television advertising. Our blind test was the same game played with prose.
Our human reviewer read a 1,000-word excerpt from the middle of each draft without knowing which was which. He found A natural and put it ahead of B. He called C very smooth, the best of the lot or on par with A. And he could not tell D from C. The draft that had gained 21 points sounded to him like the draft before it. His summary was that all four sat within five to ten points of each other by his ear, on a set whose scores ran from 58 to 92. So what gives?
The limits of human review
The two measures were looking at different things, and both were right about what they saw. We built W10V4 to count habits across a whole text, among them the share of main clauses with a category subject, the spread of sentence lengths and how often a sentence opens before its subject. Those are proportions, and they only show up in aggregate. A person reading 1,000 words meets one sentence at a time. Sixteen edits spread evenly across 2,500 words would put six or seven inside a 1,000-word excerpt, and none of the sentences they replaced was ugly on its own. A thing given a mind once makes a vivid line. Given a mind in every other paragraph, the same trick turns a post into brochure copy, and only a count can show where that tipping point sits. Our human reviewer, for his part, read the four drafts at midnight on a Friday (which is what happened), after a long day of reading almost identical versions of the same article. He did not read the whole 2,500 words four times, or even the 1,000-word excerpt four times. He spent a couple of minutes on each, and from that distance the drafts felt within five to ten points of each other, closer than they are.
Malcolm Gladwell made the same point about the cola test in Blink. On a single sip, tasters pick the sweeter drink, even when they would prefer the other one over a whole can. Two minutes on 1,000 words of prose is a sip.
Over months of weekly posts, we'd bet, those habits add up into a texture readers feel without being able to name it. We have not tested that on readers. It stays a bet, and it is the reason we still care about a score that one careful reading could not confirm.
The pass brings a tic of its own
There was one thing our human reviewer did spot at once. We wrote into the revise prompt that the model should fall back on "We think" or "Our bet is" whenever nobody in the brief had done the thing a sentence described. The intent was sound, since an opinion we own beats an invented customer every time. The result was an infestation. Counting "We think", "Our bet" and "we'd", draft A had 20 of them, C had 10 and D had 15. Five of D's 16 edits added one, and the scorer, which has no measure for them, waved every one through.
Four of the 12 edits of the first pass on this post. W10V4 marked each sentence and named the habit, and the Writer wrote the replacement.
| What W10 flagged | Before | After |
|---|---|---|
| A thing doing a person's verb: the score "tells you". | A score tells you a draft has too many category subjects. | W10V4 can count how many category subjects a draft carries. |
| A thing doing a person's verb: a month "will tell us". | A month will tell us whether that holds on briefs we did not pick, and whether the plain-claim fix kills the new tic or just moves it somewhere less obvious. | After a month we will know whether that holds on briefs we did not pick, and whether the plain-claim fix kills the new tic or just moves it somewhere less obvious. |
| A kind standing for every case, in the timeless present: "prompts behave". | Prompts behave a lot like airport security. | Our prompt had started to behave like airport security. |
| A staged scene in the timeless present, with nobody in it: "the extra lines make everyone feel safer". | The extra lines make everyone feel safer, and the queue gets longer. | We felt safer with every extra line, and the queue got longer. |
Our human reviewer caught the "Our bet" that opened A the moment he read it, and called it a tell. That is the uncomfortable part. We built a pass to scrub one family of machine habits out of our drafts, and it installed a new one, with our name on it, in the exact spot a reader looks first. The scorer could not object, since you cannot penalise a phrase you do not count. And the model, handed a fallback, used it as often as the rules allowed.
The fix is small. We will drop both phrases from the prompt and keep a single instruction, to say the thing as one plain claim when nobody in the brief did it. A plain claim carries no signature, and that is the whole idea. The wider lesson is one that anyone who has built a metric learns eventually. A model rewriting toward a score will find the phrases the score cannot see, and the only defence is a rule about what the rewrite may add.
Plain Opus 5.5 writes well
The most humbling result in the test belonged to B. Opus 5.5, with no system prompt, no house style and no rules, scored 58 and carried eight red measures. Our human reviewer called its text great. On a good brief, the bare model wrote prose he would have been happy to publish, with eight habits the scorer could still name.
