AI Writing Prompts: Fix the Rhythm, Not the Word List

AI writing prompts that ban a list of words fix the wrong layer. You can delete "delve" and the draft still reads the same, because what gives generated copy away is not the vocabulary. It is the evenness.

The Word List Is the Layer Everyone Fixes

The banned-word list is real, it is the most visible layer of the problem, and it has two flaws no list can repair. It expires: every time a word becomes the famous tell, the next generation of models stops leaning on it. And it misses the cause, since banning a word leaves untouched the process that produced it.

Write down the vocabulary you use and the vocabulary you avoid as a specification instead of a wish. The catalog already carries that half, so this post is about the half it cannot carry.

What the Largest Study of This Actually Measured

The most cited study on AI-flavored vocabulary looked at more than 15 million biomedical abstracts indexed by PubMed between 2010 and 2024 and measured an abrupt increase in the frequency of certain style words after 2022. The authors read that excess as a lower bound: at least 13.5 percent of 2024 abstracts had been processed with an LLM, a share reaching 40 percent in some subcorpora (Kobak et al., 2024). Read it as a dated symptom, not a style specification.

Nothing Here Is About Escaping a Detector

This is not a guide to passing an AI detector, and the reason is not modesty. Detectors fail in batches: a study of widely used GPT detectors found they consistently misclassified writing samples from non-native English writers as AI-generated, while native samples were identified accurately (Liang et al., 2023). A tool that flags a careful second-language writer measures how ordinary the text looks, not who wrote it. What you can measure is the draft: sentence count, the shortest sentence, and how many sections end by repeating what you already said.

The Three Scales Where AI Writing Prompts Show Their Hand

Evenness is not one defect but three, and they stack. Fix only the first and the reader still recognizes machine output by the third paragraph.

Scale 1 - The Sentence

The problem is not the long sentence, it is that there is no short one. Rhythm lives in the variance of sentence length. Detection literature names two statistics for this, perplexity and burstiness, and the second measures exactly that variance. Left alone, a model drifts toward the average length.

Scale 2 - The Paragraph

Paragraphs come back in one mold: topic sentence, three supporting sentences of similar length, and a closing line that restates the opening - "in other words", "the key takeaway is". That closing sentence is the most audible tell in the draft and the cheapest to delete. A paragraph that ends on its last piece of evidence reads like a person who trusts you to keep up.

Scale 3 - The Set

Evenness also lives between texts. A calendar generated in one session opens every piece with the same question, closes every piece the same way, and ends every section with a list of three. Individually the posts are fine; in a feed they read as one long document.

Scale	What evenness looks like	The one move that breaks it
Sentence	Every sentence roughly the same length	Write one sentence under six words per paragraph
Paragraph	Topic, three supports, a wrap-up	Delete the wrap-up sentence
Set	Same opening, same closing, a list of three per section	Change one piece so it breaks the pattern

Stop Describing the Voice - Show Three Samples

Asking for "warm but direct" or "conversational" does not produce your voice. It produces the average of everything the model has seen described that way. Voice is a pattern that already exists in something you wrote, and it can be extracted if you paste that something in.

The Rhythm Brief (Four Fields)

Four fields, pasted in front of any writing prompt you already use. Each takes two minutes to produce and can be checked in the draft.

1. SAMPLES
   Paste 2-3 pieces of your own writing in the SAME format you want back
   (not an essay if the deliverable is a 120-word email).

2. THE PATTERN
   "Report the typical length of my sentences, the longest and the
   shortest, how I open, how I close, and what I never do. Then write
   following that pattern."

3. THE MATERIAL
   The facts, numbers and names the model is allowed to use. Yours.

4. THE CONSTRAINT
   On the revision pass the model MAY cut, and MAY NOT add facts,
   figures or names that are not in field 3.

Ask for the Pattern, Not the Adjectives

Field 2 is the one that does the work, because it produces a checklist you can verify: "opens with a short declarative sentence, longest sentence runs 34 words" can be counted in the draft and again after an edit.

The Unevenness Pass - Four Moves on the Draft

Four countable moves, run on the draft rather than the prompt, in this order.

1. Break One Paragraph Into a Single Sentence

Take the flattest paragraph and leave one short sentence standing alone. It is the cheapest of the four and the one that changes the reading most.

2. Delete the Wrap-Up Sentence

Find the section that ends by restating itself and delete that line. What disappears is the tic readers hear as machine output.

3. Keep One Specific Per Section

Every section should carry one number, date or name from your material that the model could not have invented. A section with none is a candidate to be cut, not padded.

4. Leave One Limit Standing

State one real limitation without rescuing it with "but luckily". Generated text resolves every tension it raises, which is why it reads like a brochure.

The Set-Level Version

When the deliverable is a series, change one piece so it does not follow the house mold - the one that does not open with a question. Contrast is what reads as judgment. And what this pass is not: a way to beat a detector or a word list. The only reader who matters is the one who replies to the email.

Three Prompts From the Catalog, and Which Half of the Voice Each One Owns

The structural half is already written down, and each of these covers a different scale.

The Long Form - the process and its structural rules

SEO Article Writer System runs a five-step process with the rules of form inside the prompt: paragraphs of no more than three sentences, one idea per paragraph, concrete numbers and examples, and a line telling the model not to invent statistics. That last rule is field 3 of the rhythm brief, written by someone else. What the prompt cannot supply is the sample - a list of cliches to avoid is still a list.

The Set - where uniformity multiplies

Social Media Content Manager is the scale 3 case in a box: four pillars, a four-week calendar, eight fully written posts in one pass. Its own rule, that every post must be saveable or shareable, is the limit the user has to enforce on the draft, and enforcing it across eight pieces is where the same opening and closing creep in.

The Shortest Format - where the tell costs the most

Cold email template that gets replies caps everything at 120 words and asks the first line to reference something specific about the recipient's business. With a budget that small, a filler sentence does not read as filler. It reads as a template.

All three carry the half of the voice that can be specified: structure, fields, limits. The other half is two or three samples of your own writing. Start with the SEO Article Writer System for a long piece: the rhythm brief is what turns its structural rules into your voice.