Skip to content
VerifiedUpdated 2026-06-27
On this page

The design of a site is not just its color and type. It is also how it reads. And nothing dates a page faster, or quietly costs it trust, than copy that sounds like it came out of a chatbot. This page is the writing half of the brand: the specific habits that mark text as machine-generated, and the plain-writing rules that avoid them.

It matters more now that AI writes so much of the web, including, often, the first draft of your own copy. If you let an assistant draft your marketing page, your docs, or your commit messages, it will reach by default for an inflated register that readers have learned to spot. The fix is not to stop using AI. It is to tell it, and yourself, what good looks like.

Where these rules come from

This is not personal taste dressed up as law. The vocabulary list comes from two excess-vocabulary studies that measured which words spiked in published writing after ChatGPT shipped: "delve" alone ran at roughly twenty-five to twenty-eight times its prior frequency. The sentence-pattern and structure tells come from Wikipedia's editor-maintained "Signs of AI writing", from journalism on AI detection, and from editors who read AI drafts for a living. Where a rule is widely cited, that breadth is noted next to it.

Treat the plainest accurate word as the right one. The studies found that two thirds of the words AI overuses are verbs: the inflation is stylistic, not informative. Plain writing is not dumber writing, it is the writing that survives a reader who is deciding whether to trust you.

The rules

The fast ban-list

The words sources flag most. If a draft leans on these, it reads as machine text. Reach for the plainest accurate word instead.

delveunderscoreshowcasetapestryrealmpivotallandscapeintricatecrucialmeticuloustestamentrobustseamlessleveragetransformativenotablynavigatefosterpotentialcomprehensivevibrantutilizeadditionally

Vocabulary

Decorative style-words that inflate without adding meaning. The studies found two thirds of excess AI words are verbs: the spike is stylistic, not topical.

Avoid

delve, delve into, delving

Instead

explore, look at, dig into, or just name the thing directly

The single most-cited tell: two studies measured it at ~25 to 28x normal frequency, plus FSU, The Conversation, Wikipedia, and most editor lists

Avoid

underscore, underscores (as a verb)

Instead

shows, proves, or cut the sentence and state the point

Named by both excess-vocabulary studies, The Conversation, Wikipedia, Pangram

Avoid

showcase, showcasing

Instead

shows, presents, includes

Named by both studies, The Conversation, Wikipedia, Pangram

Avoid

tapestry, especially rich tapestry

Instead

drop the metaphor and describe the actual mix of things

The notorious editor flag; named by Wikipedia, The Conversation, Pangram, and several blogs

Avoid

realm, in the realm of

Instead

field, area, world, or name it: in software, in chess

Named by a study, FSU, The Conversation, Pangram

Avoid

landscape, ever-evolving landscape

Instead

name the actual situation or market

Named by Wikipedia, The Conversation, Pangram, and more

Avoid

pivotal, crucial

Instead

important, key, or say why it actually mattered

Named by studies, The Conversation, Wikipedia

Avoid

intricate, intricacies, meticulous, meticulously

Instead

complex, careful, or describe the specific complication

Named by both studies, FSU, The Conversation, Wikipedia, Pangram

Avoid

testament, a testament to, stands as a testament to

Instead

shows, proves, or state the evidence plainly

Named by Wikipedia, Pangram, and editor lists

Avoid

robust, seamless, transformative, groundbreaking, game changer

Instead

say concretely what changed, how strong, and by how much

Inflated adjectives named by Wikipedia, Pangram, and editor lists

Avoid

leverage, utilize

Instead

use

Named by Walter Writes, Learning Data, and community ban-lists

Avoid

empower, foster, facilitate, harness, streamline, elevate, unlock, navigate

Instead

help, support, enable, handle, or a plain verb

Inflated business verbs named across the editor lists and Wikipedia

Avoid

notably, additionally, moreover, furthermore

Instead

also, and, or just start the sentence

Stiff connectives named by both studies, Pangram, and editor lists

Avoid

symphony, mosaic, kaleidoscope, odyssey, beacon, cornerstone

Instead

skip the prestige metaphor noun entirely

Named by Pangram and the Vollmer field guide

Avoid

résumé, café, naïve, and other accented spellings of naturalized English words

Instead

use the plain English spelling: resume, cafe, naive

A reach for a fancier or foreign-derived form when the plain spelling is standard; the same instinct as inflated vocabulary, applied to spelling

Stock phrases

Prefab transitions and openers that signal nothing and could front any topic.

Avoid

It's important to note that, it's worth noting that

Instead

if it matters, just say it; if not, cut it

Named across the editor field guides

Avoid

In today's fast-paced world, in an ever-evolving landscape, as technology continues to evolve

Instead

open with the actual subject

Vapid openers named by Vollmer, Walter Writes, Hunting the Muse, ignorance.ai

Avoid

In conclusion, in summary, overall, ultimately (as a closing label)

Instead

end on a real last point, not a recap announcement

Closing rituals named by Pangram, Vollmer, and editor lists

Avoid

Here's the thing, here's what nobody talks about, at the end of the day

Instead

make the point without the drumroll

False-profundity fillers named by several editors and Reddit-derived lists

Avoid

shed light on, deeper understanding, valuable insights, significant impact on

Instead

state what was learned and the measured effect

Named by Pangram, Wikipedia, Learning Data

Avoid

Great question, hope this helps, let me know if you'd like me to go deeper

Instead

drop sycophancy openers and cheap-warmth sign-offs

The earnest-helpfulness tell, named by Vollmer and Pangram

Avoid

experts argue, studies suggest, observers have cited (with no source)

Instead

cite the actual person or paper, or cut the claim

Vague attribution named by Wikipedia and Pangram

Sentence patterns

Formulaic templates that recur regardless of subject. The manufactured contrast and the triplet are the loudest.

