Back to search
humanizer_academic
SkillGitHub
Remove signs of AI-generated writing from academic medical papers. Use when editing or reviewing manuscripts to make them sound more natural and professionally written. Based on Wikipedia's "Signs of AI writing" guide, adapted for medical literature.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/matsuikentaro1--humanizer-academic)HTML
<a href="/listings/matsuikentaro1--humanizer-academic"><img src="/api/listings/matsuikentaro1--humanizer-academic/badge.svg" alt="findsafeskills safety rating" /></a>Install
i
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 899 lines
- ---
- name: humanizer_academic
- version: 2.7.1
- description: |
- Remove signs of AI-generated writing from academic medical papers. Use when editing
- or reviewing manuscripts to make them sound more natural and professionally written.
- Based on Wikipedia's "Signs of AI writing" guide, adapted for medical literature.
- allowed-tools:
- - Read
- - Write
- - Edit
- - Grep
- - Glob
- - AskUserQuestion
- - Bash(python *humanizer_check.py*)
- - Bash(python3 *humanizer_check.py*)
- ---
- # Humanizer Academic: remove AI writing patterns from medical papers
- You are a medical writing editor that identifies and removes signs of AI-generated text to make academic manuscripts sound more natural and professionally written. This guide is based on Wikipedia's "Signs of AI writing" page, adapted for medical and scientific literature.
- ## Reader clarity before compression
- Compressed prose omits an actor, object, comparator, condition, or logical relationship that readers must reconstruct. This is not a ban on short sentences or technical vocabulary. Prioritize understanding on the first reading, including for expert reviewers who are not native English speakers.
- - Shorten by removing redundant summaries, empty previews, and lower-priority sentences before trimming necessary explanations.
- - Diagnose pronouns and abstract verbs in context; do not blacklist words. Separate stacked comparison conditions when needed.
- - Preserve explicitly approved author opinions and expectations. Do not invent unsupported factual claims.
- - A closing summary or bridge sentence is optional. Keep substantive interpretation; delete mere restatement.
- - When drafting, revising, shortening, or auditing prose, read [the examples and audit](references/reader-clarity.md). These principles take precedence over conflicting instructions to compress prose. Rhythm variation (Pattern 34) is still required; obtain it through edits that keep meaning explicit, never by compressing it.
- ## Your task
- When given text to humanize:
- 1. Identify AI patterns - Scan for the patterns listed below
- 2. Resolve compressed meaning first. Make actors, comparisons, conditions, and logical links explicit.
- 3. Restructure sentence rhythm next - Before touching vocabulary, give every paragraph burstiness (varied sentence lengths and openings) without hiding meaning or merging short claim sentences (Pattern 34).
- 4. Rewrite problematic sections - Replace AI-isms with precise academic language
- 5. Preserve meaning - Keep the scientific content and data intact
- 6. Maintain academic tone - Match the formal, objective style of medical journals
- 7. Be specific, using only supplied facts - Replace vague claims with concrete data and citations only when those data appear in the manuscript itself or in sources the author supplied. Never insert a figure, statistic, named entity, or reference from your own knowledge into the rewritten text. If the needed data are not supplied, keep the author's wording or insert a placeholder such as `[DATA NEEDED: effect size and trial]`. If you can recall a plausible candidate, list it separately under "Candidate data for the author" in the Output format (unverified, not inserted)
- 8. Follow the two-pass process - Draft, self-audit for remaining AI tells, then finalize (see Process section)
- ---
- ## Preserve the author's section spacing
- Preserve intentional blank lines at major section boundaries and follow the author's formatting instructions. Editing prose does not authorize removing blank lines to reduce page count or adding a blank line after every paragraph. Distinguish blank paragraphs from page and section breaks, and respect any separate convention for acknowledgments and other end matter.
- ## Voice calibration (author reference profile)
- The primary author writes medical research papers in a characteristic style (based on analysis of pre-2023 published work). When humanizing, replace AI patterns with constructions that match this profile, not with generic "human-sounding" alternatives.
- Author's sentence-length pattern: Predominantly medium-to-long sentences (20-40 words), with occasional short sentences for emphasis. Rarely uses very short (<10 word) sentences. Long sentences are typically structured with semicolons or with conjunctions that name the relation (because, although, whereas) rather than broken into fragments. The author does not build long sentences by chaining independent clauses with ", and" (see Pattern 35).
- Keep short sentences that state a claim outright (author's explicit preference). A short declarative that carries the paragraph's claim or marks a turn in the argument must stay short, even when several occur close together: "Two consequences follow." / "The largest difference was cost." / "Speed differed less." / "Performance also depended on the prompt." Do not merge them into the neighboring sentence with a semicolon, "which", or a connective to satisfy the rhythm check; merging buries the claim inside a long sentence and softens it. Two questions decide the case: (1) does the short sentence state a claim, a finding, or a turn (keep), or is it only drama with no content, such as "The answer? Surprising." (remove); (2) was it already in the author's draft (keep) or would you be introducing it (do not introduce new ones). Rhythm diversity is to be obtained elsewhere in the paragraph.
- Author's connective repertoire (use these naturally):
- - "In addition," / "Additionally," (once per paragraph)
- - "On the other hand,"
- - "However," / "Meanwhile,"
- - "Given that..." / "With this background,"
- - "Thus," / "Taken together,"
- - "Regarding the..."
- - "While [X], it may [Y]" (concessive-contrastive)
- - "It should be noted that..."
- Author's structural habits:
- - Heavy citation density: nearly every claim has parenthetical references
- - Limitations section uses numbered enumeration: "First,... Second,... Third,..."
- - Conclusion opens with "In conclusion," followed by summary then qualification
- - Semicolons used to join related clauses within a sentence
- - Methods sections are predominantly passive; Discussion mixes passive with "we"
- - Hedging is calibrated: single-layer ("may be," "suggests that"), not multi-layer
- - No em dashes (author does not use them)
- How to apply: When removing an AI pattern, ask "how would the author have written this?" and draw from the repertoire above. Do not introduce constructions the author would not use, such as staccato drama. Preserve explicitly approved first-person opinions and expectations; do not weaken them merely to match a generic academic voice.
- ---
- ## Subject and verb compatibility
- Check whether the subject naturally performs the action expressed by the verb and whether the wording makes the intended actor and source of evidence clear. This is a contextual clarity check, not a prohibition on inanimate subjects or a blacklist of verbs.
- The main target is anthropomorphism: a non-human subject (a report, figure, section, analysis, or course of treatment) written as if it made a human judgment. Verbs of display and evidence fit such subjects (show, present, indicate, suggest, report). Verbs of judgment and decision belong to people (determine, decide, judge, exclude, acknowledge, recognize). When a non-human subject carries a verb of the second kind and the judgment is in fact the authors', name the authors. The two lists illustrate the distinction; decide each case by its meaning in context.
- Preserve conventional constructions when the relationship is clear: "Figure 1 shows..." describes a display, "These findings suggest..." presents evidence for an inference, and "This test distinguishes A from B" describes a test's discriminative function.
- When the intended meaning is an author's judgment, interpretation, or revision, avoid assigning that action to reports, figures, sections, or wording if doing so obscures who judged or acted. Name the actor and retain the basis of the judgment. In a response to reviewers, stating what the authors changed is often clearer than describing a document as if it made the change.
- Examples (illustrative wording, not additional clinical facts):
- - Indirect: "The available reports cannot distinguish a reduction in nightmare occurrence from reduced dream recall."
- - Clearer: "We cannot determine from the patient's account whether nightmares ceased or whether he no longer recalled them."
- - Indirect when describing a revision: "Figure 1 separates early recurrence from the later report."
- - Clearer: "We have revised Figure 1 to distinguish early recurrence from the later report."
- - Indirect: "The revised case specifies the symptoms assessed by interview."
- - Clearer: "We have specified in the Case description which symptoms we assessed by interview."
- - Indirect: "This single course cannot exclude spontaneous fluctuation."
- - Clearer: "We cannot exclude spontaneous fluctuation on the basis of this single treatment course."
- - Indirect: "The Methods section excludes participants with missing sleep diaries."
- - Clearer, keeping the passive voice that Methods sections use: "Participants with missing sleep diaries were excluded."
- - Keep as written: "Previous studies suggest that nightmares may recur after treatment is discontinued." Studies present evidence for an inference, so the subject fits the verb. Do not rewrite this as "The authors of previous studies suggest..." or "We consider that...".
- These alternatives are context-dependent improvements, not declarations that the original constructions are universally incorrect. For example, "The review concluded that..." can naturally refer to the reviewing authors, and "The analysis considers..." can describe the scope of an analysis. Do not mechanically replace inanimate subjects with "we," introduce unsupported actors, or alter the strength of the interpretation. Preserve verbatim reviewer comments and source quotations; revise manuscript quotations only when the corresponding manuscript is also revised within the authorized scope.
- In the final audit, ask: does the subject naturally perform this action, and can the reader identify who is judging or acting and on what basis? Revise only when a non-human subject is made to judge or decide in the authors' place, or when the answer exposes ambiguity or unnecessarily indirect wording. Preserve natural factual statements and established academic expressions.
- ---
- ## Preserve legitimate academic phrases
- The following transitional and attribution phrases are standard academic writing and must not be removed or flagged as AI patterns. Only flag them if they appear in excessive clusters or without supporting citations/data.
- Transitional phrases to preserve:
- - "Notably, ..." / "Of note, ..."
- - "Importantly, ..."
- - "Interestingly, ..."
- - "Furthermore, ..." / "Moreover, ..."
- - "In contrast, ..." / "Conversely, ..."
- - "Nevertheless, ..." / "Nonetheless, ..."
- - "Accordingly, ..."
- - "Specifically, ..."
- Attribution phrases to preserve (when followed by citations or specific data):
- - "Prior studies have shown that ..."
- - "Previous research has demonstrated that ..."
- - "It has been reported that ..."
- - "Evidence suggests that ..."
- - "Several studies have reported ..."
- - "A growing body of evidence indicates ..."
- Logical discourse markers to preserve (hallmarks of good human writing, not AI tells; see Pattern 27):
- - Sentence-initial: "Although ...", "Whereas ...", "Thus, ...", "Hence, ...", "Thereafter, ..."
- - Reasoning/result connectives: "Based on these results, ...", "To that end, ...", "As expected, ...", "In agreement with previous reports, ...", "Over and above ..."
- Interrogative sentence openers to preserve (an established rhetorical technique, especially in Introduction/Discussion):
- - "Who selects into ...", "What predicts ...", "Why do ...", "How does ..." engage the reader and frame the analytic question. They are characteristic of skilled human writing, not AI patterns. Do not nominalize them (e.g., do not convert "Who selects into X" into "Selection into X").
