AI Tell Scrubber

A prompt for cleaning AI-generated markers from prose

⬇ Download the Markdown version

How to Use This

Paste this document into your context window, then submit the text you want cleaned with the instruction: “Apply the AI tell scrubber to the following text.”

First, decide whether to run it at all

This operation is not free. A full scrub of a long document is real work and real token cost, and most of that cost is wasted on text nobody will scrutinize. AI tells only matter when the reader’s trust is at stake, which means external-facing writing: anything going to a customer, a prospect, the public, a publisher, or anyone who would think less of the work if they clipped it as machine-made. For internal notes, scratch drafts, personal logs, code comments, and anything the writer alone will read, the tells cost nothing and the scrub is wasted effort. Default to not running it.

So before applying either pass, settle the question of whether it is warranted:

• If the scrub was explicitly requested (“apply the scrubber,” “clean this up,” “remove the AI tells”), proceed. Intent is already clear; do not ask.
• If the text is obviously internal or low-stakes (a short note, a scratchpad, something the user has said is just for them), skip the scrub and say so in one line, rather than burning tokens on it.
• If external exposure is plausible but unstated, and the document is long enough that a full scrub is non-trivial, ask one short question before proceeding: “Is this going somewhere external where reading as AI-generated would hurt? If it’s internal, I’d skip the scrub and save the effort.” Then act on the answer.
• If the document is short, do not bother asking. The cost of the question exceeds the cost of just doing it or skipping it. Use judgment and move on.

The goal of this gate is to spend one cheap question, at most, to avoid a large amount of pointless work. When in doubt on a short document, lean toward skipping rather than interrogating.

The two passes

The scrubber works in two passes:

1. Mechanical pass: find and fix specific flagged words, punctuation patterns, and constructions
2. Judgment pass: identify structural and tonal problems that require rewriting rather than substitution

Both passes are described below. Apply them in order.

For long documents, work section by section. The mechanical pass can be run against the full text; most of the judgment pass requires reading each paragraph.

One exception: the frequency watch (section 11) cannot be done section by section. A pet phrase used once per section reads as fine locally and as a tic globally, so section-by-section reading is the exact procedure that hides it. Run the frequency watch as a final, separate pass across the whole document, after the section-by-section work is done.

Pass One: Mechanical

1. Em Dashes

The single highest-confidence AI tell in prose. Flag every em dash ( — ) in the document.

For each one, determine what relationship exists between the clauses it connects, then replace it with the correct punctuation:

Exception: Em dashes used as intentional stylistic devices (a sentence deliberately broken mid-thought to show a character stopping themselves, a transmission interrupted by static, a diary entry that can’t find its ending) may be kept if they are doing specific narrative work. These should be rare. One per page at most. If they are appearing every paragraph, they are not intentional; they are habit.

2. The Slop Lexicon

Flag and replace every instance of the following. Replacement should use a more specific, concrete, or accurate word for the context, not a synonym from the same register.

Adjectives / adverbs to replace: crucial, pivotal, intricate, meticulous, vibrant, seamless, robust, groundbreaking, transformative, compelling, invaluable, paramount, commendable, nuanced, comprehensive, stark, genuinely, certainly, notably, importantly

Verbs to replace: delve, underscore, bolster, foster, harness, leverage, navigate (when metaphorical), illuminate (when metaphorical), facilitate, embark, endeavor, unlock, elevate, empower, utilize (use “use”), commence (use “begin” or “start”)

Nouns to replace: tapestry, realm, beacon, landscape (when metaphorical), testament, interplay, paradigm, cacophony, journey (when metaphorical)

Phrases to cut or rewrite:

• “ever-evolving” → say what is changing and how
• “not just X but Y” → pick one or rewrite entirely
• “it’s not X, it’s Y” → use sparingly; once per document at most
• “more than just” → cut; state what it is
• “at its core” → cut; state the thing directly
• “in today’s world” → cut entirely
• “as we navigate” → cut entirely
• “in conclusion” → cut; end without announcing that you’re ending
• “in summary” / “overall” → cut; the reader was there
• “it is worth noting that” → cut; just note it
• “it is important to” → cut; just say the thing
• “a deep dive” → cut; say what you’re doing specifically

3. Physical Reaction Tells (Fiction Only)

In fiction, flag every instance of the following stock physical reactions. These are not wrong individually; they are overused to the point of becoming meaningless, and they appear wherever a beat is needed rather than where they are earned.

