奇链 奇链
← Agent Skills 目录 / 文档处理
Sketchjar

detect-ai-text Skill

评估一份文档的文本是否由 AI 撰写,基于语言学特征判断,对非文本内容会诚实弃权。当用户询问某篇文章、报告、简历、投稿或论文是否为 AI 生成时使用——仅用于初步筛查,不能作为证据。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

AI-Text Detection

Estimate the probability that a document's prose was written by AI, with the specific linguistic tells and an honest abstention when the document isn't prose. Uses the Stipple API (free anonymous tier).

When to use

  • Educators screening student submissions
  • Publishers and platforms triaging inbound content
  • HR reviewing AI-drafted CVs (flag, don't reject)
  • Researchers checking source provenance

Instructions

  1. Get the document. URL, local file path (PDF, DOCX, TXT, MD), or raw text via --text.
  1. Run detection.
   curl -X POST https://www.stipple.sh/v1/detect-ai-text \
     -F "file=@essay.pdf" \
     -H "Authorization: Bearer $STIPPLE_API_KEY"

For raw text: POST JSON {"text": "..."} to the same endpoint.

  1. Interpret the response.
  • applicable: false — the document is not prose (forms, tables, scans, spreadsheets). Detection is deliberately refused rather than guessed. Report this and stop.
  • probability — model confidence (0–1), NOT a calibrated truth
  • lean — "ai" | "human" | "unsure"
  • tells[] — the specific phrases/patterns flagged (e.g. "It is important to note that", uniform sentence length, low burstiness)
  • reasoning — why the model reached its verdict
  • limitations — always present; the API states its own noise profile
  1. Report honestly. This measures style, not authenticity:

| Question | Tool |
|---|---|
| Was this written by AI? (style) | this skill |
| Is this document genuine/tampered? (forensics) | verify-document skill |

A human can write generically; an AI can write plainly. One triage signal, never a verdict.

Output format

AI-written probability: 0.87  (lean: ai)
prose ratio: 0.82

linguistic tells:
  - "It is important to note that, in today's fast-paced world" (stock phrase)
  - uniform sentence length across paragraphs
  - low burstiness; no authorial asides

reasoning: The text relies entirely on formulaic transition clichés...
limitations: The probability is the model's CONFIDENCE, not a calibrated truth.

Notes

  • Non-prose documents (forms, tables, payslips) get applicable: false — the engine refuses rather than guessing
  • Can false-flag templated/coached or non-native-English writing — always present as one signal alongside human review
  • Free anonymous tier works without a key; free key at https://www.stipple.sh