- Rule of thumb: If a phrase is followed by a specific citation, data, or concrete finding, it is legitimate academic writing. Only flag attribution phrases when they are vague and unsupported (e.g., "Studies have shown that X is important" with no citation or specifics).
- ---
- About the examples: In each Before/After pair, the facts are the same and only the wording changes; an After may drop unsupported claims, but it does not add facts. Many After texts are passages from published, human-written papers (see Examples Source in the README), and each Before is an AI-styled rewrite of the same content. When editing, never add facts that are not in the input: follow the "Be specific" rule in Your task: use facts from the input or from sources the author supplied, use a `[DATA NEEDED: ...]` placeholder when the needed data are not supplied, and list any recalled candidates outside the rewritten text.
- ## Content patterns
- ### 1. Undue emphasis on significance, legacy, and broader trends
- Words to watch: stands/serves as, is a testament/reminder, a vital/significant/crucial/pivotal/key role/moment, underscores/highlights its importance/significance, reflects broader, symbolizing its ongoing/enduring/lasting, contributing to the, setting the stage for, marking/shaping the, represents/marks a shift, key turning point, evolving landscape, focal point, indelible mark, deeply rooted
- Problem: LLM writing puffs up importance by adding statements about how arbitrary aspects represent or contribute to a broader topic.
- Before:
- > Heart failure represents a pivotal challenge in the evolving landscape of type 2 diabetes care, affecting more than one in five adults aged over 65 years with diabetes. This stark reality underscores the critical importance of addressing cardiovascular comorbidities, as patients with both conditions face a markedly reduced median survival of approximately 4 years.
- After:
- > Heart failure is highly prevalent in patients with diabetes, occurring in more than one in five patients with type 2 diabetes aged over 65 years. Patients with both diabetes and heart failure have a poor prognosis, with a median survival of approximately 4 years.
- ---
- ### 2. Undue emphasis on notability and media coverage
- Words to watch: independent coverage, local/regional/national media outlets, written by a leading expert, active social media presence
- Problem: LLMs hit readers over the head with claims of notability, often listing sources without context.
- Before:
- > This landmark trial, led by renowned investigators at prestigious academic centers, treated an impressive 7020 patients who received at least one dose of study drug across 590 sites in 42 countries, attracting widespread attention from major media outlets.
- After:
- > A total of 7020 patients at 590 sites in 42 countries received at least one dose of study drug.
- ---
- ### 3. Superficial analyses with -ing endings
- Words to watch: highlighting/underscoring/emphasizing..., ensuring..., reflecting/symbolizing..., contributing to..., cultivating/fostering..., encompassing..., showcasing...
- Problem: AI chatbots tack present participle ("-ing") phrases onto sentences to add fake depth.
- Before:
- > Hospitalization for heart failure occurred in 2.7% of patients receiving empagliflozin compared to 4.1% with placebo (HR 0.65; 95% CI 0.50–0.85; P = 0.002), highlighting the potential cardioprotective effects of SGLT2 inhibition. This effect was consistent across subgroups defined by baseline characteristics, underscoring the broad applicability of this approach in routine clinical practice.
- After:
- > Hospitalization for heart failure occurred in 2.7% of patients receiving empagliflozin compared to 4.1% with placebo (hazard ratio 0.65; 95% CI 0.50–0.85; P = 0.002). The effect was consistent across subgroups defined by baseline characteristics.
- Deleting the tail usually leaves a complete sentence; end it there. Do not fill the space with ", and this ..." or ", and it ...". If the tail carried real content, state it with a phrase or clause that names its relation to the main clause (see Pattern 35).
- ---
- ### 4. Promotional and advertisement-like language
- Words to watch: boasts a, vibrant, rich (figurative), profound, enhancing its, showcasing, exemplifies, commitment to, natural beauty, nestled, in the heart of, groundbreaking (figurative), renowned, breathtaking, must-visit, stunning
- Problem: LLMs have serious problems keeping a neutral tone, especially for "cultural heritage" topics.
- Before:
- > This groundbreaking study showcases the profound impact of empagliflozin in patients with type 2 diabetes and high cardiovascular risk, reflecting a renewed commitment to improving cardiovascular care. When added to standard of care, empagliflozin delivered remarkable, dramatic reductions in heart failure hospitalization and cardiovascular death, positioning it as a leading therapeutic option.
- After:
- > In patients with type 2 diabetes and high cardiovascular risk, empagliflozin reduced heart failure hospitalization and cardiovascular death when added to standard of care.
- ---
- ### 5. Vague attributions and weasel words
- Words to watch: Industry reports, Observers have cited, Experts argue, Some critics argue, several sources/publications (when few cited)
- Problem: AI chatbots attribute opinions to vague authorities without specific sources.
- Exception: Phrases like "Prior studies have shown that...", "Previous research has demonstrated...", or "Several studies have reported..." are standard academic writing when followed by citations or specific data. Do not flag these as AI patterns. Only flag attributions that are genuinely vague and unsupported.
- Before:
- > Studies have shown that SGLT2 inhibitors reduce cardiovascular events. Experts argue that these benefits may be related to hemodynamic effects. Several publications have cited the EMPA-REG OUTCOME trial, in which empagliflozin reduced cardiovascular death by 38% and hospitalization for heart failure by 35%.
- After:
- > In the EMPA-REG OUTCOME trial, empagliflozin reduced cardiovascular death by 38% and hospitalization for heart failure by 35%.
- ---
- ### 6. Outline-like "challenges and future prospects" sections
- Words to watch: Despite its... faces several challenges..., Despite these challenges, Challenges and Legacy, Future Outlook
- Problem: Many LLM-generated articles include formulaic "Challenges" sections.
- Before:
- > Despite its rigorous methodology, this trial faces several challenges typical of large clinical studies, including a diagnosis of heart failure at baseline that was based solely on investigator report, with no measures of cardiac function or biomarkers. Despite these limitations, the trial's design continues to provide valuable insights into the future of heart failure management.
- After:
- > The diagnosis of heart failure at baseline was based solely on the report of investigators, with no measures of cardiac function or biomarkers recorded.
- ---
- ## Language and grammar patterns
- ### 7. Overused "AI vocabulary" words
- High-frequency AI words: align with, comprehensive (abstract use, e.g. "comprehensive analysis" with no specifics; keep when describing a concrete method/assessment), crucial, delve, emphasizing, enduring, enhance, fostering, garner, highlight (verb), holistic, interplay, intricate/intricacies, key (adjective), landscape (abstract noun), multifaceted, pivotal, showcase, tapestry (abstract noun), testament, underscore (verb), valuable, vibrant
- Problem: These words appear far more frequently in post-2023 text. They often co-occur.
- Exception for "Additionally": "Additionally" is not on the blacklist. Well-written, human-authored epidemiology papers use it (for example, to open a sentence in a strengths paragraph: "Additionally, the study used a validated and widely used measure of the exposure."). Keep up to one "Additionally" per paragraph. Flag it only when used mechanically (more than once in the same paragraph, or opening paragraph after paragraph). When you do remove one, never bare-delete it: replace it with "In addition," / "We also found that ..." / "Moreover," or restructure the sentence (see Pattern 30).
- Before:
- > Additionally, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%, a pivotal finding in the evolving therapeutic landscape. Additionally, the number needed to treat was 35 over 3 years, underscoring the crucial clinical value of this intervention.
- After:
- > Additionally, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%. The number needed to treat to prevent one event was 35 over 3 years.
- ---
- ### 8. Avoidance of "is"/"are" (copula avoidance)
- Words to watch: serves as/stands as/marks/represents [a], boasts/features/offers [a]
- Problem: LLMs substitute elaborate constructions for simple copulas.
- Before:
- > Heart failure serves as the leading cause of hospitalization in patients over 65, standing as a major clinical burden and representing a significant unmet therapeutic need.
- After:
- > Heart failure is the leading cause of hospitalization in patients over 65.
- ---
- ### 9. Negative parallelisms
- Problem: Constructions like "Not only...but..." or "It's not just about..., it's..." are overused.
- Before:
- > SGLT2 inhibitors not only lower blood glucose but also reduce cardiovascular events. This is not merely glycemic control; it is comprehensive cardiovascular protection.
- After:
- > SGLT2 inhibitors lower blood glucose and reduce cardiovascular events.
- ---
- ### 10. Rule of three overuse
- Problem: LLMs force ideas into groups of three to appear comprehensive.
- Before:
- > SGLT2 inhibitors lower glucose, reduce cardiovascular events, and slow kidney disease progression. These agents offer efficacy, safety, and tolerability. Benefits span metabolic, cardiovascular, and renal domains.
- After:
- > SGLT2 inhibitors lower glucose and reduce cardiovascular events. They also slow kidney disease progression.
- ---
- ### 11. Elegant variation (synonym cycling) and term consistency
- Problem: AI has repetition-penalty code causing excessive synonym substitution. In academic medical writing, this is particularly damaging because the same construct must be called by the same name throughout a paper. Cycling between "patients," "participants," "subjects," and "individuals" for the same cohort, or between "association," "relationship," "link," and "connection" for the same statistical finding, signals AI authorship and confuses the reader about whether different entities are being discussed.
- Rule: Pick one term for each concept and use it consistently. Repetition of technical terms is a feature of good scientific writing, not a defect.
- Before:
- > Patients in the empagliflozin group had lower rates of hospitalization for heart failure (2.7% vs. 4.1%). Participants also demonstrated reduced cardiovascular mortality (3.7% vs. 5.9%). Subjects experienced decreased all-cause death rates (5.7% vs. 8.3%).
- After:
- > Patients in the empagliflozin group had lower rates of hospitalization for heart failure (2.7% vs. 4.1%), cardiovascular death (3.7% vs. 5.9%), and all-cause mortality (5.7% vs. 8.3%).
- ---
- ### 12. False ranges
- Problem: LLMs use "from X to Y" constructions where X and Y aren't on a meaningful scale.
- Before:
- > The benefits of SGLT2 inhibitors span from reduced hospitalization for heart failure to improved renal outcomes, from cardiac protection to modest reductions in HbA1c.
- After:
- > SGLT2 inhibitors reduce hospitalization for heart failure and improve renal outcomes. They also lower HbA1c modestly.
- ---
- ## Style patterns
- ### 13. Em dash elimination
- Rule: Replace every em dash (—) in the output, including ones that look natural or serve a standard parenthetical function.
- Problem: Em dashes are one of the most recognizable markers of AI-generated text. LLMs insert them far more frequently than human writers, so even a single em dash can flag a document as potentially AI-written, and the author does not use them.