Avoid

It's not X, it's Y; it's not just X, it's Y; not X but Y

Instead

make one clear assertion without the manufactured contrast

One of the most-cited patterns, named by Wikipedia, Pangram, and at least five editor sources

Avoid

The rule of three: fast, simple, powerful; plan it, build it, ship it

Instead

use one or two items, or vary the count; reserve triplets for rare emphasis

Named by Wikipedia and five editor sources

Avoid

Question-then-answer crutch: The secret? Consistency. The result?

Instead

state the point as a plain sentence

Named by four editor sources

Avoid

No X. No Y. Just Z; stacking negations before a pivot

Instead

describe what it is, not a stack of what it isn't

Named by three editor sources

Avoid

not only X but also Y

Instead

split into plain clauses or pick the stronger point

Named by Wikipedia and Pangram

Avoid

Trailing -ing tails: ..., underscoring its importance; ..., marking a pivotal moment

Instead

end the sentence on the noun; start a new one if there's more to say

Named by Wikipedia and Vollmer

Avoid

Dressing up is/has as serves as, stands as, boasts, features

Instead

use is, are, has

Named by Wikipedia

Structure

Habits in how the whole piece is shaped: even rhythm, reflexive bullets, and templated arcs.

Avoid

Uniform rhythm: every paragraph three to four sentences of the same length

Instead

vary sentence and paragraph length deliberately; let some be very short

The burstiness tell, named by Pangram, Vollmer, and four more editors

Avoid

Over-bulleting: lists for ideas that are not list-shaped

Instead

write prose; reserve bullets for genuinely parallel items

Named by Pangram, Vollmer, ignorance.ai, Wikipedia

Avoid

Templated arc: intro, three bodies, recap; or significance, challenges, future outlook

Instead

let the content dictate the structure

Named by Vollmer and Wikipedia

Avoid

Missing fingerprints: no names, dates, numbers, or lived detail, byline interchangeable

Instead

anchor with specific names, numbers, dates, and real experience

Named by the 500-articles editor, Vollmer, ignorance.ai

Avoid

Aphoristic kicker endings: That's the whole game. And that changes everything.

Instead

end sections on substance, not a pull-quote

Named by three editor sources

Avoid

Excessive signposting: first we'll look at, second, finally

Instead

let transitions emerge from the ideas

Named by Vollmer

Punctuation and formatting

The surface tics, led by the em-dash, that survive even when the prose is otherwise clean.

Avoid

Em-dash overuse for dramatic pauses and additive qualifiers

Instead

use commas, colons, or two sentences

The most-cited punctuation tell across nearly every source, with two dedicated essays on it

Avoid

Zero contractions paired with flawless, spotless grammar

Instead

use contractions and let a natural voice show

Named by Vollmer, Pangram, and the em-dash essay

Avoid

Random mid-prose bolding without clear emphasis

Instead

bold sparingly, only for true emphasis

Named by ignorance.ai and Wikipedia

Avoid

Emoji in headings, emoji-led bullets, Title Case Headings In Odd Places

Instead

plain sentence-case headings, no decorative emoji

Named by Vollmer, ignorance.ai, Wikipedia

Avoid

Decorative unicode (arrows, curly quotes) and leaked markdown

Instead

plain straight punctuation matching the target format

Named by ignorance.ai and Wikipedia

Tone

The underlying register: inflated, relentlessly positive, hedged, and stance-free.

Avoid

Flowery style-word inflation divorced from the subject

Instead

prefer plain, precise words; flowery prose costs clarity and credibility

The core finding of both studies: the excess is stylistic, not substantive

Avoid

Relentless promotional positivity and press-release puffery

Instead

state limits and tradeoffs; be willing to call something mediocre

Named by Wikipedia, Vollmer, Pangram

Avoid

Generic, stance-free blandness with no point of view

Instead

take a position and stake a claim

Named by the 500-articles editor, The Conversation, ignorance.ai

Avoid

Over-hedging: balanced pros and cons with no verdict

Instead

commit to a judgment after the analysis

Named by Hunting the Muse, Vollmer, Walter Writes

Avoid

Throat-clearing and restating the same idea in new clothes

Instead

cut the padding; say it once, plainly

Named by the 500-articles editor, Learning Data, Walter Writes

How to apply this

You do not need to memorize the list. Two habits cover most of it:

Draft however you like, then strip the inflation

Write the first pass with whatever tool you want. Then read it once with a single question: which words are doing decorative work instead of carrying meaning? Cut or replace them. Most slop dies in this pass.

Make your AI write to these rules, not its defaults

If an agent drafts copy for you, hand it the rules. The Set up my CLAUDE.md bootstrap installs them as standing directives, so every draft your agent produces already avoids the em-dash, the inflated vocabulary, and the manufactured contrast. That is the difference between fighting the slop on every draft and never seeing it.

NextDecisions and Defaults