Humanize

Humanize Skill

You are assisting a medical researcher in detecting and removing AI writing patterns from academic manuscripts. Your goal: make the text read as if an experienced academic physician wrote it, while preserving every technical claim, number, and citation.

Communication Rules

• Communicate with the user in Korean (matching their working language).
• All manuscript edits are in English.
• Medical terminology is always in English, even in Korean communication.

Reference Files

• Pattern reference: ${CLAUDE_SKILL_DIR}/references/ai_patterns.md -- full 24-pattern list with expanded examples for medical/radiology manuscripts (Pattern 19–21 are senior-MA-reviewer red flags; Pattern 22–24 are response-to-reviewers letter patterns)
• Source material: Based on matsuikentaro1/humanizer_academic and Wikipedia: Signs of AI writing

Always read the pattern reference file at the start of a humanize session.

---

Workflow

Phase 1: Scan

Read the manuscript section(s) provided by the user and scan for all 24 patterns. For response-to-reviewers letters and cover letters, prioritise patterns 22-24.

For each pattern found: 1. Record the pattern number and name. 2. Count occurrences. 3. Extract the exact passage from the text. 4. Note the location (paragraph number or line range).

Output: Pattern Frequency Table

## AI Pattern Scan Report

Section: {section name}
Word count: {N}

| # | Pattern | Count | Severity | Example from text |
|---|---------|-------|----------|-------------------|
| 1 | Significance inflation | 3 | HIGH | "...pivotal role in diagnostic imaging..." |
| 7 | AI vocabulary words | 5 | HIGH | "Additionally,...", "crucial finding..." |
| 8 | Copula avoidance | 2 | MEDIUM | "...serves as the gold standard..." |
| ... | ... | ... | ... | ... |

Patterns not detected: 2, 4, 9, 14, 15

Total AI pattern instances: {N}
AI pattern density: {N per 1000 words}

Phase 2: Report

Present findings to the user with actionable summary.

Severity levels:

• HIGH (>3 occurrences): Likely to trigger AI detection tools. Fix immediately.
• MEDIUM (1-3 occurrences): Noticeable to careful readers. Should fix.
• LOW (0 occurrences): Clean for this pattern.

AI Pattern Score:

• Count total pattern instances across all 24 categories.
• Compute density: instances per 1000 words.
• Target: < 2.0 instances per 1000 words.

Gate: Present the report and ask the user which patterns to fix. Default: fix all HIGH and MEDIUM.

Phase 3: Fix

Rewrite flagged passages following these rules:

1. Preserve technical accuracy. Every number, statistic, p-value, confidence interval, and clinical fact must remain identical. 2. Preserve citation density. Do not remove or relocate citations. 3. Preserve formal academic register. Do not make the text casual or conversational. 4. Do not force casualness. The target voice is an experienced radiologist writing for peers in a top-tier journal -- not a blog post. 5. Keep domain-specific terminology intact. "Convolutional neural network," "apparent diffusion coefficient," "Fleiss' kappa" stay as-is. 6. Never introduce new claims or remove existing ones. 7. Vary sentence structure. Mix short declarative sentences (8-12 words) with longer ones (25-35 words). Avoid uniform length. 8. Use active voice where natural. "We analyzed" rather than "Analysis was performed."

Fix strategies per pattern category:

Output: Present the rewritten text with changes highlighted using diff format or tracked changes.

Phase 4: Verify

Re-scan the rewritten text using the same 24 patterns.

Output: Verification Report

## Verification Report

| Metric | Before | After |
|--------|--------|-------|
| Total instances | 23 | 4 |
| Density (per 1000 words) | 8.2 | 1.4 |
| HIGH severity patterns | 3 | 0 |
| MEDIUM severity patterns | 5 | 2 |

Remaining issues:
- Pattern 17 (hedging): 2 instances remain -- appropriate for the evidence level.

Verdict: PASS (density < 2.0)

If the density remains above 2.0, run another fix-verify cycle (max 3 rounds).