We took two things from that. The first is that the model does most of the work. Our prompt, our score and our review pass sit on top of something that already writes well, and what they do is take the stink down by trimming the reflexes that make a reader suspect a machine. None of it turns Opus 5.5 into one of the writers we score it against. Those writers had something to say and years of practice saying it, and no quantity of find-and-replace manufactures either.
The second is about what we sell. Our pitch to a founder is a good model, kept out of its habits, at a price that lets them run it every week. It is a pitch about the floor. The ceiling belongs to the model and to whoever has something worth saying, and a style score has no business pretending to raise it. If you are weighing one more model call per post, weigh it against that floor. A review pass lifts the bottom of the range, and the top was always going to come from you.
Self-review beat a second reviewer
The short version: the writer fixing its own flagged sentences scored within three points of a separate reviewer, in a ninth of the time, at a quarter of the cost. We ran the two against each other on 31 October, on the same 2,500-word draft, from the same starting score of 77, with the same sentences marked by W10V4. In one arm the Writer read the marked sentences and rewrote them itself. In the other a second Opus 5.5 call, under the round-four reviewer prompt, read the whole draft against the brief, wrote up to thirty notes, and the Writer applied them. Two passes each. Self-review finished at 96, the reviewer at 93, and we read that gap as noise.
Writer fixing itself against a separate reviewer: score, time and cost for two passes on the same draft Writer fixes its own flagged sentences Separate reviewer, same flags Writer alone, no flags W10V4 score, out of 100 96 93 85 all started at 77 Time, both passes 24 s 217 s 52 s the reviewer is two calls a pass Cost, both passes 1× 4× 1.3× unaided, the writer thinks longer Same 2,500-word draft, Opus 5.5, two passes each. 31 October 2026.
The reviewer earned its keep in one place only. It argued with the scorer, kept five of the fourteen marked sentences as they were, and caught facts the scorer cannot see: "read the drafts" became "read a stretch from the middle of each", which is what happened. Then it cut the voice. Eighteen sentences went in two passes, 11% of the words, among them the airport-security line and the one that said you would be mad to skip a second pass. Round four had cut 9% the same way. The bill is where the two arms part. A reviewer reads the whole draft, thinks for six thousand tokens and writes three thousand more, and output tokens cost five times what input does. The Writer reading fourteen marked sentences thinks for a thousand and answers in nine hundred. Self-review took 24 seconds for both passes. The reviewer took 217.
The scorer finds what the model cannot
That left one question we had never asked. Was it the instructions in the revise prompt doing the work, or the list of sentences W10V4 handed over? So we ran the same prompt a third time with the list removed. The Writer was told what a bad sentence looks like, a thing with a mind, a kind doing a person's verb in the timeless present, and asked to find them itself. Same draft, same two passes.
It finished at 85. In the first pass it made eighteen edits, and five of them landed on a sentence the scorer had marked. The other thirteen went to sentences the scorer had no complaint about, so most of the habits it was sent to remove stayed where they were. The second pass found four more. It also thought for twice as long as the flagged pass, 2,500 reasoning tokens against 1,000, and cost a third more for ten points less. The model knows the rule and cannot see its own breaches of it, which is the whole case for a scorer that counts instead of reads. We think that is also why the round-four reviewer did no better: it was the same model reading the same prose with the same blind spot, and a longer prompt does not give a reader eyes it does not have. Pointed at a sentence, the model fixes it, and asked to find one, it fixes something else.
The pass is still on the bench
So the second pass stays a finding, and the Writer runs without it today. We have not decided to build it. Shipping it would mean one more Opus 5.5 call after every draft, a parser on the box to flag the sentences, and the two fallback phrases gone from the prompt. None of that is hard engineering. What we lack is evidence from real work at real length.
Before any of that, we want a month of real posts at 2,500 words, scored as they land. One brief and one blind reading showed us a pass that moves the number by a lot and the experience by very little. After a month we will know whether that holds on briefs we did not pick, and whether the plain-claim fix kills the new tic or just moves it somewhere less obvious.
For anyone who wants to try the pattern before we finish, the recipe has four parts. You need something that names the weak sentences, a scorer if you have one and a checklist if you don't. You need a model that rewrites only those sentences and leaves the rest alone, ideally as exact find-and-replace edits so you can see every change. You need a rule that the rewrite may not add a phrase the brief did not support, and that includes the friendly fallback phrases you wrote into the prompt yourself. And you need your own eyes after the score. Ours found the tic the scorer waved through, in the first two words of the highest-scoring draft.