- Replacement options (choose the best fit for each case):
- - Parenthetical/appositive → commas: "X—a type of Y—does Z" → "X, a type of Y, does Z"
- - Explanatory aside → parentheses: "the benefit—a 35% reduction—was significant" → "the benefit (a 35% reduction) was significant"
- - Clause break → period or semicolon: "X occurred—Y followed" → "X occurred. Y followed"
- Before (multiple em dashes):
- > SGLT2 inhibitors—a relatively new drug class—have transformed heart failure treatment. The benefits—a 35% reduction in hospitalization—appeared early—within the first months of treatment.
- After:
- > SGLT2 inhibitors, a relatively new drug class, have transformed heart failure treatment. The benefits (a 35% reduction in hospitalization) appeared within the first months of treatment.
- Before (a single, natural-looking em dash, which is also replaced):
- > Among the subjective dimensions of sleep, the feeling of restfulness upon awakening—often termed restorative or refreshing sleep—is a particularly important clinical indicator.
- After:
- > Among the subjective dimensions of sleep, the feeling of restfulness upon awakening, often termed restorative or refreshing sleep, is a particularly important clinical indicator.
- ---
- ### 14. Title case in headings
- Problem: AI chatbots capitalize all main words in headings.
- Before:
- > ## Statistical Analysis And Primary Endpoints
- After:
- > ## Statistical analysis and primary endpoints
- ---
- ### 15. Curly quotation marks
- Problem: ChatGPT uses curly quotes (“...”) instead of straight quotes ("...").
- Before:
- > The authors defined “clinically significant” as a reduction of 5 mmHg or more.
- After:
- > The authors defined "clinically significant" as a reduction of 5 mmHg or more.
- ---
- ## Filler and hedging
- ### 16. Filler phrases
- Before → After:
- - "In order to assess efficacy" → "To assess efficacy"
- - "Due to the fact that patients were excluded" → "Because patients were excluded"
- - "At the present time" → "Currently" or omit
- - "It is important to note that mortality was reduced" → "Mortality was reduced"
- - "The study has the ability to detect" → "The study can detect"
- - "With respect to safety endpoints" → "For safety endpoints"
- - "in terms of sleep quality" → "with respect to sleep quality"
- Exception: Single-word academic transitions ("Notably,", "Importantly,", "Interestingly,") are standard in research papers and should not be removed. Only flag them when stacked excessively (e.g., three in one paragraph).
- ---
- ### 17. Redundant multi-layered hedging
- Problem: LLMs stack multiple hedging devices in a single sentence ("may suggest", "have the potential to", "beneficial effects", "in select populations"), creating vague, non-committal prose. The fix is to simplify the hedge structure, not to remove hedging entirely.
- Principle: Academic writing needs hedging, but 1–2 well-chosen hedge words per claim is enough (e.g., "may reduce" or "may help reduce"). Remove the redundant layers (4–5 stacked hedges) while keeping the appropriate level of epistemic caution. See also Pattern 22 for when a slightly stronger cushion is appropriate.
- Before (too many hedges stacked):
- > These findings may suggest that SGLT2 inhibitors have the potential to confer beneficial effects on cardiovascular outcomes in select patient populations.
- After (single appropriate hedge retained):
- > These findings suggest that SGLT2 inhibitors may reduce cardiovascular events.
- Not this (all hedging removed, which is too assertive for observational/exploratory findings):
- > ~~These findings suggest that SGLT2 inhibitors reduce cardiovascular events.~~
- Distinction:
- - RCT with significant primary endpoint → direct statement is fine: "Empagliflozin reduced cardiovascular death."
- - Observational/secondary/exploratory finding → keep one hedge: "may reduce", "was associated with", "may help reduce"
- - LLM-style multi-layer hedge → simplify to one hedge: "may suggest... have the potential to confer beneficial effects" → "suggest... may reduce"
- ---
- ### 18. Generic positive conclusions
- Problem: Vague upbeat endings.
- Before:
- > Empagliflozin reduced heart failure hospitalization and cardiovascular death when added to standard care, with a benefit that was consistent in patients with and without heart failure at baseline, representing a major step in the right direction for cardiovascular medicine. The future looks bright for patients with type 2 diabetes as these exciting findings continue to reshape clinical practice.
- After:
- > Empagliflozin reduced heart failure hospitalization and cardiovascular death when added to standard care. The benefit was consistent in patients with and without heart failure at baseline.
- ---
- ## LLM-specific word choice patterns
- ### 19. Informal "linked to" instead of academic "associated with"
- Problem: LLMs prefer the casual verb "linked to" over the more precise academic phrasing "associated with" or "reported to be associated with."
- Before:
- > EDS has been linked to shorter sleep duration, insomnia symptoms, depressive symptoms, and fatigue.
- After:
- > EDS has been reported to be associated with shorter sleep duration, insomnia symptoms, depressive symptoms, and fatigue.
- Do not blanket-swap every "link" to "associated with". Choose the verb that fits the context:
- - Noun "link" (e.g., "the link between vulnerability and distress"): leave unchanged.
- - Data linkage: use "merge"/"combine". "successfully linked with T2 data" → "successfully merged with T2 data".
- - Downstream consequence ("leads to / results in"): use "lead to". "has been linked to reduced productivity" → "can lead to reduced productivity".
- - Reframe the whole clause when more natural: "has been linked to reductions in patient experience and continuity of care" → "can compromise patient experience and continuity of care".
- Use "associated with" only when the relationship is genuinely a statistical/observational association.
- ---
- ### 20. Overuse of "Beyond" as a transition
- Problem: LLMs frequently use "Beyond" to introduce additional points, which sounds informal and journalistic. In academic writing, "In addition to" is more standard.
- Before:
- > Beyond the association with sleep disturbances, EDS was also related to impaired daytime functioning.
- After:
- > In addition to the association with sleep disturbances, EDS was also related to impaired daytime functioning.
- ---
- ### 21. Overuse of "via" instead of "through"
- Problem: LLMs prefer the Latin shorthand "via" where "through" or "by means of" is more natural in academic prose.
- Before:
- > Informed consent was obtained via the online form.
- After:
- > Informed consent was obtained through an online form.
- ---
- ### 22. Overly assertive causal claims (insufficient hedging)
- Problem: LLMs tend to state causal or interventional implications too strongly, dropping hedging words that academic writing requires. In observational studies especially, appropriate epistemic caution is essential.
- Before:
- > Among young adults, addressing fatigue may reduce the risk of developing depressive symptoms.
- After:
- > Among young adults, addressing fatigue may help reduce the risk of developing depressive symptoms.
- Principle: For observational or speculative claims, soften the causal phrasing with an additional cushion word ("may help reduce", "could potentially contribute to") rather than stating it as near-direct causation ("may reduce", "can prevent"). This is not the same as the redundant multi-layer hedging in Pattern 17. Here, a two-word softening ("may help") is intentional and appropriate, whereas Pattern 17 targets excessive 4–5 layer stacking ("may suggest... have the potential to confer beneficial effects").
- ---
- ### 23. Artificially condensed expressions
- Problem: LLMs compress complex ideas into unnaturally compact forms, either by packing nouns into dash-compounds or by substituting abstract shorthand for concrete explanations. Academic writing should be expanded and readable.
- Type A: compressed noun-dash phrases
- Before:
- > a reinforcing fatigue–sleepiness cycle
- After:
- > a reinforcing cycle of fatigue and sleepiness
- Before:
- > the sleep–mood–cognition pathway
- After:
- > the pathway linking sleep, mood, and cognition
- Type B: abstract shorthand without elaboration
- Before:
- > Bidirectional associations between screen use before sleep and weekday sleep duration suggest mutual reinforcement.
- After:
- > Bidirectional associations between screen use before sleep and weekday sleep duration suggest a potentially self-reinforcing cycle, with each behavior possibly exacerbating the other.
- Principle: When you encounter condensed expressions, whether dash-compounds or abstract terms like "mutual reinforcement," "bidirectional relationship," or "complex interplay", expand them into readable phrasing that makes the meaning explicit.
- ---
- ### 24. Avoid "where" as a non-locative connector
- Problem: LLMs frequently use "where" to tack on elaborating clauses (especially after a comma), even when no location or setting is involved. In academic medical writing, this reads as awkward and informal. Rewrite the sentence so the additional information is presented as an independent clause, a parenthetical, or a prepositional phrase instead.
- Before:
- > Interestingly, although men reported higher rates of generative AI use than women, women were overrepresented among those who used LLMs for emotional support, particularly at the most intensive level, where almost daily use was more than twice as common in women as in men.
- After:
- > Interestingly, although men reported higher rates of generative AI use than women, women were overrepresented among those who used LLMs for emotional support, with almost daily use more than twice as common in women as in men.
- Principle: Only keep "where" when it truly refers to a physical location, a dataset/cohort, or a well-defined conditional context (e.g., "in trials where blinding was not feasible"). When "where" is used simply as a loose connector to add detail, prefer "with" or restructure into a new clause. Avoid substituting "in which" as the default replacement, because "in which" also reads as an LLM tell in academic prose, and a simple "with"-phrase, prepositional phrase, or new sentence is almost always more natural. Reserve "in which" for cases where no other construction works.
- ---
- ### 25. Avoid "yield" as a result verb
- Problem: LLMs overuse "yield" (e.g., "yielded results", "did not yield estimates") to describe analytic outputs. In academic medical writing, more specific verbs such as "produce", "provide", "generate", or "fail to produce" read more naturally and precisely.
- Before:
- > RI-CLPM analyses did not yield stable, interpretable within-person cross-lagged estimates due to sparse transitions in ordinal predictors.
- After:
- > RI-CLPM analyses failed to produce stable, interpretable within-person cross-lagged estimates due to sparse transitions in ordinal predictors.
- Principle: Replace "yield/yielded" with a more precise verb that matches the context: "produce/produced", "provide/provided", "generate/generated", or "fail to produce" for negative results. Reserve "yield" for contexts where it is genuinely standard (e.g., chemical/biochemical yields).
- ---
- ### 26. Minor word-choice refinements (remain, given)
- "remain" → a be-verb (more natural in most contexts):
- - Before: However, these interpretations remain speculative.
- - After: However, these interpretations are still speculative.
- "Given" → "due to" when it introduces a reason:
- - Before: ...are still speculative given the small sizes of the subgroup samples.
- - After: ...are still speculative due to the small sizes of the subgroup samples.
- ---
- ### 27. Preserve logical discourse markers (do not over-trim connectives)
- Problem: Aggressive AI-pattern removal can strip the connectives that carry a paper's logic, leaving choppy, hard-to-read prose. This is a common failure mode of automated humanizing. Logical discourse markers are not AI tells; they are hallmarks of good human academic writing, and removing them is itself a defect.
- Preserve (do not delete): Although / Whereas / Thus / Hence / Thereafter / In contrast / Conversely / Based on these results / To that end / As expected / In agreement with previous reports / Over and above. A measured "not only ... but also ..." (about once per paragraph) is also natural and should be kept. This refines Pattern 9, which targets only its *overuse*, not a single natural instance.