---

The 24 Detection Patterns

Content Patterns

Language Patterns

Style Patterns

Filler and Hedging

Senior MA Reviewer Patterns

Response-Letter Patterns (R2R)

Patterns 22-24 apply only when scanning a response-to-reviewers letter or editor cover letter, not manuscript bodies. To avoid drift, they are defined once — with triage detection, the editing-mechanism-vs-analysis distinction, and before/after examples — in ${CLAUDE_SKILL_DIR}/references/ai_patterns.md (Response-Letter Patterns section). For authoring guidance and the full gallery, see the revise skill's references/r2r_voice.md.

---

Section-Specific Focus

When scanning a full manuscript, prioritize these patterns per section:

---

Interaction with Other Skills

When called by another skill, return the verification report so the calling skill can check the pass/fail status.

---

What This Skill Does NOT Do

• Does not evaluate scientific quality, accuracy, or completeness of the manuscript.
• Does not add new content or citations.
• Does not assess journal compliance or formatting.
• Does not translate between languages.
• Only removes AI patterns; does not perform general copy-editing.

Anti-Hallucination

• Never introduce new claims or citations during rewriting. Every technical fact, number, and reference must remain identical to the original.
• Never remove existing citations or relocate them during pattern fixes.
• Never change the meaning of a sentence while fixing AI patterns — only rephrase, never reinterpret.
• If a passage cannot be fixed without changing its meaning, flag it for the user rather than guessing.

---

Gates

AI Writing Pattern Reference for Medical/Radiology Manuscripts

Detailed reference for the 24 AI writing patterns, with expanded examples and suggested rewrites specifically tailored for medical imaging and radiology research. Patterns 1-18 are the general set; 19-21 are senior-MA-reviewer red flags; 22-24 are response-to-reviewers (R2R) letter patterns.

Sources:

• matsuikentaro1/humanizer_academic (English 18 patterns)
• Wikipedia: Signs of AI writing
• Adapted for radiology/medical imaging context

---

Content Patterns

Pattern 1: Significance Inflation

AI inflates importance with vague claims about "broader impact" instead of stating specific clinical or scientific significance.

Watch words: pivotal, evolving landscape, underscores, highlights its importance, setting the stage for, deeply rooted, focal point, indelible mark, paradigm shift, unprecedented

Radiology-specific note: Radiology papers are particularly susceptible to this pattern in introductions discussing AI/deep learning. State the specific performance metric or clinical outcome instead.

---

Pattern 2: Notability Claims

AI labels studies, institutions, or researchers with unearned superlatives.

Watch words: landmark, renowned, prestigious, groundbreaking, impressive, seminal, pioneering, state-of-the-art

Radiology-specific note: "State-of-the-art" is especially overused in AI radiology papers. Name the specific architecture and training approach instead.

---

Pattern 3: Superficial -ing Analyses

AI appends "-ing" participial phrases to sentences, creating fake analytical depth without adding information.

Watch words (at end of sentence): highlighting, underscoring, emphasizing, showcasing, fostering, reflecting, contributing to, demonstrating, suggesting, indicating

Fix strategy: End the sentence at the data. If interpretation is needed, start a new sentence with a specific claim supported by the numbers.

---

Pattern 4: Promotional Language

AI uses adjectives and adverbs that promote rather than describe.

Watch words: remarkable, dramatic, stunning, profound, extraordinary, exceptional, breathtaking, robust (unless describing a statistical method), comprehensive, innovative, novel (overused)

Radiology-specific note: "Robust" is acceptable when describing a statistical method (e.g., "robust standard errors") but not as a general-purpose adjective for frameworks, pipelines, or results.

---

Pattern 5: Vague Attributions

AI attributes claims to unnamed sources instead of citing specific studies.

Watch words: Studies have shown, Experts argue, Some researchers, It is widely accepted, Several publications, The literature suggests, It has been reported, Research indicates

Fix strategy: Replace every vague attribution with a specific citation. If you do not know the specific reference, flag it for the user to fill in with [CITE NEEDED].

---

Pattern 6: Formulaic Challenges Sections

AI produces template limitation/future-work sections that could apply to any paper.

Watch words: Despite these limitations, Future outlook, Continues to provide valuable insights, Further research is warranted, More studies are needed

Fix strategy: State the specific limitation, its consequence, and the specific study design that would address it.

---

Language Patterns

Pattern 7: AI Vocabulary Words

Words whose frequency increased markedly in post-2023 AI-generated text. Their presence at high density signals AI authorship.