The sigh family: sighed, exhaled slowly, let out a long breath, released a breath he/she didn’t know he/she had been holding

The pause family: paused, hesitated, fell silent, took a moment, let the silence stretch, considered for a long moment, “for a moment”

The nod family: nodded, gave a slow nod, nodded thoughtfully, nodded once

The chuckle family: chuckled, gave a short laugh, let out a humorless laugh, snorted (when used as a reaction beat rather than a described action)

The eye family: raised an eyebrow, arched a brow, met his/her eyes, held her/his gaze (as a reaction beat)

For each flagged instance, ask: is this reaction carrying subtext, or is it filling a beat? If it is filling a beat, cut it and let the next line of dialogue or action carry the weight. If the reaction is genuinely doing work (revealing character, carrying specific meaning), rewrite it with more precision. What exactly does this character do? What is specific to them, in this moment?

4. Transition Word Audit

Flag every paragraph that opens with: furthermore, moreover, additionally, in addition, consequently, subsequently, therefore, thus, hence, indeed, certainly, notably, significantly

Most of these can be cut entirely. The logical connection should be carried by the content, not announced by a transition word. If the logic doesn’t flow without the word, the sentences need to be reordered or rewritten, not bridged.

Pass Two: Judgment

These problems cannot be fixed by find-and-replace. They require reading each paragraph and making editorial decisions.

5. Uniform Sentence Rhythm

Read the text aloud, or simulate reading it aloud. Flag sections where sentence length and structure are consistent across three or more consecutive sentences.

AI generates prose with even rhythm. Human writers are irregular. Short sentences break long ones. Fragments appear. The pattern of long-medium-long-medium is a tell even when no individual sentence is wrong.

Fix by breaking the pattern. Insert a short sentence where the rhythm has been running long. Combine two short sentences into one complex one. Vary the position of the main clause. The goal is not randomness but the specific rhythm of a human mind thinking through a problem in real time.

6. Compulsive Summarizing

Flag any sentence that restates something the preceding sentence or paragraph already established. Common forms:

• “This demonstrates that…”
• “In other words…”
• “What this means is…”
• “As we can see…”
• A concluding sentence that summarizes the paragraph it closes

Cut all of these. Trust the reader to have read the preceding content. If the point needs restating, the original statement was not clear enough. Fix the original, don’t add a summary.

7. The Rule of Three

Flag every list of exactly three items, every sentence with three adjectives, every paragraph with three examples.

AI defaults to triadic structure because it feels complete. Count how many times the pattern appears. If it appears more than twice in a document, break it up. Use two-item lists. Use four. Use one strong example instead of three weak ones.

8. Over-Explained Emotion (Fiction Only)

Flag any sentence that names an emotion and then explains it:

“She felt a complex mixture of grief and relief, the two emotions intertwined in ways she could not fully articulate.”

Cut the explanation. Name the emotion at most, or cut the naming and show the behavior. The reader will do the emotional arithmetic.

Also flag: any sentence that tells the reader how to interpret a scene that the scene itself is already conveying. The narrative voice stepping in to say something was significant, troubling, or ironic when the scene itself already makes this clear. Cut the editorial layer. Trust the scene.

9. Performative Balance

Flag any passage where a position, character, or argument is presented with both its strengths and its weaknesses in even measure, producing a “on the one hand… on the other hand” structure.

This is not always wrong. But AI produces it reflexively, even in contexts where a human writer would commit to one side. In persuasive writing, commit. In fiction, characters should be partial; they should not explain themselves fairly. Flag any character who seems to be presenting a Wikipedia article about their own motivations.

10. Generic Sensory Detail (Fiction Only)

Flag sensory description that uses expected, interchangeable details:

• cold air
• warm light
• heavy silence
• the smell of rain / coffee / something burning
• a feeling of weight in the chest

These are not wrong. They are placeholders. The question is whether this specific character, in this specific moment, would notice this specific detail, or whether the detail is simply what the scene calls for in general.

Replace with something that could only appear in this scene. The sensory detail that is specific to this character’s history, training, or state of mind. If you cannot identify what that would be, leave the description out entirely. Absence is better than placeholder.

