eliminate-ai-writing-tropes

Overview

A passive ruleset that strips detectable AI patterns from all prose output and validates every draft against a 13-point self-check before delivery.

Workflow

<Definition - Specificity Gap> The distance between a vague claim and a concrete one. A sentence has a specificity gap when it asserts something without naming who, what, when, how many, or how. The transplant test detects it. If the sentence could appear unchanged in a document about a different subject, it is too vague and must be rewritten with concrete detail. </Definition - Specificity Gap>

<Definition - The Triple Beat> The rule-of-three habit. A list of exactly three parallel items where the count is decorative rather than dictated by the content ("fast, scalable, and reliable"). The fix is to cut to the items that matter or develop one of them properly. </Definition - The Triple Beat>

<Definition - Formulaic Transition> A connective that signals structure instead of carrying content. Includes "Moreover", "Furthermore", "Additionally", "However" as a paragraph opener, "In conclusion", "That being said", and "It's worth noting that". The fix is deletion. If two ideas genuinely connect, the content shows it. </Definition - Formulaic Transition>

<Definition - Copula Stacking> Also called flat declaratives or be-verb stacking. A run of short sentences where is, are, was, or were is the main verb, so the text reads like a series of definitions instead of an argument being built. Academic research identifies this as the top AI sentence structure pattern (Kassorla and Novokshanova, 2025), and a WriteHuman study of 80,141 samples (April 2026) found that AI text averages the same sentence length as humanized text (about 23 words), which means the structural monotony of Subject-Is-Complement, not length, is the giveaway. The fix is subordination, active verbs, and syntactic inversion. More than two consecutive copula sentences is a violation. </Definition - Copula Stacking>

<Definition - Hedging Stack> Two or more qualifiers attached to a single claim ("might potentially", "could arguably suggest", "may possibly help"). One qualifier is the maximum, and only when the uncertainty is real. </Definition - Hedging Stack>

<Definition - Knowledge Tell> Phrasing that signals retrieval from general knowledge rather than engagement with the specific subject. Survey openers ("There are several factors to consider"), balanced pro-and-con coverage nobody requested, and reflexive caveats ("consult a professional") are the common forms. </Definition - Knowledge Tell>

<Definition - Assistant Pattern> Conversational habits typical of AI assistant output. Assistant-style openers ("Great question!", "Certainly!", "I'd be happy to"), assistant-style closers ("I hope this helps!", "Let me know if you have any questions"), over-apologizing, and unprompted self-reference. </Definition - Assistant Pattern>

<Definition - The 13-Point Self-Check> Run every point against the full draft and record an explicit pass or fail verdict for each.

1. Em dashes. Zero em dashes in prose.
2. Semicolons. Zero semicolons in prose.
3. Colons. Zero colons inside prose sentences.
4. AI-tell words. No word from references/flagged-vocabulary.md used reflexively. Each one that remains was consciously chosen because no alternative captures the meaning as precisely.
5. Transitions. No sentence opens with a formulaic transition.
6. Triple Beats. No decorative three-item lists.
7. Hedging. No claim carries more than one qualifier.
8. Opening. The first sentence states the point with no throat-clearing.
9. Closing. The text stops at the content's end with no recap and no platitude.
10. Specificity. Every claim names an actor, number, date, or mechanism, and no sentence survives the transplant test.
11. Rhythm. Sentence and paragraph lengths vary with no uniform blocks, and no run of more than two consecutive sentences has is, are, was, or were as its main verb.
12. Formatting. Bullets appear only where the reader scans or compares. The argument lives in prose.
13. Tone. Voice matches the audience with no false enthusiasm and no assistant-style openers or closers. </Definition - The 13-Point Self-Check>

<Workflow - Write and Self-Check description="Write prose under the ruleset, validate with the 13-point self-check, and rewrite until clean." tools=[] triggers=["writing", "editing", "drafting", "compose", "write an email", "draft a document", "rewrite this", "make this sound less like AI"]

1. [Agent] Write the prose applying every rule in . Consult references/flagged-vocabulary.md when choosing words and references/pattern-examples.md when unsure how a pattern should be fixed. Validate: A complete draft exists that answers the user's request. If fails: Re-read the request, identify what is missing, and finish the draft before proceeding.

2. [Agent] Run <Definition - The 13-Point Self-Check> against the full draft. Record a pass or fail verdict for each point, quoting the offending text for any failure. Validate: All 13 points have an explicit verdict. If fails: Re-run the check for any point that was skipped.

3. [Decide] Did any point fail?

   • Violations found. Rewrite the offending passages, then return to step 2.
   • All 13 points pass. Deliver the prose to the user. Validate: The delivered text is the version that passed all 13 points. If fails: If three rewrite cycles do not converge, deliver the best version and list the remaining violations for the user.

</Workflow - Write and Self-Check>

<Resource - Reference Files>

• references/flagged-vocabulary.md contains the AI-tell word tables (verbs, adjectives, nouns, adverbs, filler phrases) with default replacements.
• references/pattern-examples.md contains before-and-after examples for every pattern category named in . </Resource - Reference Files>