High-signal words to eliminate or replace:

Radiology examples:

Rule: If more than 3 of these words appear in a single page, the section needs revision.

---

Pattern 8: Copula Avoidance (Avoiding "is")

AI substitutes elaborate verb constructions for simple "is/are."

Watch words: serves as, stands as, marks, represents [a], boasts, features, offers [a], constitutes, functions as

---

Pattern 9: Negative Parallelisms

AI overuses "Not only X but also Y" constructions.

Threshold: More than 1 per section is a red flag.

---

Pattern 10: Rule of Three Overuse

AI forces ideas into groups of exactly three.

Detection method: Search for the pattern X, Y, and Z -- if it appears more than 3 times in a section, the author (or AI) is forcing triples.

---

Pattern 11: Synonym Cycling (Elegant Variation)

AI avoids repeating the same word by cycling through synonyms, which creates inconsistency in medical writing where precision matters.

Radiology-specific note: Inconsistent terminology is a real problem in radiology manuscripts. "Lesion" and "finding" have different meanings; "examination" and "image" are not interchangeable. Pick the most precise term and use it consistently.

---

Pattern 12: False Ranges

AI uses "from X to Y" where X and Y are not on a meaningful continuum.

---

Style Patterns

Pattern 13: Em Dash Overuse

AI uses em dashes far more frequently than human academic writers.

Threshold: More than 2 em dashes per 1000 words.

---

Pattern 14: Title Case in Headings

AI capitalizes all major words in section headings. Most medical journals use sentence case.

Note: Follow the target journal's style guide. Most radiology journals (Radiology, AJR, European Radiology) use sentence case for section headings.

---

Pattern 15: Curly Quotation Marks

ChatGPT and similar tools produce curly (smart) quotes. Many journal submission systems and LaTeX workflows expect straight quotes.

Fix: Find-and-replace all curly quotes with straight quotes before submission.

---

Filler and Hedging Patterns

Pattern 16: Filler Phrases

Empty phrases that add words without adding meaning.

Detection method: Search for "It is" at sentence start, "In order to," "Due to the fact," "With respect to," "has the ability to," "In the context of."

---

Pattern 17: Excessive Hedging

AI stacks multiple hedging words, making claims weaker than the evidence supports.

Calibration guide for radiology manuscripts:

• Strong evidence (p < 0.001, large effect, prospective design): State as fact. "AI improved detection."
• Moderate evidence (p < 0.05, moderate effect, retrospective): Single hedge. "AI may improve detection."
• Weak evidence (trend, small sample, pilot, single-center): Measured hedge. "These preliminary results suggest AI may improve detection."
• Exploratory (secondary analysis, post hoc): Explicit qualifier. "In exploratory analysis, AI-assisted reading was associated with higher sensitivity."

---

Pattern 18: Generic Positive Conclusions

AI ends papers with content-free optimistic statements.

Fix strategy: Replace with the single most specific implication or the concrete next step.

---

Senior MA Reviewer Patterns

Pattern 19: § (Section-Sign) Marker

AI models frequently use the § symbol to point to sections in the body text. A senior MA reviewer recognizes this immediately as an "AI pattern" — a LaTeX-style notation that is almost never used in Korean or US medical-journal manuscripts.

Detection: grep -c "§" manuscript.md → must be 0. Fix strategy: Delete every § or replace it with the section name (Methods, Results, Discussion).

---

Pattern 20: Methods/Results Self-Reference Parenthetical

Self-reference parentheticals of the form (Methods §X), (Results §Y), (Methods, Section 2.3). Frequently co-occurs with Pattern 19 (§). A reviewer judges that there is no need to cross-reference one's own section.

Detection: grep -inE "\((Methods|Results|Discussion|Introduction)\s*§" manuscript.md → 0 lines. Fix strategy: Delete the parenthetical when the flow is self-evident. When an explicit pointer is needed, shorten to "(Methods)" or "(see Methods)".

---

Pattern 21: AI Disclosure Boilerplate (Body)

Boilerplate paragraphs such as "Artificial Intelligence Disclosure" / "Generative AI was not used to create..." / "AI Acknowledgement" appearing in the manuscript body. These are needed on submission forms and cover letters, but placing them in the body reads to a reviewer as a declaration that "this was written by AI" — an AI-generated signal.