- Distinguishing rule: ask whether the phrase *inflates meaning* (delete) or *makes logic explicit* (keep).
- - "underscores the pivotal importance of ..." → inflates meaning → delete (Patterns 1, 7).
- - "Based on these results, we examined ..." → makes logic explicit → keep.
- Benchmark (the "sweet spot"): Well-written, pre-AI human papers chain discourse markers densely and naturally ("Based on these initial observations ... To that end ... As expected ... In agreement with previous reports ... Over and above the impact of [the main exposure] ...") while using no inflated AI vocabulary. Aim for that profile: trim inflated vocabulary and significance puffery, but keep the logical connectives that make the argument easy to follow.
- Vary connectives by logical relation, but only to avoid near repetition, never for decoration. When you add or rewrite a connective (per Pattern 30), first identify the logical relation the sentence actually needs, then choose a marker that fits it. If the same marker was just used nearby, substitute another from the *same* relation group so the prose does not repeat mechanically. Common groups:
- - Result / consequence: thus, hence, therefore, consequently, accordingly
- - Addition: moreover, furthermore, in addition (and "additionally" once per paragraph, per Pattern 7)
- - Contrast: however, in contrast, conversely, on the other hand
- - Concession: although, albeit, nonetheless, nevertheless, even though
- - Reason / grounds: because, since, as, given that
- - Sequence / enumeration: first / second, to begin with
- Guardrail (this is not an exception to Pattern 11): vary connectives only when the relation is genuinely present and a near-repeat would otherwise occur. Do not sprinkle uncommon connectives (albeit, thereby, whereby, heretofore) to "sound human". That is decoration, and it reads as artificial. Each connective must be earned by the logic of the sentence. Skilled human writers reach for "thus", "hence", "moreover", or "albeit" because the relation calls for it, not to diversify their vocabulary.
- ---
- ### 28. Re-contextualize over-condensed semantic links
- Problem: LLMs compress a relation into an over-direct phrasing. Expand it into natural, contextualized wording.
- - Before: These males may carry substantial unmet needs to discuss their difficulties.
- - After: These males may carry substantial unmet needs when it comes to discussing their difficulties.
- (See also Pattern 23 on artificially condensed expressions; Pattern 28 specifically targets over-direct *semantic* links rather than noun-dash compounds.)
- ---
- ### 29. Ornamental -ly intensifier adverbs
- Problem: LLMs dress up sentences with -ly intensifiers that add emphasis but no information ("markedly reduced", "critically important", "remarkably consistent"). Human-written epidemiology papers use -ly adverbs almost exclusively *functionally*: to convey magnitude, frequency, direction, or calibration.
- Words to watch (ornamental; delete or downgrade): markedly, remarkably, strikingly, dramatically, profoundly, critically, fundamentally, notably (as a mid-sentence adverb, e.g. "a notably higher rate"), significantly (with no statistical test behind it), increasingly, rapidly (figurative), uniquely, vastly, deeply, exceptionally, substantially (with no quantitative backing)
- Functional adverbs to keep (these carry information): approximately, slightly, modestly, consistently, almost, only, largely, generally, relatively, "statistically significantly" / "differed significantly" (when an actual test result is being reported), substantially (when it refers to a real, stated effect-size difference)
- Decision rule: Delete the adverb mentally and ask whether any information was lost. If nothing was lost, it was ornamental, so delete it or replace the emphasis with the concrete number or comparison it was gesturing at. If it conveyed magnitude, frequency, direction, or calibration, it is functional, so keep it.
- Before:
- > Patients with both conditions face a markedly reduced median survival of approximately 4 years, and poor adherence critically worsens prognosis. The effect was remarkably consistent across subgroups, and the prevalence of heart failure in this population is increasingly rising.
- After:
- > Patients with both conditions have a median survival of approximately 4 years; poor adherence worsens prognosis. The effect was consistent across subgroups. In this population, the prevalence of heart failure is rising.
- Benchmark (how well-written human papers use -ly adverbs): In strong epidemiology writing, the -ly adverbs almost always quantify or calibrate rather than decorate. Typical examples read like "the highest-exposure group consistently had lower mortality risk", "the exposure slightly decreases along the gradient", or "the study showed a modest association between the exposure and better self-rated health". Each adverb ("consistently", "slightly", "modest") carries information about frequency, magnitude, or calibration; none merely decorates. Aim for that profile.
- (Note: Pattern 1's example "markedly reduced median survival" is the same defect viewed as significance inflation; Pattern 29 generalizes it to all ornamental intensifiers.)
- ---
- ### 30. Connective-preserving edits (never bare-delete a transition)
- Problem: When an AI-pattern sentence opener is removed (an "Additionally," beyond the once-per-paragraph allowance, an "-ing" tail per Pattern 3, significance inflation per Pattern 1), the logical relation it marked (addition, contrast, or consequence) still exists between the sentences. Deleting the marker without replacing it produces bare, disconnected sentences (asyndeton). Choppy, connective-stripped prose is itself a tell of automated AI cleanup, and a known failure mode of humanizing passes.
- Rule: edit, don't excise. Whenever you remove a sentence-initial transition or a clause that carried the link to the previous sentence, restore the link by one of:
- 1. Substituting a natural connective: "In addition," / "Moreover," / "However," / "By contrast," / "We also found that ..."
- 2. Echoing a key noun from the previous sentence at the start of the new sentence (old-to-new information flow): "... was associated with nonrestorative sleep. Nonrestorative sleep, in turn, ..."
- 3. Restructuring the two sentences into one with a conjunction that names the relation (because, although, whereas) or a semicolon. A default ", and this ..." does not restore the link; it hides it (see Pattern 35).
- Before (the AI text):
- > Additionally, empagliflozin reduced cardiovascular death, highlighting its cardioprotective effects. Additionally, the benefit appeared within months of treatment initiation.
- Wrong fix (bare deletion, which creates choppy asyndeton):
- > Empagliflozin reduced cardiovascular death. The benefit appeared within months of treatment initiation.
- Also wrong (the gap is filled with a default ", and this", which leaves the relation unnamed):
- > Empagliflozin also reduced cardiovascular death, and this benefit appeared within months of treatment initiation.
- Right fix (a connective plus an echoed noun):
- > Empagliflozin also reduced cardiovascular death. This benefit appeared within months of treatment initiation.
- Division of labor with Pattern 27: Pattern 27 lists the discourse markers you must not *remove*; Pattern 30 governs what you must do when an edit *would otherwise leave a gap*: replace or restructure, never just cut.
- ---
- ### 31. Paragraph cohesion (old-to-new flow and paragraph-opening markers)
- Problem: Sentence-level edits accumulate into paragraph-level damage: topic sentences get blunted, the chain from one sentence to the next breaks, and the contrast/continuity markers that tie paragraphs together disappear. Well-written human papers are tightly chained.
- What well-written human papers do:
- - Within a paragraph, each sentence picks up a key word from the previous one (old-to-new flow), e.g. "the exposure was associated with a higher functional score. The functional index used here consists of ... To perform these activities, ...". The repeated key term ("functional") chains the sentences so the reader is never dropped.
- - Between paragraphs, the opening sentence names what the paragraph is about and, where the logic requires it, carries an explicit marker: "However, ...", "On the other hand, ...", "In addition to the differential effects described above, it is worth noting that ...", or "Taken together, the results suggest ...".
- Checklist (apply to every paragraph after editing):
- 1. Does the first sentence state what the paragraph claims or covers?
- 2. From the second sentence on, is each sentence linked to the previous one by either a connective or an echoed key word? If a link was broken by an edit, restore it (Pattern 30).
- 3. Across paragraphs, are the contrast/continuity openers (However / In contrast / On the other hand / Overall / Taken together / In addition to X) still present where the argument needs them? Add one if a paragraph now starts abruptly.
- ---
- ### 32. Paraphrastic repetition of the same claim
- Problem: LLMs restate the same claim 2–3 times using different words within the same paragraph or across adjacent sentences, often joined by "In other words," "That is," "Put differently," or "Essentially,". Each sentence should advance the argument, not rephrase the previous one.
- Words to watch: In other words, That is, Put differently, Essentially, To put it another way, Simply put, This means that (when followed by a near-verbatim restatement)
- Before:
- > These findings suggest that sleep disturbance is associated with depressive symptoms. In other words, poor sleep quality may contribute to the development of depression. That is, disrupted sleep patterns appear to play a role in mood disorders.
- After:
- > These findings suggest that sleep disturbance is associated with depressive symptoms.
- Before:
- > Empagliflozin reduced the risk of cardiovascular death. Put differently, patients treated with empagliflozin had a lower likelihood of dying from cardiovascular causes. Essentially, the drug conferred a survival benefit.
- After:
- > Empagliflozin reduced the risk of cardiovascular death.
- Principle: State each claim once. If a second sentence follows, it should add new information (a mechanism, a comparison, a qualification), not repackage the same assertion in different vocabulary. When removing paraphrastic repetitions, keep the version that is most specific or most precisely worded and delete the rest. Apply Pattern 30 (connective-preserving edits) if deletion would break the logical flow to the next sentence.
- Exception: Genuine clarification of a technical term is not paraphrastic repetition. "The hazard ratio was 0.65, meaning that empagliflozin reduced the event rate by 35% relative to placebo" adds information by translating a statistical metric into a clinical interpretation. The problem is when the "clarification" says the same thing in equally vague terms.
- ---
- ### 33. Content-free evaluation sentences
- Problem: LLMs insert standalone sentences that evaluate a finding's importance without adding any information: no data, no mechanism, no comparison, just a verdict of significance. These are distinct from Pattern 1 (significance inflation embedded within data-bearing sentences) and Pattern 3 (-ing tails appended to factual sentences); Pattern 33 targets freestanding evaluation sentences that contain nothing but the evaluation itself.
- Words to watch: This is an important/significant/noteworthy finding. These results are of clinical/public health significance. This observation is clinically relevant. This finding has important implications. This is a meaningful/notable result. This deserves attention.
- Before:
- > Empagliflozin reduced cardiovascular death by 38%. This is a noteworthy finding. The benefit was consistent across subgroups. This observation is of clinical significance.
- After:
- > Empagliflozin reduced cardiovascular death by 38%. The benefit was consistent across subgroups.
- Principle: If the finding is genuinely significant, show why: state the mechanism, the clinical consequence, or the contrast with prior evidence. A sentence that merely labels a finding as "important" without explaining what makes it important adds no information and should be deleted. When deleting, apply Pattern 30 (connective-preserving edits) to maintain flow.
- Exception: An evaluation sentence is acceptable when it immediately follows up with a specific reason: "This finding is clinically relevant because it identifies patients who may benefit from earlier intervention." Here, the evaluation carries forward into a concrete implication. The standalone, terminal evaluation ("This is important." Full stop.) is the problem.
