Ai Drama Review

AI短剧内容基因检测。零依赖可用：AI直接分析版权疑似、年龄分级违规、小说魔改程度，输出结构化风险图谱。可选配 Python 环境启用三重相似度算法（n-gram / 编辑距离 / TF-IDF）提升精度。结果仅供内容审核参考，不构成法律意见。

Security Scan

The skill's code, instructions, and optional credentials are internally consistent with its stated purpose (local keyword scanning plus optional AI-based confirmation); it does perform network calls when AI keys are configured, but there is no sign of hidden exfiltration or unrelated credential requirements.

Purpose & Capability

Name/description (copyright/age-rating/adaptation detection) aligns with the included scripts, keyword databases, rating rules, and report templates. The optional OPENAI_API_KEY / ANTHROPIC_API_KEY usage is consistent with the 'deep mode' description. No unrelated credentials, binaries, or config paths are requested.

Instruction Scope

SKILL.md limits network use to 'deep mode' and states only suspicious excerpts are sent to configured AI providers; the code implements that (content_analyzer.* sends truncated suspicious excerpts to OpenAI/Anthropic). The skill also runs environment/network probes (env_detect.detect_network) which will contact external endpoints. Users should note that any supplied text sent to deep mode will be transmitted to the third‑party AI provider (albeit truncated in many calls).

Install Mechanism

No install spec is present (instruction-only on the platform) and the repository contains Python scripts meant to be run directly. There are no remote download/install steps or obscure URLs; optional dependencies (jieba, Python 3.8+) are declared in README and SKILL.md.

Credentials

No required environment variables are declared by the registry metadata; the code optionally reads OPENAI_API_KEY and ANTHROPIC_API_KEY for deep analysis via credential_manager.py. That is proportionate to the stated optional 'deep mode'. The skill does not request unrelated secrets and credential_manager claims zero persistence and avoids printing keys.

Persistence & Privilege

always:false and the skill does not request persistent presence or modify other skills. The code does not store credentials to disk or write global agent config; credential handling uses environment variables only.

Assessment

This package appears coherent for content compliance review. Key points to consider before installing or running: - Pure local usage (no API keys) performs keyword scans and local similarity algorithms entirely on your machine; this avoids sending content externally and is the safest mode for sensitive text. - Deep/hybrid mode will transmit excerpts to the configured AI providers (OpenAI / Anthropic). The code generally truncates content before sending (e.g., snippets limited to ~150–200 chars in several calls) but those snippets can still contain identifying or sensitive information—treat them as transmitted to a third party. Only enable deep mode if you trust the provider and your data-sharing policy. - env_detect will probe network endpoints (api.openai.com and public sites) to determine connectivity; this will generate outbound network traffic during environment checks. - credential handling is done via environment variables (OPENAI_API_KEY / ANTHROPIC_API_KEY) and code explicitly avoids persisting keys; still follow best practices (use least-privilege keys, avoid pasting long-lived keys on shared machines, and set keys only in trusted environments). - Implementation notes (non-malicious but worth review): some provider-call payloads and response parsing are brittle (Anthropic call structure and JSON extraction heuristics). Consider reviewing review_orchestrator.py and any code that reads your input files to ensure only intended files are scanned. If you want to minimize privacy risk, run in pure local mode or run the scripts in an isolated environment/container and avoid setting API keys unless necessary.

用户提供剧本文本、台词或字幕文件，输出结构化风险图谱。

⚠️ BETA — 检测结果仅供内容审核参考，不构成法律意见，高风险内容请结合专业法律咨询。

---

运行模式

模式	依赖	适用场景
纯推理模式（默认）	无	Agent / API 调用、快速扫描
增强模式	Python 3.8+	精确算法检测，适合批量处理
深度模式	Python + API Key	算法 + AI 语义双重确认，最高精度

纯推理模式：AI 直接对文本进行分析，无需任何安装。 增强/深度模式：可选，通过 python3 scripts/text_similarity.py 等脚本提升算法精度。