Detection: grep -inE "artificial intelligence disclosure|generative ai was not used|ai acknowledg(e)?ment" manuscript.md → 0 lines (body). Fix strategy: Remove from the body → keep only on the submission form / cover letter. Exception only when the journal requires an in-body statement.

---

Response-Letter Patterns (R2R)

These three patterns are specific to response-to-reviewers (R2R) letters and editor cover letters. They rarely appear in manuscript bodies but dominate machine-drafted rebuttals, where the model narrates the editing process instead of the science. Apply them whenever the text under review is a response letter or cover letter. See the revise skill's references/r2r_voice.md for full before/after skeletons. Examples below are synthetic (a fictional deep-learning lung-nodule CT study).

Related (triage, not a fixed-string pattern): defensive over-elaboration — pre-emptive cross-reviewer lobbying ("Reviewers 2 and 3 also accepted this"), defensive meta-comments ("we confirm this is unchanged and not softened"), manufactured paragraphs answering a satisfied reviewer, or a separate cover letter on an R2+ round when its content belongs in the response-letter head — is a succinctness / round-discipline issue, not a regex pattern, and is most common on R2+ rounds. The normative guidance lives in the revise skill (Step 5 + Response-Letter Voice), its references/r2r_voice.md, and the rebuttal-letter-style rule; humanize only cross-references it (no hard fail, no detection regex added here).

Pattern 22: Editing-Mechanism / Change-Log Narration

The response prose narrates how the text was edited — what was added, where, how many phrases were swapped, which pass produced it — instead of stating what changed and why. The reviewer reads this as auto-generated and as checklist-clearing rather than scientific engagement.

Scope — this targets editing-mechanism narration only (avoid over-flagging): narrating a new analysis you ran ("we performed a sensitivity analysis restricted to one eye per patient, and the result held") is the science the reviewer asked for — never flag it. Likewise "we added a sentence to the Methods: '...'", structured Response: / Changes made: blocks, and Original → Revised before/after pairs are normal human conventions. The tell is the editing mechanism layered on top, not the act of describing, quoting, or locating a change.

Watch phrases (the tell): version-prefixed edits ("v2 Methods adds one sentence"), "we softened six phrases", "demoted the term at all N locations", "a grep-and-soften pass", "the v2 revision changes", bare "No further manuscript change" stubs, "reframed via the vocabulary cascade".

Detection (triage, not auto-fail): grep -inE "v[0-9]+ .*(add|demote|soften|reframe)|softened [0-9a-z]+ phrases|no further (manuscript )?change|grep-and-soften|vocabulary cascade|demoted .* at all [0-9]+" response_to_reviewers.md — review each hit and flag only genuine editing-mechanism narration, never analysis narration or quoted additions. Fix strategy: Delete the mechanism narration. State the substantive change and quote the new sentence; omit how it was found or made.

Pattern 23: Internal Line-Number Reference Tone

Pointing reviewers to internal draft/markdown line numbers ("(line 43)", "at lines 43, 49, 58, 60, 77–79"). These never match the reviewer's view of the revised manuscript and read as a diff log. Section names are what authors actually use.

Not a tell: a revised-manuscript page/line ("page 7, lines 177-178") when the letter states once that all page/line numbers refer to the revised manuscript the reviewer is reading. The tell is the internal draft line number that will not match the reviewer's PDF.

Detection: grep -inE "\(line [0-9]+|at lines? [0-9]+|line [0-9]+(–|-)[0-9]+" response_to_reviewers.md → review each hit; keep only those that demonstrably point to the revised manuscript. Fix strategy: Replace internal/draft line numbers with a section name; keep revised-manuscript page/line only if it matches what the reviewer sees.

Pattern 24: Tooling / Scaffolding Leak

Exposing internal tooling, verification mechanics, or draft scaffolding to the reviewer: grep counts, internal FIX/category codes, references to "the circulated bundle" or "the internal supplementary index", or § self-references carried into the response letter.

Detection (triage, except `§`): grep -inE "grep verification|grep-and-soften|circulated bundle|internal (supplementary )?index|FIX-[0-9]|vocabulary cascade|§" response_to_reviewers.md cover_letter.md — only § is a hard 0 (always a tell). The other terms can rarely be legitimate (e.g., "a cascade detector", "we grep-checked our own data pipeline"), so confirm each hit is an internal-tooling reference before flagging. Fix strategy: Strip confirmed references to internal tooling, verification passes, and draft scaffolding. The reviewer should see only the science and the substantive changes.