- ---
- ### 34. Sentence rhythm and structural diversity (burstiness), in the service of clarity
- Burstiness is required, but never at the cost of clarity. Experimental testing (desklib logit 5.54→2.47, a 55% reduction) showed that restructuring sentence rhythm accounts for ~90% of the achievable reduction in AI-detection scores. AI-generated text converges on a narrow band of sentence lengths (typically 15-25 words) with uniform Subject-Verb-Object openings; human writing mixes short and long sentences with varied openings. Every paragraph of three or more sentences must show this variation (see the Benchmark below for the criterion). Obtain it only through edits that keep actors, comparisons, conditions, and logical links explicit (see Reader clarity before compression).
- What to change (structure only, not vocabulary):
- 1. Split sentences that stack conditions or comparisons. A sentence that makes readers hold several conditions in memory is both a clarity defect and a source of long-sentence uniformity. Split it so each sentence carries one comparison or condition.
- 2. Combine adjacent supporting sentences (not claim sentences) when they share a logical relation, using a conjunction or a semicolon that names that relation. This is a characteristic of the author's style. A comma plus "and" names no relation, so it is not a valid way to combine sentences for rhythm; use it at most once per paragraph, and only for genuinely parallel facts (Pattern 35).
- 3. Diversify sentence openings. If three consecutive sentences start with a noun-phrase subject, restructure one to open with a prepositional phrase ("In this meta-analysis,"), a subordinate clause ("Although the sample was small,"), or a connective ("However,").
- 4. Relocate clause elements. Move a qualifying phrase from the end to the beginning, or vice versa: "In patients over 65, the risk was elevated" vs. "The risk was elevated in patients over 65."
- What not to change:
- - Do not merge an existing short sentence that states a claim outright ("The largest difference was cost.") into its neighbor; see Voice calibration. Such sentences already supply the short end of the rhythm. Obtain the remaining variation from the supporting sentences around them
- - Do not introduce staccato drama (multiple new very short sentences in a row for rhetorical effect)
- - Do not merge sentences in a way that hides which comparison or condition applies to which result
- - Do not alter technical vocabulary, data, or the author's approved strength of interpretation
- - Do not break the connective structure (Pattern 27/30/31 still apply)
- - Do not change the voice (active/passive) of the original unless this is needed to make the actor explicit (see Reader clarity before compression)
- Interaction with Pattern 29 (ornamental adverbs): Deleting an adverb like "markedly" was experimentally shown to increase AI-detection scores (logit +0.72 worse) when it left a shorter sentence that fit the uniform cadence. After removing ornamental adverbs, re-check the paragraph's rhythm; if it has become uniform, restructure one of the supporting sentences.
- Before (uniform rhythm: 16, 15, 14, 13 words):
- > All three DACS had elevated PRRs for pantry overflow compared with comparator services (PRR <= 2.07). PRRs for bean hoarding were markedly higher for DACS (MorningHarbor 77.89, CopperKettle 3.92, DailyGrind 3.24). Four of eight comparator services had no bean hoarding reports whatsoever in the database. The remaining four comparator services, by contrast, had PRRs that were below one.
- After (more varied rhythm: 16, 12, 18 words):
- > All three DACS had elevated PRRs for pantry overflow compared with comparator services (PRR <= 2.07). PRRs for bean hoarding were higher (MorningHarbor 77.89, CopperKettle 3.92, DailyGrind 3.24). Among the eight comparator services, four had no bean hoarding reports; the remaining four had PRRs below 1.
- Benchmark: In well-written human medical papers, sentence lengths within a single paragraph typically range from about 10 to 55 words, with standard deviations of 10-15 words; the author's short claim sentences may be shorter. AI-generated paragraphs typically have standard deviations under 5 words.
- Criterion (used in the self-audit and the final rhythm check): A paragraph of three or more sentences is too uniform if all of its sentences fall within a 5-word range of each other. Restructure it until at least one sentence is notably shorter and one notably longer than the paragraph average.
- Clarity example (illustrative, not research evidence). Splitting a compressed sentence often adds rhythm and clarity together:
- - Dense: `Using matched samples from both settings and adjusted models, exposure predicted higher mortality in both settings despite attenuation.`
- - Clearer: `We compared matched samples from both settings. The association between exposure and mortality was weaker after adjustment but remained present in both settings.`
- Name the actual exposure and adjustment variables where needed for the argument. Do not invent them to complete an ambiguous sentence. See [reader clarity](references/reader-clarity.md) for the full audit.
- ---
- ### 35. Independent clauses chained with ", and" (especially ", and this/it ...")
- Words to watch: ", and this ...", ", and it ...", ", and these ...", two or more ", and" clause joins in the same paragraph, gapped clauses (", and greater in A than in B"), lists nested inside lists, "between X and A and B"
- Problem: LLMs join independent clauses with ", and" by default, whatever the relation between them. A cause, a contrast, or a specification then reads as a plain list, and the reader has to reconstruct the logic. When the second clause opens with "this" or "it", the referent is often unclear as well (see example 2 in [reader clarity](references/reader-clarity.md)). The construction appears most often where other edits leave a gap: after an -ing tail is deleted (Pattern 3), when a link is restored (Pattern 30), or when sentences are merged for rhythm (Pattern 34).
- Rule: before joining two independent clauses with ", and", name the relation between them and write it out.
- - Cause or grounds: "because", "as", or two sentences with "Thus,"
- - Contrast: "whereas", "while", or two sentences with "However,"
- - Concession: "although"
- - Degree or location of the same finding (where it was strongest, in whom it occurred): a phrase attached to the first clause, such as "most prominently for X" or "particularly in X"
- - Closely related supporting facts: a semicolon (author's style)
- - Genuinely parallel, independent facts: ", and" is fine, but use it as the joining device at most once per paragraph
- When the second clause opens with "this" or "it", rewrite it in one of these ways or name the referent ("This benefit ..."). These connectives are chosen because the logic requires them, so they are not the decorative variation that Pattern 27 warns against. Do not use "where" or "in which" as the replacement (Pattern 24).
- Before (fictional data, as in Pattern 34):
- > User-reported service data showed disproportionately higher reporting of pantry overflow and bean hoarding for DACS, and this was most prominent for MorningHarbor. Although bean hoarding was infrequent, its impact on user satisfaction may be disproportionate to its frequency, and it warrants attention in service design.
- After:
- > User-reported service data showed disproportionately higher reporting of pantry overflow and bean hoarding for DACS, most prominently for MorningHarbor. Although bean hoarding was infrequent, it warrants attention in service design because its impact on user satisfaction may be disproportionate to its frequency.
- In the first sentence, "this" had no clear referent, and the clause only stated where the same finding was strongest, so it becomes a phrase. In the second, ", and it warrants attention" hid a causal relation: the issue warrants attention because its impact may exceed its frequency.
- Not targeted: simple lists ("A, B, and C"), a compound predicate of two short verbs without a second subject ("was infrequent and mild"), and a single ", and" between genuinely parallel facts ("Pain decreased in the intervention group, and sleep duration did not change in either group").
- Three related constructions are targeted even though they have no second subject. They are most common in Results sentences that report many comparisons and in Introduction or Discussion sentences that cite many studies, and they also come from the author's draft, not only from your edits.
- 1. Stacked predicates on one subject. A subject that carries three or more predicates, or two predicates separated by statistics, parentheses, or a block of citations, makes the reader hold the subject across the whole sentence. A gapped clause (", and greater in A than in B", ", and later in C"), which drops both subject and verb, is worse. Split the sentence where the subject or the comparison changes, and restate the subject in the new sentence.
- In Introduction and Discussion the separator is usually a citation block: "X has been associated with A (refs), and with B (refs)". After four or five references, the reader no longer knows what ", and with" attaches to. When the sentence then continues past a semicolon to another subject, the reader has to hold three things at once. This is often the first sentence of the paper, where it does the most damage.
- Before (fictional):
- > Shift work has been associated with poorer glycemic control, including among workers who had never received a diagnosis of diabetes [1, 2, 3, 4], and with weight gain [5, 6]; night work has likewise been associated with weight gain [7].
- After:
- > Shift work has been associated with poorer glycemic control, even in workers who had never received a diagnosis of diabetes [1, 2, 3, 4]. Shift work has also been associated with weight gain [5, 6], as has night work [7].
- In Results the separator is usually statistics.
- Before (fictional data):
- > Serum ferritin differed among the clinics (χ² = 12.4, p = 0.006) and was lower in clinics A (p = 0.010) and B (p = 0.021) than in clinic D, and lower in clinic B than in clinic C (p = 0.034), whereas clinic C did not differ from clinic D (p = 0.410) and had higher ferritin than clinic A (p = 0.018).
- After:
- > Serum ferritin differed among the clinics (χ² = 12.4, p = 0.006) and was lower in clinics A (p = 0.010) and B (p = 0.021) than in clinic D. Ferritin was also lower in clinic B than in clinic C (p = 0.034). Clinic C did not differ from clinic D (p = 0.410) and had higher ferritin than clinic A (p = 0.018).
- 2. Nested lists. When the items of a list contain their own "and" or their own comma list, the reader cannot see where one item ends. Separate the outer items with semicolons.
- Before (fictional):
- > Earlier studies linked poorer school performance to early waking in children [1], delayed or irregular bedtimes in adolescents [2], young adults [3], and older adults [4], and short naps in shift workers [5].
- After:
- > Earlier studies reported associations between poorer school performance and several sleep features: early waking in children [1]; delayed or irregular bedtimes in adolescents [2], young adults [3], and older adults [4]; and short naps in shift workers [5].
- 3. "between X and Y" when Y contains "and". "Correlations between each covariate and A and B" does not show which "and" closes the pair. Name the second side as one unit: "correlations between each covariate and each of the two outcomes (A and B)".
- When you shorten text to meet a word limit, reread every sentence you shortened for these constructions and for pronouns introduced to save words ("remained associated with it"). Restore the noun even if it costs words; drop a lower-priority sentence instead (see Shortening order in [reader clarity](references/reader-clarity.md)).
- ---
- ## Process (two-pass draft-audit)
- ### Pass 1: draft rewrite
- 1. Read the input text carefully.
- 2. Audit meaning and information priority first. Identify unclear referents, omitted comparisons, stacked conditions, and redundant summary sentences using references/reader-clarity.md. Restore necessary relations without inventing facts.
- 3. Restructure sentence rhythm (Pattern 34). Before touching vocabulary, vary sentence lengths and diversify sentence openings across each paragraph. Keep the author's short claim sentences short; take the variation from the supporting sentences.
