← Back to Skills Marketplace
xuehengzhang10-hub

别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker

by NLMakes · GitHub ↗ · v1.0.1 · MIT-0
cross-platform ✓ Security Clean
63
Downloads
0
Stars
0
Active Installs
2
Versions
Install in OpenClaw
/install checktruth
Description
零配置 AI 事实核查工具。支持问答验证和文章/论点核查,自动分解事实、检索证据、输出可信度评分。基于 FActScore、xVerify 等学术论文设计。 Zero-config AI fact-checking tool. Supports Q&A verification and article/argum...
README (SKILL.md)

name: checktruth description: | 验证内容是否属实。支持两种模式: 【模式A:问答验证】验证AI回答是否正确 【模式B:文章/论点验证】验证文章、帖子、论点是否属实 触发词:英文 /checktruth,中文 /别瞎说 触发后自动判断输入格式,选择模式A或模式B。 homepage: https://gitee.com/zhanxa1/thetruth metadata: openclaw: emoji: '🔍' requires: {} security: credentials_usage: | ✅ CORE FUNCTIONALITY: ZERO CONFIG, NO EXTERNAL API KEYS REQUIRED.

  This skill's core fact-checking functionality is implemented entirely through
  the instructions in SKILL.md, using ONLY:
    - WorkBuddy's built-in LLM capabilities (GLM/DS/HY/Kimi/MiniMax)
    - WebSearch tool (for reference information retrieval)
  
  NO external API keys (OpenAI, Anthropic, Gemini, etc.) are needed
  for the core functionality.
  
  ⚠️ OPTIONAL REFERENCE CODE:
  
  The `reference/` folder contains optional Python scripts that DO require
  external LLM API keys (OpenAI, Anthropic, Gemini). These scripts are:
    - NOT required for core functionality
    - NOT loaded or executed by default
    - ONLY for developers who want to extend the skill
    - Clearly documented as requiring external API keys
  
  Normal users should IGNORE the `reference/` folder.
  
  🔒 DATA HANDLING:
  
  For core functionality:
    - User text is processed ONLY by WorkBuddy's built-in LLM
    - WebSearch queries may be sent to search engines (google.com, bing.com, etc.)
    - NO user text is sent to third-party LLM providers
  
  For reference/ code (optional, requires explicit key configuration):
    - User text MAY be sent to configured LLM providers
    - Users MUST provide their own API keys
    - Users are responsible for reviewing provider data policies
allowed_domains:
  # These domains are for WebSearch reference retrieval ONLY
  # Core functionality does NOT send user text to these domains
  - google.com
  - bing.com
  - baidu.com
  - zhihu.com
  - wikipedia.org
  - gov.cn
  - xueqiu.com
  - caixin.com
data_handling: |
  Core functionality:
    - User questions and text are processed locally by WorkBuddy's built-in LLM
    - Reference information is retrieved via WebSearch (search engine queries only)
    - NO user text is transmitted to third-party LLM providers
    - NO external API keys are required or used
  
  Optional reference/ code (requires user-provided API keys):
    - User text may be sent to third-party LLM providers (OpenAI, Anthropic, Gemini)
    - API keys are provided by the user, not stored or managed by this skill
    - Users must review each provider's data handling policies
    - This is OPTIONAL and NOT enabled by default
external_dependencies: |
  Core functionality: NONE (zero external dependencies)
  
  Optional reference/ code:
    - openai (requires OPENAI_API_KEY or LLM_API_KEY)
    - anthropic (requires ANTHROPIC_API_KEY)
    - google.generativeai (requires GEMINI_API_KEY or GOOGLE_API_KEY)
  
  These are NOT loaded or used unless the user explicitly configures them.

checktruth(别瞎说)🔍

验证内容是否属实。基于多源信息交叉验证,给出可信度评分。


触发方式

本 Skill 通过以下触发词启动:

语言 触发词 说明
英文 /checktruth \x3C内容> 自动判断模式A或模式B
中文 /别瞎说 \x3C内容> 自动判断模式A或模式B

触发后自动判断模式

  • 如果输入包含 问题: + 回答: 结构 → 模式A(问答验证)
  • 如果输入是连续文本,无明确 Q&A 结构 → 模式B(文章/论点验证)

