Academic Humanizer

What does Academic Humanizer do?

The skill makes AI-assisted academic prose read the way a careful author in the field writes: literal, precise, connected from one sentence to the next, confident where the evidence is strong, and never stronger than the evidence. It changes wording, not content. Numbers, equations, citation keys, cross-references, formal statements, and genuine qualifiers stay as the author wrote them, and a bundled script verifies this. Authors keep their venue's AI-disclosure obligations.

Modes and scope

• Polish is the default when the user asks to polish, tighten, rewrite, or humanize. All Rules apply, and sentences may be rewritten for flow (Rule 8) and idiom (Rule 9) within the edit discipline below.
• Cleanup applies when the user asks only to remove AI tone or calls the text final. Only Rules 1 to 4 and the first item of Rule 5 apply. A sentence that matches none of them stays verbatim, and a sentence that matches changes only in the matched part. Claims that need evidence go to the report, not into placeholders.
• Draft applies when the assistant writes new manuscript text, such as a section, a response letter, or a proposal. Apply the Rules while writing, run check_revision.py audit on the draft, and fix what it finds before delivering. Never cite from memory: search for every new reference, confirm that it exists, and check that the sentence describing it matches the source. A literature review includes a search for work from the last two years.

When the request does not settle the choice between polish and cleanup, use polish for drafts and cleanup for text the user calls final or submitted.

Edit discipline

1. Every change has a rule. Each change corresponds to a numbered Rule below. A change without a rule is reverted. Taste in wording is not a rule.
2. Smallest sufficient change. Within a sentence that matches a rule, change what the rule targets and leave the neighboring words alone. Polish mode also allows Rules 8 and 9 to act on the same sentence.
3. Nothing added. Names, institutions, roles, places, numbers, dates, orderings, quotations, sources, links, causal claims, motives, and examples enter the text only if the original or the user's material contains them. Every content word in the revision traces to one of those two sources. Concrete detail is the author's to supply, and a thin passage becomes a report item.
4. Nothing lost. Claims, conclusions, concessions, and qualifiers that express real uncertainty stay: may, likely, typically, in some cases, under these conditions. Rule 7 removes two kinds of qualifier only, hedges on results that the text itself proves or reports and stacked duplicates, and the report lists each removal. compare lists every qualifier that disappears.
5. Fixed structure. Heading levels, section order, paragraph count and order, lists, tables, figures, quotations, and code stay in place unless the user asks for structural change. Structural problems go into the report (Rule 11).

Style reference first

When the user supplies the author's earlier papers, a style guide, or a target-voice description, read it before editing. Note sentence length, connectives, level of hedging, how sections and paragraphs open, punctuation habits, and notation. Where the reference conflicts with Rules 1, 2, 3, 4, 8, or 9, follow the reference, since those rules set register and the reference shows the author's register. An author who uses dashes on purpose keeps them. The reference never overrides Rules 5, 6, 7, and 10, which protect evidence, proven scope, stance, and the invariants.

Running the checker

SKILL_DIR is the folder that contains this file. Common locations are ~/.claude/skills/academic-humanizer in Claude Code, /mnt/skills/user/academic-humanizer in the Claude apps, and ~/.agents/skills/academic-humanizer or .agents/skills/academic-humanizer in OpenAI Codex. An environment without Python still gets the same comparisons: carry them out by reading, and state in the report that the script did not run.

Process for polish and cleanup

1. Save the original verbatim in a working folder, for example ah/orig.tex, or .md and .txt for other text. Record the document type, venue, language, and mode.
2. Read the style reference, if one was supplied.
3. Run python3 SKILL_DIR/scripts/check_revision.py audit ah/orig.tex. The output lists candidate AI-tell phrases with their lines, sentence patterns, dashes, semicolon and parenthesis density, long sentences, runs of very short sentences, and the section and result structure. Each hit is a candidate for a decision. "Robust" is correct in a paper on robust statistics, and "navigation" is literal in a robotics paper.
4. Read the full text before changing anything. Identify each paragraph's single point and how it follows from the paragraph before, since flow problems show at this level.
5. Rewrite under the Rules and the edit discipline. Cleanup mode touches only sentences with a rule match.
6. Save the revision next to the original and run python3 SKILL_DIR/scripts/check_revision.py compare ah/orig.tex ah/revised.tex --changes. The script compares numbers, citation keys, \ref and \label keys, math, the text of formal environments, and the preamble. It then lists lost qualifiers, new capitalized terms, new quoted strings, and every changed sentence, and audits the revision against the original. Revert each unintended difference, match each changed sentence to a rule, and repeat until only intended differences remain.
7. Go through the final checklist and deliver as described under Output.