11. Frequency Watch (Repeated Pet Phrases)

The hardest tell to catch, because it is not a specific word. Any phrase becomes a tell when it repeats more often than a human writer would allow. The marker is the repetition itself, not the phrase.

This catches the tells that the slop lexicon misses. A phrase can be precise, useful, and absent from every published list, and still expose the text as machine-made simply because the model reached for it four times in two pages. Human writers vary their phrasing, partly from boredom and partly from ear; models return to a favored construction and do not notice they are doing it. Known offenders in this category include “load-bearing” (metaphorical), “it’s worth noting,” “that said,” “the reality is,” “at the end of the day,” “the key insight,” and any distinctive metaphor reused across sections. But the list is not the point. The point is the method.

How to run the check, since you cannot find-and-replace your way through it. Run it as a single pass over the entire document, not section by section. This is the one check that breaks if you work in chunks: a phrase used once per section passes every local read and only reveals itself as a tic when you see the whole thing at once. So save it for last, after the rest of the judgment pass is done, and read the full text in one go.

1. Read straight through the whole document and note any distinctive phrase or metaphor you have seen already earlier in it. Distinctive is the operative word: ordinary connective tissue (“and,” “but,” “for example”) does not count. You are hunting for phrases with a flavor.
2. For each repeated phrase, count the instances. Once is invisible. Twice is usually fine. Three or more in a short document is a tic, and the reader will feel it even if they cannot name it.
3. For every instance past the first, either cut it or say the thing a different way. Do not reach for a synonym from the same register; rephrase the idea. “Load-bearing” becomes “the part everything else depends on” or “the one assumption holding this up,” and crucially, it becomes those things only some of the time, so no single replacement turns into the new tic.

A note on legitimate reuse: a genuine term of art repeated in a technical document is not a tell. If a piece is about load-bearing walls, the phrase will recur and should. The judgment is whether the repetition serves the subject or just reflects the model’s habit. When the phrase is doing real, specific work each time, keep it. When it is filler that happens to sound smart, vary it or cut it.

A tooling aid: the judgment of which repeats are tics stays human, but finding the candidates does not have to be. A word-frequency count over the document, or just Ctrl-F on a phrase you suspect, surfaces the repeats far faster than re-reading. Use the tool to build the list of suspects; use your ear to decide which ones to fix.

Quick Reference Card

For use after the full scrubber has been applied at least once.

Apply to both the original text and any revised output. Tells introduced during rewriting are as damaging as tells in the source. Run the mechanical checks on every draft, including your own.

Mechanical checks:

• Every em dash replaced with correct punctuation (colon / semicolon / parentheses / comma / period)
• Slop lexicon scanned and replaced
• Transition word openers at paragraph starts cut or rewritten
• Physical reaction beats (sigh / nod / pause / chuckle) audited
• Revised output checked separately: rewriting prose to fix one tell can introduce another

Judgment checks:

• Sentence rhythm varies within and across paragraphs
• No compulsive summarizing at end of paragraphs or sections
• Rule-of-three not appearing in every paragraph
• Emotion shown in behavior, not explained in narration (fiction)
• Characters are partial, not self-explaining (fiction)
• Sensory details are specific, not interchangeable (fiction)
• No distinctive phrase or metaphor repeated more than twice (frequency watch)

Notes on Application

The goal is not to remove all AI influence. If a passage was AI-generated and is good, it should stay. The goal is to remove the markers that make AI-generated prose feel synthetic: the tells that break the reader’s trust before they have consciously identified why.

Some tells are era-specific and will age. The slop lexicon was current as of 2024-2026. Words that were AI-marked in that period may become neutral again as models diversify their output. The structural and tonal tells (uniform rhythm, compulsive summarizing, over-explained emotion) are more durable because they reflect how language models generate text, not just what vocabulary they favor.

This scrubber is not a style guide. It does not prescribe a particular voice. It removes patterns that undermine any voice. What replaces them should reflect the writer’s own choices, not a different set of defaults.

When in doubt, cut. Most AI-generated text is longer than it needs to be. The slop lexicon words are often filler that can be removed without replacing. The transition words are often not needed. The summarizing sentences are rarely needed. Default to shorter. The reader is not lost without the scaffolding.