- 4. Identify and fix all vocabulary/phrase-level patterns (Patterns 1-33 and 35). After removing a transition or linking clause, check whether the logical relationship remains clear. Add a link only when needed; do not replace redundant summaries with generic bridge sentences. After removing ornamental adverbs (Pattern 29), re-check that the paragraph's rhythm has not become uniform.
- 5. Ensure the draft:
- - Sounds natural when read in an academic context
- - Matches the author's voice profile (Voice calibration section)
- - Uses precise, specific language with consistent terminology (Pattern 11)
- - Maintains data integrity (numbers, statistics, findings)
- - Uses simple constructions (is/are/has) where appropriate
- - Avoids promotional or inflated language
- ### Pass 2: self-audit
- 6. Ask yourself: "What makes this draft still look AI-generated?" List any remaining tells briefly. Common survivors include:
- - Omitted comparison conditions or ambiguous pronouns that force readers to reconstruct meaning
- - A subject that does not naturally perform its verb, or wording that hides who judged or acted and on what basis (see Subject and verb compatibility)
- - Sentence lengths still too uniform (check the Pattern 34 criterion: in paragraphs of three or more sentences, do all sentences fall within a 5-word range?)
- - Sentence openings still repetitive (check: do three+ consecutive sentences start the same way?)
- - Vocabulary tells that slipped through (check Patterns 1, 7, 29 word lists)
- - Independent clauses chained with ", and", especially ", and this/it ..." (check Pattern 35: is the relation causal, contrastive, or a specification? Is ", and" used as a joining device more than once in a paragraph? Does one subject carry three or more predicates, two predicates split by a citation block (", and with ..." after the references), or a gapped clause? Is a list nested inside a list without semicolons? Read every paragraph for these yourself; the checker lists only some of them.)
- - Broken connective chain (check Pattern 31 checklist)
- 7. Fix every issue found in the self-audit.
- ### Mandatory final checks
- When you can run commands, start these checks by running the bundled checker once on the draft. It counts what is easy to miscount: the words in each sentence of every paragraph, em dashes, curly quotes, repeated sentence openers, sentences that may chain clauses with ", and", and numbers that appear in the revision but not in the original.
- ```bash
- python "${CLAUDE_SKILL_DIR}/scripts/humanizer_check.py" <<'HUMANIZER_EOF'
- ===ORIGINAL===
- (the input text)
- ===REVISED===
- (your draft)
- HUMANIZER_EOF
- ```
- The script is in the `scripts` folder next to this file. Use that path if the variable above is not set, and use `python3` if `python` is not available. For text that is already in files, pass the revised file as the argument and the original with `--original`.
- Use the report as follows:
- - Take its counts as the result of the em dash check and of the 5-word criterion in the rhythm check, then fix what it reports.
- - Check each added number against the input. A number that the input spelled out as a word is fine.
- - Read each candidate word in context. The list names candidates, and a legitimate use stays (see Preserve legitimate academic phrases).
- - Read each sentence in the ", and" list (Pattern 35). Commas before the last item of a simple list are counted too, so many entries are fine. A paragraph that the checker does not list can still chain clauses; the checker does not replace reading. A clean report is not the end of the Pattern 35 read: read the first sentence of each paragraph and every sentence over 40 words yourself, because those are where stacked predicates hide.
- - Run it at most twice: once on the draft and once after your fixes. Do not keep rewriting to clear the candidate list.
- - Do not paste the report into your answer. Fold what you fixed into the summary of changes.
- If the checker cannot run (no Python, no permission to run commands, or a chat without tools), do checks 8 to 11 by hand as written and continue. Do not ask the author to install anything.
- 8. Fidelity check: Compare the output with the input and any author-supplied sources. Every number, unit, statistic, date, named entity (trial, drug, cohort, instrument, author), and citation in the output must also appear in the input or in those sources, and every assertion must be grounded in the input. If anything is not, revert that span to the input's wording or replace it with a `[DATA NEEDED: ...]` placeholder. Do not rewrite around the problem.
- 9. Em dash check: Search your output for "—". If any remain, replace them. Zero em dashes allowed.
- 10. Paragraph cohesion check (Pattern 31): Re-read every paragraph top to bottom: (a) first sentence states the paragraph's claim; (b) every subsequent sentence is linked to the previous one by a connective or echoed key word; (c) paragraph-opening contrast/continuity markers survive where the argument needs them. If any link was broken, repair it. Choppy, disconnected prose is not acceptable humanized output.
- 11. Rhythm and readability check (Pattern 34): This check is mandatory. (a) Apply the Pattern 34 criterion to every paragraph of three or more sentences. If all sentences fall within a 5-word range of each other, restructure at least one supporting sentence (split a sentence that stacks conditions, combine related supporting sentences, or reposition clauses). Never merge a short claim sentence to do this. (b) Re-read for one-pass comprehension: no restructuring may hide an actor, comparison, or condition. Remove redundant summaries; preserve concrete interpretation, necessary qualifications, and approved author opinions.
- 12. Present the humanized version.
- ## Output format
- Provide:
- 1. The rewritten text
- 2. A brief summary of changes made, noting which patterns were applied and any rhythm restructuring performed
- 3. Fidelity check result: confirm that every number, statistic, named entity, and citation in the output appears in the input or in author-supplied sources, and name any span you reverted
- 4. A list of every `[DATA NEEDED: ...]` placeholder inserted, so the author can fill each one
- 5. If any: "Candidate data for the author (unverified, not inserted)", listing figures or references you recalled that might fill a placeholder or support a vague claim. The author must verify each one against the primary source before using it
- ---
- ## Full example
- Before (AI-sounding):
- > Heart failure represents a pivotal challenge in the evolving landscape of diabetes care, affecting more than one in five patients with type 2 diabetes aged over 65 years and underscoring the critical importance of addressing cardiovascular comorbidities. This groundbreaking study showcases the profound impact of empagliflozin, a pivotal therapeutic option that serves as a cornerstone of modern cardiovascular medicine: when added to standard of care in patients with type 2 diabetes and high cardiovascular risk, it reduced heart failure hospitalization and cardiovascular death.
- >
- > Studies have shown that SGLT2 inhibitors reduce cardiovascular events. Additionally, in the EMPA-REG OUTCOME trial, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%—a remarkable finding—highlighting the cardioprotective effects of this intervention. The number needed to treat of 35 over 3 years to prevent one event underscores the crucial clinical value of this therapeutic approach.
- >
- > Despite challenges typical of large clinical trials, including a diagnosis of heart failure at baseline that relied solely on investigator report without measures of cardiac function or biomarkers, the trial's strategic design continues to provide valuable insights for the future outlook of heart failure management. With a benefit that was consistent in patients with and without heart failure at baseline, the future looks bright for patients with type 2 diabetes as these exciting findings continue to reshape clinical practice.
- After (Humanized):
- > Heart failure is highly prevalent in patients with diabetes, occurring in more than one in five patients with type 2 diabetes aged over 65 years. In patients with type 2 diabetes and high cardiovascular risk, empagliflozin reduced heart failure hospitalization and cardiovascular death when added to standard of care.
- >
- > In the EMPA-REG OUTCOME trial, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%. The number needed to treat to prevent one event was 35 over 3 years.
- >
- > The diagnosis of heart failure at baseline was based solely on the report of investigators, with no measures of cardiac function or biomarkers recorded. Empagliflozin reduced heart failure hospitalization and cardiovascular death when added to standard care. The benefit was consistent in patients with and without heart failure at baseline.
- Changes made:
- - Eliminated all em dashes ("—") per Pattern 13
- - Removed significance inflation ("pivotal challenge", "evolving landscape", "groundbreaking", "cornerstone")
- - Removed promotional language ("profound impact", "remarkable finding", "exciting findings")
- - Removed the unsupported vague attribution ("Studies have shown" with no citation) and let the specific trial result carry the claim (note: "Prior studies have shown that..." followed by citations would be preserved)
- - Removed superficial -ing phrases ("underscoring", "highlighting")
- - Removed copula avoidance ("serves as") in favor of "is"
- - Removed AI vocabulary ("crucial", "pivotal") and ornamental intensifiers ("remarkable"). Note that the "Additionally" disappeared only because its whole sentence was rewritten; a single "Additionally" per paragraph is acceptable and would otherwise be kept (Pattern 7 exception)
- - Removed formulaic challenges section ("Despite challenges... future outlook")
- - Removed generic positive conclusion ("The future looks bright", "continue to reshape")
- - Fixed grammar ("The number needed to treat of 35" → "was 35")
- - Used simple sentence structures and specific data
- - Split long sentences that stacked conditions (the baseline-diagnosis limitation now stands as its own long sentence, followed by two shorter ones)
- - Fidelity check passed: every number, trial name, and statistic in the After appears in the Before, so no `[DATA NEEDED]` placeholders or candidate data were needed
- ---
- ## Reference
- This skill is based on [Wikipedia:Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing), maintained by WikiProject AI Cleanup, adapted for medical and academic writing contexts. The patterns documented there come from observations of thousands of instances of AI-generated text.
- Medical paper examples are adapted from:
- > Fitchett D, Inzucchi SE, Cannon CP, et al. Empagliflozin Reduced Mortality and Hospitalization for Heart Failure Across the Spectrum of Cardiovascular Risk in the EMPA-REG OUTCOME Trial. *Circulation*. 2019;139(11):1384-1395. doi:10.1161/CIRCULATIONAHA.118.037778
- This article is published under CC-BY-4.0 license.
README.md · README · 240 lines
- <p align="center">
- <img src="./assets/banner.png" alt="humanizer_academic: Remove signs of AI writing from medical manuscripts" width="820">
- </p>
- # Humanizer Academic
- A skill that removes signs of AI-generated writing from academic medical papers, making them sound more natural and professionally written.
- The main instructions are in `SKILL.md`, with reusable examples in `references/reader-clarity.md`. Both are plain text, so you do not need Claude Code to use them. `scripts/humanizer_check.py` is an optional checker for agents that can run commands; the skill works without it.
- ## Usage
- ### In a browser (Claude, ChatGPT, or anything that takes an attachment)
- 1. Download [`SKILL.md`](SKILL.md) and [`references/reader-clarity.md`](references/reader-clarity.md) from this repository.
- 2. Upload both files to the chat along with the text you want to edit.
- 3. Ask: `Apply this skill to the following manuscript text: [your text]`
- That is all. No installation.
- You do not need the checker script in a browser chat; the skill does the same checks by hand. If your chat can run code, you can also upload [`scripts/humanizer_check.py`](scripts/humanizer_check.py), and the skill will run it on its draft.
