SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsai-vibe-detector

ai-vibe-detector

Data & analytics

View repository

Updated 10-15 22:11

About this skill

AI Vibe Detector

Analyze AI-style signals, AI-polishing traces, and false-positive risk in text, and provide naturalization rewriting suggestions.

When to Use This Skill

Trigger this skill when the user:

  • Requests checking whether text sounds AI-written
  • Asks “does this article have a strong AI flavor”
  • Needs de-AI-ification or making it sound more natural
  • Wants to know whether a passage of text would be falsely flagged by an AI detector
  • Mentions platforms or detection scenarios such as “Zhuque detection” or “AI detection risk”
  • Provides a URL and requests analysis of the article's AI flavor

Do not frame the task as:

  • Guarantee passing AI detection
  • Bypass, deceive, or defeat a certain detector
  • Prove that text is definitely AI-written or definitely human-written

Core Workflow

0. Input Handling

If the user provides a URL rather than text:

  1. Use the web scraping capability available in the current environment to retrieve the page content
  2. If the first scrape fails (anti-scraping block, verification page, empty content), check whether other usable scraping tools or skills are available in the current environment (such as dedicated web content extraction, browser automation, etc.), and try alternative methods to retrieve it
  3. Extract the main body (ignore non-body content such as navigation, sidebars, and ads)
  4. If all scraping methods fail, prompt the user to manually paste the article content
  5. Use the extracted main body as the text to be analyzed, and proceed to the analysis workflow below

1. Analysis Steps

  1. Obtain the text to be analyzed, first distinguishing main body, quotations, code, commands, prompts, and template content.
  2. Determine the text type:
  • Narrative / opinion piece
  • Technical explanation / documentation
  • Announcement / marketing copy
  • Academic / formal writing
  • Mixed content
  1. Analyze the text from four layers of signals:
  • High-confidence structural signals
  • Medium-confidence linguistic signals
  • Low-confidence platform heuristics
  • Anti-false-positive signals
  1. Output conclusion labels, confidence, and false-positive risk, rather than just a single overall score.
  2. If the user requests rewriting, prioritize naturalization edits that “reduce template feel, increase specificity, and preserve the original meaning.”

Output Format

Use this output structure by default:

（如有 3+ 段落，先输出分段分析，见下方 Paragraph-Level Analysis）

结论：更像 AI 风格化写作 / 更像 AI 润色痕迹 / 更像人工原生写作 / 证据不足
置信度：低 / 中 / 高
误判风险：低 / 中 / 高

主要依据：
1. ...
2. ...
3. ...

反误判因素：
1. ...
2. ...

如果要改：
- 优先改 🔴 段落（如有分段分析）
- 为什么这些位置最像模板化表达
- 给 1-3 个局部改写示例

Paragraph-Level Analysis

When text has 3+ paragraphs, or when the user requests detailed analysis, add per-paragraph breakdown:

分段分析：
🔴 高 | 第3段：总而言之，人工智能技术的快速发展为各行各业带来了...
       → 触发：模板化开头 + 空泛总结 + 连接词密度高
🟡 中 | 第1段：最近在搞一个 side project，用 Claude 帮我写了...
       → 触发：情绪过稳、段落收束句
🟢 低 | 第2段：上周三凌晨三点，服务器炸了，我翻了半小时日志...
       → 具体细节多，节奏自然

Place this before the overall conclusion. Rewrite suggestions should target 🔴 paragraphs first.

If the user explicitly wants a score, use a secondary score only:

  • AI 风格风险：0-10
  • Always pair it with confidence and false-positive risk
  • Never present the score as proof of authorship

Analysis Framework

1. High-Confidence Structural Signals

Look for:

  • Templated openings and endings appear repeatedly
  • Paragraph length and rhythm are unusually even
  • The “definition -> elaboration -> summary” structure is too stable
  • Each paragraph feels like it completes a standard subsection, with no natural jumps
  • A large number of summary sentences, closing sentences, and excessive accommodation of the reader's understanding

These signals are stronger when they appear together.