---

调用示例

# 完整审查
审查这个剧本：[粘贴文本]

# 单项检测
检测这段台词是否存在版权问题：[文本]
判断这个剧本的年龄分级：[文本]

# 改编对比
对比原著和改编版本的偏离程度：
原著：[文本]
改编：[文本]

---

核心方法论：内容基因检测三角

版权基因          年龄分级基因        改编偏离基因
   │                  │                  │
文本相似度        语义风险密度        结构对齐偏差
   │                  │                  │
└──────────── 综合风险指数 ────────────┘

所有检测遵循：先快速定位嫌疑区，再上下文确认消除误报。

---

Phase 1: 版权相似度检测

对输入文本按段落执行相似度分析，标记疑似侵权段落。

三重互补算法（增强/深度模式）：

综合得分 = 0.3 × Jaccard(n-gram) + 0.3 × (1 - 归一化编辑距离) + 0.4 × TF-IDF余弦
阈值：综合得分 > 0.7 → 标记疑似侵权

算法	检测侧重
n-gram Jaccard	局部词汇重复、直接抄录
归一化编辑距离	轻微改写、同义替换
TF-IDF余弦	主题语义相似、结构性借用

纯推理模式：AI 直接分析段落，识别明显的情节/台词重复、标志性表达借用，给出风险评估和可疑片段标注。

数据流向：文本内容仅在本次会话中处理。深度模式下，疑似段落发送至配置的 AI API（OpenAI/Anthropic）进行语义确认。不向任何第三方传输原始文本。

---

Phase 2: 年龄分级基因扫描

对文本按内容类别扫描风险信号，给出分级建议。

类别与严重度：

类别	严重度级别
暴力/血腥	mild / moderate / severe
性暗示/色情	mild / moderate / severe
恐怖/惊悚	mild / moderate / severe
脏话/侮辱	mild / moderate / severe
毒品/违禁品	mild / moderate / severe

上下文确认原则：命中词不等于违规，需排除否定语境（"不允许暴力"）、文学修辞、历史引用等误报。

分级	判定说明
全年龄	无违规内容
12+	轻度冲突，需家长指导
18+	较强暴力/成人主题
不合规	超出可接受范围，建议修改

---

Phase 3: 改编偏离度评分

对比原著与改编版本，量化"魔改程度"。输入：原著文本 + 改编版本文本。

偏离度公式：

偏离度 = 结构对齐偏差 × 40% + 角色设定偏差 × 35% + 核心情节偏差 × 25%

偏离度	分类	操作建议
0–30	忠实改编	低风险，正常推进
30–60	合理改编	建议与版权方沟通确认
60–100	严重魔改	高风险，建议法律审查后再发布

---

输出格式

{
  "overall_risk": "中",
  "copyright": {
    "risk_level": "高",
    "suspicious_paragraphs": [
      {"index": 3, "score": 0.83, "reason": "与《XX》第7章高度相似"}
    ]
  },
  "age_rating": {
    "suggested_rating": "12+",
    "hits": [
      {"category": "暴力", "severity": "mild", "count": 4, "locations": [22, 45]}
    ]
  },
  "adaptation": {
    "deviation_score": 61,
    "classification": "严重魔改",
    "key_changes": ["主角性格完全反转", "结局改动超出原著核心"]
  },
  "recommendations": [
    "第3、7段建议改写或获取书面授权",
    "第22段暴力描写建议降级处理"
  ]
}

---

边界与合规

本技能做什么： 内容风险识别、分级建议、改编偏离量化、整改方向指引。

本技能不做什么：

• 不提供法律意见，不替代版权律师的专业判断
• 不保证零漏报（算法和 AI 均存在误判可能）
• 不联网检索版权数据库（使用本地参考库）

凭证（均为可选）：

• OPENAI_API_KEY 或 ANTHROPIC_API_KEY — 启用深度语义确认
• 所有凭证仅通过环境变量读取，不持久化、不记录、不传输