Human Touch Rewriter

The Ubiquitous "AI Smell" in Generated Text

Does your text drip with template connectors like "notably", "firstly/secondly/lastly", or "not only...but also..."? Do paragraphs invariably start with "However" or "Furthermore"? Does it habitually conclude with a summary or call-to-action? Combined with homogenous sentence structures (repetitive Subject-Verb-Object) and abstract jargon (empowerment, closed-loop, underlying logic), these hallmarks create the distinct "AI smell" of LLM-generated content—polished yet impersonal, like a report rather than human communication.

Deconstructing the "AI Smell": Diagnosis & Rewriting Logic

The Human Touch Rewriter (formerly Libai.Skill) systematically dismantles these artifacts to restore a natural human voice. Its approach is a structured "de-AI-ification" process, not mere synonym replacement:

1. Precise Diagnosis: It first scans the text across seven dimensions (Connector Density, Sentence Homogeneity, Emotional Temperature, Closing Pattern, Abstract Word Density, Paragraph Uniformity, etc.) to quantify AI traces. You can set the Rewrite Intensity to "Conservative", "Standard", or "Aggressive" based on context.
2. Iron-Clad Rules: The rewriting adheres to three inviolable principles: Stance Conservation (never alter facts, judgments, or sentiment), Preservation of Blood & Flesh (retain all specific details, colloquialisms, and personal tone), and a Tolerance for Length Variation (±10%).
3. The "Pattern-Breaking" Five-Layer Method: This is the core engine. It involves Deleting all AI connectors and elevated endings, Breaking symmetrical syntax and the "excessive 'de' character" problem, Downgrading formal vocabulary (e.g., utilize -> use), Infusing human anchors (first-person perspective, sensory details, subjective feelings), and boldly Restructuring—merging paragraphs, removing subheadings, or even inserting seemingly off-topic personal details. A Self-Quality-Check list with four key questions ("Does it affect information?", "Can it be filmed?", "Would a real person say this?", "Is the 'human touch' still there?") validates every edit.

Usage Scope & Key Considerations

This is a textual expression optimization skill. Please note:
- No Factual Verification: It only polishes style, not accuracy. For critical domains like medicine, law, or finance, human review of content is mandatory.
- Input Constraints: Processing text under 3000 characters is recommended. For longer documents, segmenting is advised.
- Two Operating Modes: The default B-mode delivers the final text and a concise change summary. For full diagnostic reports, revision diffs, and retrospectives, specify the A-mode.
- Resource Dependency: Performance relies on the bundled zh_rules.json core rule set and synonym library. The system will automatically degrade or declare its operating mode if resources are missing.