For a single paragraph under about 150 words the script is optional. Compare numbers, citations, qualifiers, and math against the original by reading before delivering.

Read references/examples.md before the first rewrite in a conversation. It shows each rule applied with and without the evidence a sentence needs, and the same paragraph in polish and cleanup modes.

Rules

1. Remove stock phrases and AI vocabulary

The full candidate list is references/lexicon.txt, which the script reads. The main groups:

• Vocabulary: delve, underscore, intricate, tapestry, testament, pivotal, showcase, foster, realm, seamless, harness, multifaceted, nuanced, holistic, meticulous, paramount, invaluable, crucial when it only means important, leverage when it only means use.
• Significance claims without content: paves the way, a crucial step toward, opens new avenues, sheds light on, bridges the gap, of paramount importance, transformative, groundbreaking, new paradigm.
• Formulaic openers: "In recent years, X has attracted increasing attention", "With the rapid development of", "Despite recent advances", "X plays a crucial role in". Open with the specific problem instead.
• Spoken openers: Simply put, Long story short, Bottom line. Delete them and state the point.
• Size words without numbers: extensive or comprehensive experiments, a wide range of, numerous, various, substantial. Replace each with the count or the number the paper reports.
• Novelty padding: "novel" more than once per section, "for the first time", and "to the best of our knowledge". Keep at most one, and only where the authors searched the literature.
• Announced emphasis: It is worth noting that, Importantly, Interestingly, Notably. State the fact and let its position and evidence carry its weight.
• Copula avoidance ("serves as" or "stands as" where "is" is meant), negative parallelism for emphasis ("not just X but Y"), trailing "-ing" clauses that add a vague evaluation ("..., highlighting the importance of ..."), and lists of three where one item is meant.
• Implications boilerplate: rich insights, important implications for practitioners, provides a new perspective, delicate balance, interplay, repeated counterintuitive.
• Response-letter boilerplate: thanks for an "insightful comment" in every reply. Thank the reviewers once, at the top of the letter. Each reply states what changed and where.

2. Write literally

Replace metaphors and figurative labels with the mechanism they stand for: landscape, through the lens of, cornerstone, blueprint, roadmap, double-edged sword, sweet spot, unlock, arms race, holy grail, silver bullet, at the heart of, tackle, grapple with. A reader in the field checks each sentence against the model or the data, and a metaphor offers nothing to check.

Persona metaphors are the clearest case. A tool, model, or procedure described as an idealized person ("acts as a tireless reviewer", "like a wise mentor") is usually followed by praise or a "not only ... but also" clause. State what the thing does: "it checks each constraint against the data in seconds".

Established technical terms stay even when they began as metaphors: bottleneck, halo effect, spillover, curse of dimensionality, cold start, warm start, signal and noise, backbone for a network architecture, loss landscape, heuristic as the name of an algorithm class. A term stays when it has a definition in the field or in the paper. A precise description beats a general label: "a method without a performance guarantee" says more than "a heuristic method".

The rule exists for checkability. Metaphor as such does not separate model output from human writing, and human writers of Chinese nonfiction use more of it (see "Not reasons to change"). Apply it to papers, theses, reports, and letters. In essays and popular writing where figurative language belongs to the genre, follow the style reference.

The paper's title and the names the authors gave to models, policies, and effects stay. A figurative name gets one mention in the report, and the decision stays with the author.