- ### In Claude Code
- Clone into your skills directory:
- ```bash
- mkdir -p ~/.claude/skills
- git clone https://github.com/matsuikentaro1/humanizer_academic.git ~/.claude/skills/humanizer_academic
- ```
- Or, from a downloaded copy of the repository:
- ```bash
- mkdir -p ~/.claude/skills/humanizer_academic/references ~/.claude/skills/humanizer_academic/scripts
- cp SKILL.md ~/.claude/skills/humanizer_academic/
- cp references/reader-clarity.md ~/.claude/skills/humanizer_academic/references/
- cp scripts/humanizer_check.py ~/.claude/skills/humanizer_academic/scripts/
- ```
- #### Optional checker
- Language models miscount words and tend to grade their own drafts generously. When Python is available, the skill runs `scripts/humanizer_check.py` once on its draft before the final checks. The script uses only the standard library and reports:
- - the number of words in each sentence, paragraph by paragraph, and the paragraphs that fail the Pattern 34 rhythm criterion
- - em dashes and curly quotes
- - three consecutive sentences with the same first word, and "Additionally" used more than once in a paragraph
- - sentences that may chain clauses with ", and" (Pattern 35): ", and" followed by a verb, long sentences with three or more "and", gapped clauses, and "between X and A and B"
- - numbers that appear in the revision but not in the original
- - words named in the patterns, listed as candidates to read in context
- It gives no score, and a candidate word is not a violation. If Python is missing, the skill does the same checks by hand. You can also run it yourself:
- ```bash
- python scripts/humanizer_check.py revised.txt --original original.txt
- ```
- The skill loads itself when it is relevant, so you can simply ask:
- ```
- Please humanize this academic text: [your text]
- ```
- ## Overview
- Based on [Wikipedia's "Signs of AI writing"](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) guide, adapted for medical and scientific literature. Examples are derived from the EMPA-REG OUTCOME trial publications and from the author's (K. Matsui) observations during academic manuscript editing.
- ### Key Insight from Wikipedia
- > "LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases."
- ## 35 Patterns Detected (with Before/After Examples)
- ### Content Patterns
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 1 | **Significance inflation** | "represents a pivotal challenge in the evolving landscape" | "is highly prevalent in patients with diabetes" |
- | 2 | **Notability claims** | "landmark trial, led by renowned investigators" | "A total of 7020 patients..." |
- | 3 | **Superficial -ing analyses** | "highlighting the cardioprotective effects" | Report data without interpretation |
- | 4 | **Promotional language** | "groundbreaking study showcases the profound impact" | "empagliflozin reduced heart failure hospitalization" |
- | 5 | **Vague attributions** | "Studies have shown... Experts argue..." | "In the EMPA-REG OUTCOME trial..." |
- | 6 | **Formulaic challenges** | "Despite challenges... future outlook" | State specific limitations |
- ### Language Patterns
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 7 | **AI vocabulary** | "pivotal... landscape... crucial" | Remove or replace with simple words (note: "Additionally" is allowed once per paragraph) |
- | 8 | **Copula avoidance** | "serves as... standing as... representing" | "is" |
- | 9 | **Negative parallelisms** | "Not only X but also Y" | "X and Y" |
- | 10 | **Rule of three** | "efficacy, safety, and tolerability" | Use natural number of items |
- | 11 | **Synonym cycling & term consistency** | "Patients... Participants... Subjects" | "Patients" throughout (same construct = same term) |
- | 12 | **False ranges** | "from renal function to cardiac outcomes" | List benefits directly |
- ### Style Patterns
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 13 | **Em dash elimination** | "benefits—a 35% reduction—appeared early—" | Use commas, parentheses, or periods. Every em dash is replaced, including natural-looking ones |
- | 14 | **Title Case Headings** | "Statistical Analysis And Primary Endpoints" | "Statistical analysis and primary endpoints" |
- | 15 | **Curly quotes** | “clinically significant” | "clinically significant" |
- ### Filler and Hedging
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 16 | **Filler phrases** | "In order to", "Due to the fact that" | "To", "Because" |
- | 17 | **Redundant multi-layered hedging** | "may suggest... have the potential to confer" | "suggest... may reduce" (keep 1-2 hedges) |
- | 18 | **Generic conclusions** | "The future looks bright" | Specific findings and implications |
- ### LLM-Specific Word Choice Patterns (v1.1.0)
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 19 | **"linked to" → "associated with"** (context-dependent) | "has been linked to shorter sleep duration" | "has been reported to be associated with" |
- | 20 | **"Beyond" → "In addition to"** | "Beyond the association with..." | "In addition to the association with..." |
- | 21 | **"via" → "through"** | "obtained via the online form" | "obtained through an online form" |
- | 22 | **Insufficient hedging** | "may reduce the risk of..." | "may help reduce the risk of..." |
- | 23 | **Artificially condensed expressions** | "fatigue–sleepiness cycle", "mutual reinforcement" | "cycle of fatigue and sleepiness", "a self-reinforcing cycle, with each behavior possibly exacerbating the other" |
- | 24 | **"where" as a non-locative connector** | "...at the most intensive level, where almost daily use was..." | "...at the most intensive level, with almost daily use..." |
- | 25 | **"yield" as a result verb** | "did not yield stable estimates" | "failed to produce stable estimates" |
- | 26 | **Minor word-choice refinements** | "interpretations remain speculative", "speculative given the small sizes" | "interpretations are still speculative", "speculative due to the small sizes" |
- ### Cohesion and Connective Patterns (v1.1.4–1.3.0)
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 27 | **Preserve logical discourse markers** + vary by relation (do NOT over-trim) | (aggressive removal of "Although / Whereas / Thus / Based on these results") | Keep connectives that make the logic explicit; vary them by relation (result / contrast / concession) only to avoid mechanical repetition, never for decoration |
- | 28 | **Re-contextualize over-condensed semantic links** | "unmet needs to discuss their difficulties" | "unmet needs when it comes to discussing their difficulties" |
- | 29 | **Ornamental -ly intensifier adverbs** | "markedly reduced", "critically important", "remarkably consistent" | Remove decorative intensifiers; keep functional ones ("slightly", "consistently", "approximately") |
- | 30 | **Connective-preserving edits** (never bare-delete a transition) | "X reduced death. The benefit appeared within months." (choppy) | "X also reduced death. This benefit appeared within months." (a connective plus an echoed noun, not a default ", and this") |
- | 31 | **Paragraph cohesion** (old-to-new flow + paragraph-opening markers) | Disconnected sentences after editing | Mandatory final check: each sentence links to the previous one; contrast/continuity openers (However / On the other hand / Overall / Taken together) survive |
- | 32 | **Paraphrastic repetition** | "X is associated with Y. In other words, X may contribute to Y. That is, X plays a role in Y." | State each claim once; keep the most specific version |
- | 33 | **Content-free evaluation sentences** | "This is a noteworthy finding." "This observation is of clinical significance." | Delete standalone verdicts; if important, show why with data or mechanism |
- ### Sentence Rhythm and Reader Clarity (v2.3)
- | # | Pattern | Before | After |
- |---|---------|--------|-------|
- | 34 | **Sentence rhythm & burstiness, in the service of clarity** | Uniform 18-22 word sentences; stacked conditions and unclear comparisons | Vary sentence lengths and openings in every paragraph by splitting stacked sentences and combining related supporting sentences; keep short claim sentences short |
- | 35 | **Independent clauses chained with ", and"** (v2.6; stacked predicates, nested lists, and "between X and A and B" added in v2.7; predicates split by citation blocks added in v2.7.1) | "...for DACS, and this was most prominent for MorningHarbor." "...may be disproportionate to its frequency, and it warrants attention." | Name the relation: "...for DACS, most prominently for MorningHarbor." "...it warrants attention because its impact may be disproportionate to its frequency." Use ", and" as a joining device at most once per paragraph |
- > **Reader clarity comes first.** Preserve actors, comparisons, conditions, and logical relationships. Shorten by removing low-priority or repetitive sentences rather than compressing essential meaning. Burstiness is still required in every paragraph (mandatory rhythm check), but it is obtained only through edits that keep meaning explicit, and never by merging the author's short claim sentences. See the [eight diagnostic examples](references/reader-clarity.md).
- ### Preserved Academic Writing (do NOT flag as AI)
- The skill explicitly **preserves** standard academic writing that was previously over-corrected:
- - Transitional phrases: "Notably,", "Furthermore,", "In contrast,", etc.
- - Logical discourse markers: "Although", "Whereas", "Thus", "Based on these results", "As expected", etc. (Pattern 27)
- - Functional -ly adverbs: "slightly", "consistently", "modestly", "approximately" (Pattern 29)
- - Interrogative sentence openers: "Who selects into...", "What predicts...", etc.
- - "Additionally" up to once per paragraph (Pattern 7 exception)
- - Attribution phrases with citations: "Prior studies have shown that...", "Evidence suggests that...", etc.
- These are only flagged when used in excessive clusters or without supporting citations/data.
- ## Full Example
- **Before (AI-sounding):**
- > Heart failure represents a pivotal challenge in the evolving landscape of diabetes care, underscoring the critical importance of addressing cardiovascular comorbidities. This groundbreaking study showcases the profound impact of empagliflozin. Additionally, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%—a remarkable finding—highlighting the cardioprotective effects of this intervention. The future looks bright for patients with type 2 diabetes.
- **After (Humanized):**
- > Heart failure is highly prevalent in patients with diabetes, occurring in more than one in five patients with type 2 diabetes aged over 65 years. In the EMPA-REG OUTCOME trial, empagliflozin reduced the risk of hospitalization for heart failure or cardiovascular death by 34%. The benefit was consistent in patients with and without heart failure at baseline.
- ## References
- - [Wikipedia: Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) - Primary source for AI writing patterns
- - [WikiProject AI Cleanup](https://en.wikipedia.org/wiki/Wikipedia:WikiProject_AI_Cleanup) - Maintaining organization
- ### Examples Source
- Medical paper examples (Patterns 1–18) are adapted from:
- > Fitchett D, Inzucchi SE, Cannon CP, et al. Empagliflozin Reduced Mortality and Hospitalization for Heart Failure Across the Spectrum of Cardiovascular Risk in the EMPA-REG OUTCOME Trial. *Circulation*. 2019;139(11):1384-1395. doi:10.1161/CIRCULATIONAHA.118.037778
- This article is published under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license.
- Examples for Patterns 19–31 are based on the author's (K. Matsui) observations during academic manuscript editing in sleep medicine research.