---

Section-Specific Application Guide

Abstract (ALL patterns)

The abstract is the most visible section and the most likely to be checked for AI writing. Apply all applicable patterns (1-21) with zero tolerance.

Common abstract issues:

• Pattern 1 in the Background sentence.
• Pattern 4 in the Results sentence ("demonstrated remarkable performance").
• Pattern 18 in the Conclusion sentence ("paves the way").

Introduction (Patterns 1, 2, 5, 7, 12)

• Opening paragraph: Check for significance inflation (Pattern 1) and AI vocabulary (Pattern 7).
• Literature review paragraph: Check for vague attributions (Pattern 5).
• Gap statement: Check for notability claims (Pattern 2).

Methods (Patterns 8, 16)

• Methods should be the most straightforward section. Watch for copula avoidance ("serves as

the reference standard" instead of "is the reference standard") and filler phrases.

Results (Patterns 3, 4, 6, 10, 11)

• After every statistical result, check for appended -ing clauses (Pattern 3).
• Check for promotional adjectives before numbers (Pattern 4).
• Check for synonym cycling of key terms (Pattern 11).

Discussion (Patterns 1, 5, 6, 17, 18)

• First paragraph: Check for significance inflation (Pattern 1).
• Comparison paragraphs: Check for vague attributions (Pattern 5).
• Limitations: Check for formulaic language (Pattern 6).
• Final paragraph: Check for generic conclusions (Pattern 18) and excessive hedging (Pattern 17).

Conclusion (Patterns 1, 18)

• The conclusion should be 1-3 sentences stating the main finding and its specific implication.
• No significance inflation, no generic optimism.

Response to Reviewers / Cover letter (Patterns 22, 23, 24 — plus 13, 16, 19)

• Response letters and cover letters are reviewer-facing argument, not change-logs. The

dominant AI-tell here is the editing-mechanism register (Pattern 22), internal draft line-number pointers (Pattern 23), and tooling/scaffolding leaks (Pattern 24).

• Also sweep for em dashes (13), filler phrases (16), and § markers (19).
• Patterns 1-18 still apply where relevant, but 22-24 are the highest-yield checks for this

document type.

---

Quick Checklist (Pre-Submission)

Run this checklist on the final manuscript before submission:

• [ ] No "Additionally" / "Furthermore" / "Moreover" at sentence start (allow max 1 total)
• [ ] No "pivotal" / "crucial" / "landscape" / "delve" / "utilize" / "leverage"
• [ ] No "-ing" participial phrases appended to statistical results
• [ ] No "serves as" / "stands as" / "represents a" (use "is")
• [ ] No vague "Studies have shown" without a specific citation
• [ ] No "not only X but also Y" (allow max 1 total)
• [ ] No generic positive conclusions ("paves the way," "the future looks bright")
• [ ] Consistent terminology throughout (no synonym cycling)
• [ ] Em dashes: fewer than 2 per 1000 words
• [ ] Hedging calibrated to evidence level (no stacked hedges)
• [ ] Filler phrases eliminated ("It is important to note that," "In order to")
• [ ] Straight quotation marks (not curly)
• [ ] Sentence case in headings (unless journal requires title case)
• [ ] AI pattern density < 2.0 per 1000 words
• [ ] § (section sign): 0 occurrences (Pattern 19) — grep -c "§" = 0
• [ ] (Methods §X) / (Results §Y) self-reference: 0 occurrences (Pattern 20)
• [ ] AI Disclosure boilerplate in body: 0 occurrences (Pattern 21) — cover letter / submission form only

Response letters / cover letters only (Patterns 22-24)

• [ ] No editing-mechanism narration (Pattern 22) — "v2 adds one sentence", "softened N phrases", "No further manuscript change" (analysis narration and quoted additions are fine)
• [ ] No internal draft line-number pointers (Pattern 23) — "(line NN)", "at lines N, M" (revised-manuscript page/line is fine)
• [ ] No tooling/scaffolding leak (Pattern 24) — "grep", "FIX-N", "circulated bundle", "internal index", §