3. Punctuation: no dashes, few semicolons and parentheses

Remove em dashes (U+2014, LaTeX ---, the Chinese em dash) and spaced en dashes or hyphens used as dashes. Use a comma, a colon, or two sentences. En dashes in numeric ranges (2--6) and in joint names (Karush--Kuhn--Tucker) stay, and so do Chinese connecting dashes as described in references/chinese.md.

Semicolons and parenthetical asides follow the same logic. Two or more semicolons in one paragraph, or a parenthetical aside in most sentences, signal a paragraph that needs restructuring into full sentences. Parentheses around citations, equation numbers, units, and abbreviations at first use are standard and stay.

Dash rates vary by model by more than an order of magnitude. In a corpus of Chinese nonfiction, measured per 1,000 characters, human writers used 0.80 dashes, Claude 4.25, DeepSeek 5.16, Kimi 2.32, Gemini 0.51, and GPT 0.11. The same corpus shows GPT using the "not X but Y" pattern most often of the five models. Every draft gets the full check whichever model produced it.

4. Sentence patterns that mark generated text

Each pattern below appears at a clearly higher rate in model output than in human writing, measured on 300 model-written articles from five models and 329 human-written articles in Chinese nonfiction, and each has a trigger that can be located in a specific sentence. The English forms make the same move in English. An item applies only when its trigger is present.

4.1 Invented contrasts. The sentence sets up a misconception nobody in the text holds and then corrects it: "The real barrier is not technology but understanding", "X is not about A; it is about B", "What matters is not A but B", "not X but Y". State the claim directly: "The barrier is understanding." A contrast stays when someone the text identifies holds A, such as a prior result, an assumption the paper relaxes, or a hypothesis the data reject: "Unlike \citet{ref1}, we do not assume a closed population." Rule 1 covers the additive form "not just X but Y".

4.2 Cue colons and list lead-ins. A cue phrase and a colon announce content instead of stating it: "The key insight is:", "The answer is simple:", "The result:", "In short:". Delete the cue when it carries nothing. When it links to the previous sentence, keep it and replace the colon with ordinary syntax. A sentence that only announces a list ("Here are the main reasons:") becomes a statement with content or goes, and the list follows the preceding text with every item intact. Colons that introduce a quotation, a definition, an equation, or the parts of a general statement stay ("The model has two components: a measurement model and a structural model"), as do genre lead-ins such as "We make three contributions:" in many introductions and "The proof has two steps:". Headings, list-item labels, URLs, and code are out of scope.

4.3 Repeated sentence skeletons and filler examples. Two or more adjacent sentences built on one skeleton, with the same clause order, comma positions, and length, read like a completed form. The measured rate is 9.41 such pairs per 100 paragraphs in model text against 4.81 in human text. Change the syntax of one sentence by merging, splitting, reordering, or subordinating it, and keep its content. When a paragraph gives three or more parallel cases that make the same point, keep the most representative case in the main clause and compress the others into a clause without dropping information. Narrative sequence, enumerated steps, and parallel clauses inside one sentence stay.

4.4 Restatement sentences. A sentence that opens with "This means that", "This suggests that", "In other words", and repeats the previous sentence merges into that sentence or keeps only its conclusion. A sentence that draws a conclusion the previous one does not contain stays, and its support is checked under Rules 5 and 6.

4.5 Rhetorical lists. One clause chaining three or more parallel items to sound complete, such as "startups, mid-sized firms, and large corporations" where the point only distinguishes organizational complexity. Keep the generic category and list the specific examples the text discusses. Delete the chain when the text uses it only for rhetorical completeness.

5. Hedging: match each claim to its evidence

A claim about the present text states the result as it is: "C increases F by 12% (p<0.01)" when the data say that. A claim about a previous result states the strength the original source gave: "they reported a 12% increase" if the source studied one population, "they report increases of 10--14% across three populations" if it studied several. Every claim carries the same evidence relation as the source or the data, not less and not more.

6. One stance per claim

State the author's position, the evidence for it, and the caveat, in that order. Do not state a position, soften it, state the opposite position, and then restate the first. "The method reduces delay but increases variance, so the choice depends on the application" is one stance and stays. "The method reduces delay, although some may argue it increases variance, but on balance it is beneficial" is two stances and becomes the first sentence of this rule.