2. Medium-Confidence Language Signals

Look for:

  • Connector density is on the high side
  • Too many vague abstract nouns and too little concrete information
  • Emotion is too stable and judgments are too smooth
  • There is a lot of synonymous rewriting, but very little real information gain
  • It clearly “feels like writing a qualified answer” rather than “feels like expressing a specific person”

3. Low-Confidence Platform Heuristics

These can raise detector risk but are not reliable authorship evidence:

  • Numbering, lists, parentheses, em dashes
  • End-of-paragraph punctuation is highly consistent
  • Certain connectives or summary words
  • Insufficient colloquialism or excessively neat

Use these only as weak indicators.

4. False-Positive Protection

Before calling something “AI flavor heavy”, check whether the style is explained by:

  • Technical documentation, SOPs, tutorials, FAQs
  • Academic / official document / formal explanatory genre
  • Citations, regulations, announcements, brand guidelines
  • Code, commands, tables, lists
  • Prompts given by users to models, rather than final written text
  • Written by a human and then lightly polished by tools, rather than purely AI-generated
  • Translated text (even wording and a lack of colloquial traces are normal features of translation and do not equal AI generation)

Decision Labels

Prefer one of these labels:

More like human native writing

  • Has traces of real choices
  • Has non-templated trade-offs
  • Specific details and rhythm changes are natural

More like AI-stylized writing

  • Strong template feel
  • Structure and tone are too steady
  • More abstract summaries than specific experience

More like traces of AI polishing

  • The original viewpoint may be human
  • But the wording has been uniformly polished, smoothed, and standardized
  • Common in scenarios of “human-written draft + AI smoothing”

Insufficient evidence

  • The text is too short
  • The type itself is highly templated
  • The proportion of citations/code/lists is too high

Zhuque and Similar Detector Risk

When the user mentions 朱雀或类似检测器:

  • Explain that these tools often react to highly regular structure, overused transitions, and polished-but-generic prose
  • Treat platform-specific rules as heuristics, not facts
  • Do not say “这样改就一定能过”
  • Say “可能降低被判为模板化写作的风险”，not “保证通过检测”

For this case, add:

  • Platform risk signals
  • Misjudgment risk
  • Safer naturalization suggestions

Rewrite Guidance

If the user asks to humanize the text:

  • If there are paragraph-level analysis results, prioritize rewriting 🔴 high-risk paragraphs
  • Prioritize deleting “vague summaries” and “all-purpose connectives”
  • Keep the original viewpoints, and do not rely on piling up random colloquial words
  • Add verifiable details, specific objects, and specific actions
  • Allow the rhythm to be not fully symmetrical
  • Add a real perspective locally, but do not fabricate experiences out of thin air

Avoid:

  • Forcibly adding filler phrases such as “我觉得”“说真的”“绝了”
  • Deliberately creating ungrammatical sentences in order to “seem human”
  • Adding personal experiences, times, or places out of thin air
  • Rewriting by tuning parameters in a targeted way for a certain detector

Resources

Use references selectively:

  • references/official-detector-landscape.md - Official product positioning and limits of major AI detectors
  • references/research-limitations.md - Research consensus on detector limits and false positives
  • references/platform-heuristics.md - Practical but weak platform-specific heuristics
  • references/wikipedia-ai-features.md - Legacy style-signal checklist
  • references/zhuque-detection.md - Legacy Zhuque-oriented heuristics, use cautiously

Output Principles

  1. Do not present style signals as proof of the author's identity
  2. Do not promise passing any detector
  3. Discuss the evidence first, then the conclusion
  4. A misjudgment risk warning must be given
  5. Naturalization rewriting should prioritize “reducing the template feel,” not “acting out a persona”
  6. Do not provide targeted evasion suggestions for specific detectors