Machine writing has tells. Some are a single sentence shape, and some only show up when you count. We wrote our rules down on October 12, tested each change against a scorer, and learned where rules help and where they stop. This post is the story of those rules. It is one of four on how we got to our style score. The summary is a style score for AI drafts, and where it stops.
The rules
There are twenty of them, and the shortest summary of each fits on one line: no decoy lead-in, no "not X, but Y", one job per sentence, no reflex triplets, no trailing -ing phrases, never "quietly" or its family, no intensifiers, no marketing words, verbs and not nouns, numbers sourced or cut, titles that make one plain statement in sentence case, no curtain-raiser or closer, zero em dashes, no drumroll, lists only for lists, name things, one real thing beats a general truth, do not announce, take a side and say where it stops, and write for one reader as one company.
We split them into two kinds. Mechanical rules, such as em dashes and "quietly", can be checked by code, and a failing check can send the draft back.. Rules about meaning, such as "one real thing beats a general truth", get judged by one small model call, because a regular expression cannot tell whether a sentence says something. We never hard-fail a soft rule. One of us put it plainly: if all twenty were enforced in code, nothing would pass.
The first test
We ran the old prompts and the new ones on the same inputs and counted rule breaks.
| Before | After | |
|---|---|---|
| Pitch titles that broke a rule, of 8 | 2 | 0 |
| Strategy topic names that broke a rule, of 4 | 1 | 0 |
The cost was about a cent more per pitch and per draft, because the prompts got longer. Before the rules, pitch titles came out in title case ("Your AI Writer Will Make Up Product Facts"). After, they read like sentences: "Most startup blogs stop after eight posts".
Where the rules ran out
The cleanest finding came in round five. We gave the Writer a prompt with a portrait of the target style, target numbers and a self-check, and it scored 9 greens of 18 measures. Then we added our twenty rules to the same prompt. It dropped to 6 greens and 7 reds, and 31 percent of its paragraphs became flat runs of similar sentences. A longer list of rules made the draft worse by the scorer's count.
Three patterns explain most of what we saw.
A rule is obeyed where it points and paid for elsewhere. We restored the rule that steers the Writer away from "you", and "you" fell from 3.3 per hundred words to 0.7. Then the subject of its sentences moved to a category noun. Our rule said sentences should mostly have as their subject "the thing that did something: the model, the page, the invoice", and the Writer read that as permission to put things in the subject seat. When we removed the "you" rule in an earlier round, "you" came back at three to six per hundred words.
The examples teach more than the rule text. Rule 20 banned crowds as subjects and gave a model rewrite: "Keep your prices in one place and your posts stop disagreeing". A post, which is a thing, now does two verbs that need a mind, "stop" and "disagree". The passive sentence it replaced was more honest, since nobody acted in it. Every rewrite needs its new subject read back: who acts here, and could that thing do that?
A fix overshoots. We told the Writer to give the verb to a named person. "We" then took 56 percent of the subjects, and "we would" slid into habits we never had ("We round a price down to sound friendlier"). When we fixed rhythm on eight real drafts, mean sentence length went from 9.9 to 26 words, and sentences of 25 words or more went from 3 percent to 48 percent. The writers in our reference sit between 11 and 42 percent.
The fault under most of the faults
The same fault showed up under several names. We called the version with generic subjects the Jaqen habit: "A model answering a question lifts the passage that answers it". Nobody in it, and a kind of thing stands in for a particular case. The fix is particularity, such as a place, a time, a named actor, a detail with colour. "When AI models answer a question, they lift the first passage that answers it" reads as a process that keeps happening, and the first version reads like a law of nature.
We also learned it is a density fault and not a ban. "Your blog, newsletter and LinkedIn disagree" is a smart thing to say once. Said three times in a post, it reads as generated. Things can act ("the search returns nothing" is fine). The fault is a pile of mind verbs on things: decide, disagree, know, want, tell. So the rules now say cut them hard, and they stop short of forbidding them.
What house style does
It cuts the stink. With the "people do the verbs" rule on the sales-call post, things acting as subjects fell from 17.6 percent to 8.6 percent, inside the writers' range, and the draft went to 13 greens of 18. Our first eight drafts had a median of 2.5.
It does not make a draft the work of a writer people seek out. The model sets most of the ceiling, and the rules lift the floor. When we tested plain Opus on a good brief against our full prompt stack at 2,500 words, the plain draft read as good prose to its reader, and the stack's advantage was smaller than the score said. That does not make the rules pointless. A tell repeated forty times across a blog is what a reader feels over months, and no single excerpt shows it.
What we would pass on
Write the rule, then write the example that goes with it, then read the example the way the model will.
Check the new subject of every rewrite.
Test each rule as the only change. Our worst round was the one where we added everything at once.
Treat style rules as densities, never as bans, unless the shape is a character-level tell like an em dash.
Keep a scorer next to the rules, and keep your own ear above the scorer.