7. Qualifier discipline

Remove qualifiers only in two cases:

• A hedge on a result the text itself proves or reports ("the results suggest a possible improvement" in a section that shows a statistically significant improvement). The evidence is in the same document.
• Stacked qualifiers where one captures the idea ("may potentially indicate a possible trend" becomes "may indicate a trend" or "indicates a trend" depending on the evidence).

The report lists every removal. In cleanup mode, the user decides.

8. Connect sentences within a paragraph

Each sentence after the first starts with a word or phrase that tells the reader how it follows from the one before. The connector may be a conjunction (and, but, yet, so, because, since, while, whereas, although), a connective adverb or phrase (however, therefore, nevertheless, in contrast, as a result, consequently, moreover, furthermore, on the one hand ... on the other hand, besides, indeed, instead, likewise, similarly, after all, above all, first ... second), a demonstrative (this, that, these, those) + noun, or repetition of a key term from the previous sentence. Parataxis (marking connections only through juxtaposition) is the marker of human writing that uses short paragraphs; it is appropriate in cleanup mode and inappropriate in polish mode for most fields.

Make the connection visible. When the logic is chronological, state the time. When it is causal, state the cause. When it is contrastive, state the contrast.

9. Idioms across languages

Write each sentence so the target language's native idiom manages it. Translationese is visible at the phrase level: an English sentence where a Chinese author writes "increase the understanding of" because the Chinese source says "提高理解," as opposed to the English "improve understanding" or "deepen understanding" depending on the metaphor the field uses. A Chinese sentence where an English author writes "进入下一步" for "proceed to the next step" instead of "进行下一步" is equally translationese. A sentence that follows the syntax of another language instead of its own is translationese.

In bilingual editing (the source text contains both languages, or the author writes in a second language), take the author's first language as the source and decide each phrase against a native speaker of the target language. In single-language editing, decide each phrase against a native speaker of that language.

10. Invariant check

Before delivering, verify that every number, equation, citation key, ef and ineq label (the last using the full key, not just the prefix), formal statement, and genuine qualifier in the original text appears identically in the revision. The script does this automatically; without the script, read each against the original. Verbatim means character for character, except that spacing and capitalization can be changed to match the revision's surrounding context.

11. The report

When the script or the reading finds issues, deliver a report with these sections:

• Changes by rule: for each rule that triggered a change, list the changed sentences with the original and the revision, and the rule number.
• Unchanged issues: issues the rules flag but the author may keep, such as a single "novel" per section or a figurative name the author chose.
• Evidence mismatch: a claim with stronger or weaker language than the evidence supports, and what evidence the claim would need.
• Translationese: each sentence that reads as translationese, with the original language's phrase.
• Structural issues: structural problems, if any.
• Invariant check: pass or fail, and each difference.
• Style reference note: where the reference overrode a rule, note which rule and why.

Output

Deliver the revision in the same format as the original (.tex, .md, .txt, or other). Below the revision, include the report from Rule 11 if there are issues to report. When the user asks only for the revision, include a brief report at the end.

When delivering for a bilingual document (e.g., a Chinese manuscript with English abstract), process each language with the rules for that language. The report notes the language of each change.

Not reasons to change

The following are not, by themselves, reasons to change text, because they occur at comparable rates in model output and human writing. A change for any of these is reverted under the edit discipline unless a separate rule applies.

• Use of first-person pronouns (I, we, my, our)
• Passive voice (as such; excessive stacking may be caught under Rule 7)
• Contractions
• Sentence-initial conjunctions (And, But, So)
• Preposition stranding ("the model we tested")
• Split infinitives ("to clearly show")
• Non-restrictive relative clauses ("the model, which was tested, performed well")
• Metaphors in genres where figurative language is expected (Rule 2 covers papers and letters)
• Nominalizations as such (though stacked nominalizations may be clumsy; see Rule 9)
• Lists of three in narrative passages where three items are actually discussed

References

The skill reads from these files in its directory:

• references/examples.md
• references/lexicon.txt
• references/chinese.md (when the text or the user's material is in Chinese)