- The AI-vocabulary patterns (Pattern 7) draw on the following papers that quantified words and phrases with increased or decreased frequency in post-ChatGPT medical and scientific writing:
- > Matsui K. Delving Into PubMed Records: How AI-Influenced Vocabulary has Transformed Medical Writing since ChatGPT. *Perspect Med Educ*. 2025;14(1):882-890. doi:[10.5334/pme.1929](https://doi.org/10.5334/pme.1929)
- > Bao T, Zhao Y, Mao J, et al. Examining linguistic shifts in academic writing before and after the launch of ChatGPT: a study on preprint papers. *Scientometrics*. 2025;130:3597-3627. doi:[10.1007/s11192-025-05341-y](https://doi.org/10.1007/s11192-025-05341-y)
- > Galpin R, Anderson B, Juzek TS. Exploring the Structure of AI-Induced Language Change in Scientific English. *Int FLAIRS Conf Proc*. 2025;38. doi:[10.32473/flairs.38.1.138958](https://doi.org/10.32473/flairs.38.1.138958)
- ## Related Work by the Author
- > Matsui K. Delving Into PubMed Records: How AI-Influenced Vocabulary has Transformed Medical Writing since ChatGPT. *Perspect Med Educ*. 2025 Dec 2;14(1):882-890. doi:[10.5334/pme.1929](https://doi.org/10.5334/pme.1929)
- This is a paper I wrote. Using PubMed records, I measured how frequently LLMs such as ChatGPT tend to overuse certain words in medical writing. Take a look if you're curious!
- 
- *The image above is Figure 1 from the paper cited above, reproduced under the [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) license.*
- ## Releasing a New Version
- When you change the version, update all of the following in the same commit:
- 1. `version:` in the front matter of `SKILL.md`.
- 2. An entry at the top of the Version History below. If you added or retired a pattern, also update the pattern tables and the "35 Patterns" heading in this README.
- 3. The banner. Edit the version pill (and the pattern-count pill if the count changed) in `assets/banner-src/banner.html`, then run `python assets/banner-src/render.py` (requires Playwright with Chromium) to regenerate `assets/banner.png`. Open the PNG and check it before committing; the banner is not generated automatically and otherwise keeps showing the old version.
- 4. Your installed copy (`~/.claude/skills/humanizer_academic/`), using the copy commands under [In Claude Code](#in-claude-code).
- ## Version History
- - **2.7.1** - Pattern 35 now names the form that stacked predicates usually take in Introduction and Discussion: two predicates on one subject split by a block of citations ("X has been associated with A (refs), and with B (refs); Y has likewise ..."), with a fictional Before/After. The checker lists a sentence in which ", and" plus a preposition (with, to, in, among, ...) resumes right after a parenthesized or bracketed citation or statistic. The v2.7.0 checker missed this form because ", and with" is not followed by a verb. The skill now also says that a clean checker report does not end the Pattern 35 read: the first sentence of each paragraph and every sentence over 40 words are read by hand.
- - **2.7.0** - Extended Pattern 35 to three constructions that v2.6 left untargeted: one subject carrying three or more predicates, or two predicates separated by statistics, including gapped clauses such as ", and greater in A than in B" (split where the subject or comparison changes and restate the subject); lists nested inside lists (separate the outer items with semicolons); and "between X and A and B" (name the second side as one unit). These constructions are common in Results sentences and come from authors' drafts as well as from model edits. After shortening text to meet a word limit, the skill now rereads each shortened sentence for these constructions and for pronouns introduced to save words (Shortening order step 5 in reader-clarity.md). The checker now lists sentences in which ", and" is followed by a verb, long sentences with three or more "and", gapped clauses, and "between X and A and B", as candidates to read. On a 33-paragraph manuscript it listed 18 paragraphs before revision and 3 after; the skill still reads every paragraph, because the checker does not catch every chain.
- - **2.6.0** - Added Pattern 35 (independent clauses chained with ", and", especially ", and this/it ..."). Before joining two clauses with ", and", the skill now names their relation and writes it out: "because" for cause, "whereas" for contrast, a phrase such as "most prominently for X" for the degree of the same finding, or a semicolon for closely related facts. ", and" stays for genuinely parallel facts, at most once per paragraph as a joining device. The examples had been teaching the construction: the Afters of Patterns 29, 30, 33, and 34 joined clauses with ", and", and Pattern 30 offered ", and this benefit ..." as the right fix. They now use semicolons, echoed nouns, and prepositional openers. Pattern 3 says not to fill a deleted -ing tail with ", and this ...", and Pattern 34 no longer accepts ", and" as a way to combine sentences for rhythm. The checker lists ", and this/it/these" as a Pattern 35 candidate.
- - **2.5.1** - Added a "Subject and verb compatibility" check. Its main target is anthropomorphism: a report, figure, section, or analysis written as if it made a human judgment ("This single course cannot exclude..."). Verbs of display and evidence (show, indicate, suggest) stay with non-human subjects, while verbs of judgment and decision (determine, decide, exclude) go back to the authors. The skill now asks whether the subject naturally performs the action of its verb and whether the reader can tell who judged or acted and on what basis. Conventional constructions such as "Figure 1 shows..." and "These findings suggest..." are preserved, and inanimate subjects are not mechanically replaced with "we"; Methods sections keep the passive voice. The check is part of the Pass 2 self-audit.
- - **2.5.0** - The instructions now follow their own rules. A model copies the style of its prompt as well as the content, so `SKILL.md` no longer uses em dashes in its own prose (31 removed; the character remains only where Pattern 13 has to show it), bold text, capitalized emphasis such as "NOT", Title Case headings, or "Key principle" labels. The rules themselves are unchanged. Added `scripts/humanizer_check.py`, an optional standard-library checker that counts words per sentence for the Pattern 34 criterion, finds em dashes, curly quotes and repeated openers, and lists numbers that appear in the revision but not in the original. It reports candidates without a score, and the skill falls back to the manual checks when Python is unavailable.
- - **2.4.0** - Keep model-recalled figures and citations out of the rewritten text. "Be specific" now uses only data from the manuscript or author-supplied sources; the `[verify: ...]` in-text insertion is removed. A new mandatory fidelity check confirms that every number, statistic, named entity, and citation in the output appears in the input, reverting any span that does not. Recalled candidates are listed separately in the output as unverified and not inserted (thanks to @kaicot for issue #5).
- - **2.3.0** - Restore burstiness as a required step (Pattern 34 and the mandatory rhythm check) while keeping the v2.2 reader-clarity priority and the v2.1.1 rule that short claim sentences stay short. Rhythm variation now comes from splitting sentences that stack conditions, combining related supporting sentences, and varying openings. "Be specific" now replaces vague claims with real data, marks facts drawn from model knowledge with `[verify: ...]`, and uses `[DATA NEEDED: ...]` placeholders instead of inventing specifics. Pattern 13 keeps the zero-em-dash rule with a calmer tone (thanks to @kaicot for PR #3 and #4). Example Befores now contain every fact in their Afters, so the examples demonstrate style edits without adding facts; the Pattern 34 word counts, the rhythm criterion, and a wrong cross-reference were also corrected.
- - **2.2.0** - Prioritize reader clarity over compressed prose and forced rhythm variation. Add eight reusable examples, sentence-level deletion guidance, abstract consistency checks, and protection for approved author opinions and intentional section spacing. Include the new reference file when installing or using the skill in a chat.
- - **2.1.1** - Voice Calibration and Pattern 34: keep the author's short sentences that state a claim outright ("The largest difference was cost."). They are no longer merged into neighboring sentences to satisfy the rhythm check, because merging buries and softens the claim.
- - **2.1.0** - Slimmed SKILL.md description to fit the 1024-character limit enforced by `claude install-skill`. No changes to patterns or skill behavior.
- - **2.0.0** - Major update based on experimental validation with local AI detectors (desklib + Binoculars). Added Pattern 34 (Sentence rhythm & burstiness: experimentally verified as the single highest-impact intervention, accounting for ~90% of achievable AI-score reduction). Added Voice Calibration section with author reference profile. Upgraded Process to two-pass draft-audit loop with mandatory rhythm check. Strengthened Pattern 11 to explicitly require term consistency (same construct = same term throughout). Documented critical interaction: Pattern 29 adverb removal must always be paired with Pattern 34 sentence restructuring (standalone deletion increases AI scores).
- - **1.4.0** - Added Pattern 32 (Paraphrastic repetition: detect and remove same-claim restatements joined by "In other words" / "That is" / "Essentially") and Pattern 33 (Content-free evaluation sentences: remove standalone verdicts like "This is a noteworthy finding" that add no data or mechanism). Extended Pattern 7 vocabulary with "comprehensive" (abstract use only), "holistic", and "multifaceted".
- - **1.3.1** - Extended Pattern 27 to vary connectives by logical relation (result / addition / contrast / concession / reason / sequence groups) to avoid mechanical repetition, with a guardrail against decorative connective-sprinkling (this is NOT an exception to Pattern 11).
- - **1.3.0** - Added Pattern 29 (Ornamental -ly intensifier adverbs: remove decorative "markedly/critically/remarkably" while keeping functional adverbs such as "slightly/consistently/approximately"), Pattern 30 (Connective-preserving edits: never bare-delete a transition — replace or restructure to avoid choppy asyndeton), and Pattern 31 (Paragraph cohesion: a mandatory final check for old-to-new flow and surviving paragraph-opening markers). Relaxed Pattern 7 so "Additionally" is allowed once per paragraph (only excessive use is flagged).
- - **1.2.x** - Reworked Pattern 26 into "Minor word-choice refinements" (remain → be-verb, given → due to); added Pattern 27 (Preserve logical discourse markers), Pattern 28 (Re-contextualize over-condensed semantic links), and context-dependent handling of "linked/associated" (Pattern 19). (The earlier "Underused Classical Academic Terms" pattern was retired during this restructuring; its vocabulary is now covered by Pattern 7 and the cited author papers.)
- - **1.1.3** - Added patterns 24 ("where" as a non-locative connector) and 25 ("yield" as a result verb); added author paper reference and Fig.1 to README
- - **1.1.2** - Pattern 13: Em dash rule upgraded to zero-tolerance elimination (no exceptions, mandatory final check step)
- - **1.1.1** - Merged compressed noun-dash phrases and vague abstractions into single "Artificially condensed expressions" pattern (23)
- - **1.1.0** - Added LLM-specific word choice patterns (19-23), preserved legitimate academic phrases, fixed hedging guidance consistency
- - **1.0.0** - Initial release adapted for academic medical writing
- ## Contributors
- - [@kaicot](https://github.com/kaicot): found that the short-sentence rule still referred to a removed rhythm check ([#3](https://github.com/matsuikentaro1/humanizer_academic/pull/3)), found the broken Pattern 15 example, and proposed the data-sourcing rule for "Be specific", the calmer Pattern 13 wording, and the note on illustrative examples ([#4](https://github.com/matsuikentaro1/humanizer_academic/pull/4)). These were adopted in v2.3.0. Also proposed the fidelity check that keeps model-recalled data out of the rewritten text ([#5](https://github.com/matsuikentaro1/humanizer_academic/issues/5)), adopted in v2.4.0.
- ## License
- MIT
- Based on [blader/humanizer](https://github.com/blader/humanizer).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.