Academic Humanizer

Academic Humanizer

Remove the template voice, not the scholarship. Make academic prose more specific, natural, and recognizably the author's while preserving the research exactly.

This Skill improves writing quality. It does not certify that text is human-written, optimize for an AI detector, or remove an author's obligation to disclose AI assistance under the relevant journal, institution, or funder policy.

Load only what is needed

• Always read references/preservation-and-evidence.md, references/pattern-catalog.md, and references/output-contract.md.
• Read references/voice-calibration.md only when the author provides writing samples or explicitly asks to match their voice.
• When source and revision are available as local text, use scripts/check_invariants.py as a deterministic first check. A passing result never replaces the semantic audit.

Use this Skill when

The user asks to:

• remove “AI味”, AI-like phrasing, template language, or generic LLM prose from academic text;
• make a paper sound less mechanical, less inflated, or more like the author;
• audit an abstract, paper section, thesis passage, rebuttal, or reviewer response for recurring AI-writing patterns;
• calibrate a draft against the author's prior academic writing.

For ordinary Chinese-to-English translation, grammar correction, or general publication polishing without a de-templating or voice-matching goal, sci-ssci-polishing is the better entry point.

Inputs

Required:

• the passage or document to audit or revise.

Useful but optional:

• language, field, section type, and target venue;
• one or more genuine writing samples from the author;
• protected terminology or phrases;
• desired intensity: audit only, standard, or strong.

Do not block on optional context. Without author samples, use a neutral, precise, venue-appropriate academic voice and state that personal voice matching was not performed.

Workflow

1. Classify the task

Identify:

• language: Chinese, English, or mixed;
• scope: passage, paragraph set, section, or full document;
• rhetorical function: Abstract, Introduction, Methods, Results, Discussion, Conclusion, literature review, rebuttal, or mixed;
• requested mode: audit only, standard rewrite, strong rewrite, or voice matching.

Infer missing context conservatively. Do not force a paper, thesis, rebuttal, and grant proposal into the same register.

2. Freeze the scholarship

Before editing, create a private preservation ledger covering:

• numbers, statistics, units, equations, variables, citations, and figure or table references;
• named methods, datasets, instruments, entities, and author-supplied terminology;
• actors, populations, comparisons, directions, time order, uncertainty, limitations, causal strength, and conclusion scope.

Follow references/preservation-and-evidence.md. If the source is ambiguous, keep the narrowest defensible meaning and add an author query.

3. Audit patterns before rewriting

Use references/pattern-catalog.md to locate patterns and explain why they weaken the passage.

Treat the catalog as a set of editorial diagnostics, not a detector:

• do not label a passage AI-written;
• do not flag a single word or punctuation mark as proof;
• look for repeated cues or clusters across at least two categories;
• protect legitimate disciplinary conventions and phrases that carry real information.

For each material finding, record the location, pattern category, why it weakens the text, and the smallest useful fix.

4. Calibrate voice when evidence is available

If the user provides prior writing, follow references/voice-calibration.md. Match stable habits such as information order, sentence rhythm, connective density, hedging, citation integration, and use of “we”. Do not copy distinctive phrases or introduce claims from the sample.