Academic Humanizer

The academic writing problem

Academic drafts can sound too much like model output: even clause rhythm, padded abstractions, mechanical signposting, repeated three-item lists, and report-shell phrases such as this paper examines. The goal of humanizer-academic is not to make prose casual or playful. It rewrites English, Chinese, and mixed-language scholarly text into restrained, specific, and committed academic prose.

How the skill works

The core protocol is: subtract AI signals, add human texture, then verify register and facts.

• Lexical layer: reduce inflated vocabulary, promotional adjectives, vague attribution, chatbot residue, and stacked empty hedging.
• Structural layer: lower mechanical triads, overused connectors, this paper examines style meta-sentences, and reflexive not-X-but-Y balance.
• Statistical layer: vary sentence and paragraph length on purpose, raising burstiness without adding noise.
• Texture layer: replace abstract summaries with numbers, cases, or mechanisms already present in the source, while preserving calibrated confidence.
• Detector: scripts/detect_ai_signals.py returns a three-layer diagnostic signal map for before/after comparison; it does not rewrite and is not the success oracle.

Boundaries

Use it for abstracts, literature reviews, research reports, policy papers, and other scholarly or professional writing. Do not use it for poetry, speeches, fiction dialogue, satire, or casual blurbs. A hard constraint is zero net-new facts: citations, numbers, dates, named entities, and claim strength must trace back to the input. Detector deltas are diagnostic only; quality is judged by blind-evaluator scores, register preservation, and fact-invention rate.