AI Rate

Overview

Use this skill to diagnose and rewrite text so it sounds less formulaic, less generic, and more plausibly human-written. Treat "under 20%" as an editorial target, not a guaranteed detector score, because AI detectors vary and can be unreliable.

Workflow

1. Preserve the user's intent, facts, stance, and required format before changing style.
2. Identify the main AI signals:
• evenly balanced sentence lengths
• generic transitions such as "firstly", "moreover", "in conclusion", "值得注意的是"
• abstract claims without concrete detail
• overly complete structure that feels like an essay template
• repeated modifier patterns, parallel phrasing, or polished but bloodless tone
• excessive hedging, summary language, or encyclopedic neutrality
3. Rewrite with human variance:
• vary sentence length and rhythm
• keep a few imperfect but natural turns of phrase
• add concrete context only when it is supported by the source or clearly requested
• replace broad claims with specific observations
• use contractions or colloquial phrasing when appropriate for English
• for Chinese, prefer natural spoken-written hybrids over official-report phrasing unless the genre requires formality
4. Run a self-check and revise again if the text still feels too smooth, symmetrical, generic, or mechanically structured.
5. When useful, return both a short diagnosis and the revised version. For simple user requests, return only the revised text.

Rewrite Rules

Keep meaning first. Do not fabricate personal experience, citations, numbers, quotes, or events to make a text sound human. If the draft lacks detail, either keep it general or ask for real details when accuracy matters.

Prefer small asymmetries over heavy "humanization." A strong rewrite should feel edited by a person, not deliberately messy.

Avoid phrases that announce process, such as "下面是降低AI率后的版本" unless the user asks for explanation. Avoid overusing emotional particles or slang; use them only when they match the text's setting.

For academic, professional, legal, medical, financial, or public-facing text, keep the register credible and do not remove necessary precision just to lower AI feel.

AI-Rate Self-Check

Before finalizing, estimate whether the rewrite is likely below the user's target using this rubric:

• Under 20% target: varied rhythm, specific wording, non-template structure, no obvious AI transitions, tone fits the author and context.
• 20-40% risk: mostly natural but still too balanced, too generic, or too neatly summarized.
• Above 40% risk: essay-like structure, predictable transitions, generic claims, or uniform sentence cadence.

If the estimate is above 20%, revise once more before responding.

Output Patterns

When the user asks only to rewrite, output the rewritten text directly.

When the user asks for checking plus rewriting, use:

• Estimated AI-rate risk: concise percentage range or "likely below 20%"
• Main issues: 2-4 short points
• Revised version: the final text

Be transparent when a detector-specific guarantee is impossible. Say the text is optimized toward the target rather than claiming a guaranteed score.