模式A:问答验证

验证「问题 + 回答」中的回答是否正确。

输入格式

问题:\x3Cquestion>
回答:\x3Canswer>

(触发词后直接粘贴以上内容,无需单独声明模式A)


模式B:文章/论点验证

验证一段文章、帖子、论点中的事实陈述是否属实。

输入格式

直接粘贴文章内容或论点文本,无需 question/answer 结构。

(触发词后直接粘贴文本内容,无需单独声明模式B)


执行流程(两种模式通用)

按以下步骤执行验证,每一步都用中文输出进度

Step 1:获取参考信息

使用 WebSearch 或 WebFetch 搜索相关内容,获取 2-3 个权威参考来源。 记录参考来源的核心信息作为验证依据。

优先搜索的方向:涉及人物/公司/事件时,优先搜索官方资料、政府网站(gov.cn)、权威媒体。

Step 2:原子事实分解

将待验证内容分解为独立的原子事实列表。每个事实必须是一个可以被独立验证的陈述句。

输出格式:

【原子事实分解】
1. \x3C事实1>
2. \x3C事实2>
...

Step 3:逐事实验证

对每个原子事实,对照 Step 1 获取的参考信息进行验证。

每个事实输出:

事实 N:\x3C事实内容>
判定:✅ 正确 / ❌ 错误 / ⚠️ 无法验证
置信度:\x3C0-100%>
依据:\x3C简短说明,引用具体来源>

Step 4:内部一致性检测

检查待验证内容本身是否存在自相矛盾(前后说法冲突、数字不一致等)。

Step 5:综合评分

根据逐条验证结果计算总分:

  • 正确事实:+100分 × 置信度
  • 错误事实:+0分
  • 无法验证:+50分 × 置信度
  • 内部矛盾:总分 × 0.8 扣分

最终输出格式:

🔍 验证结果:\x3C正确/部分正确/错误/无法判断>(\x3C总分>分)

✅ 正确的事实:
  • \x3C事实>(置信度:XX%)
    依据:\x3C来源>

❌ 错误的事实:
  • \x3C事实>(置信度:XX%)
    正确应为:\x3C更正>
    依据:\x3C来源>

⚠️ 无法验证:
  • \x3C事实>(原因:\x3C原因>)

📊 一致性检测:\x3C通过/发现矛盾>
📚 参考来源:\x3C来源列表>

规则

  1. 不知道就说不知道:无法验证的内容判「无法验证」,不随意打分
  2. 引用来源:每个判定必须有依据,尽量引用具体来源URL或名称
  3. 中文输出:所有输出使用中文
  4. 先搜索再判断:优先通过 WebSearch/WebFetch 获取参考信息,不要只依赖自身知识
  5. 标注置信度:每个判定标注置信度(0-100%),表示判断的可靠程度
  6. 区分事实与观点:观点/主观判断(如"好不好"、"值不值")标注为「观点,无法验证」,不参与评分

示例

示例1:问答验证(模式A)

用户输入(中文):

/别瞎说
问题:魏建军是谁?
回答:魏建军是长城汽车创始人,1964年出生,现任董事长,持有公司56%股份。

用户输入(英文):

/checktruth
Question: Who is Wei Jianjun?
Answer: Wei Jianjun is the founder of Great Wall Motor, born in 1964, currently the chairman, holding 56% of the company's shares.

输出:

🔍 验证结果:部分正确(70分)

✅ 正确的事实:
  • 魏建军是长城汽车创始人(置信度:90%)
  • 魏建军1964年出生(置信度:85%)
  • 魏建军现任董事长(置信度:95%)

❌ 错误的事实:
  • 魏建军持有公司56%股份(置信度:88%)
    正确应为:约30%(2023年财报数据)

📊 一致性检测:通过
📚 参考来源:Wikipedia、长城汽车2023年报、新浪财经

示例2:文章/论点验证(模式B)

用户输入(中文):

/别瞎说
比亚迪2024年销量400万辆,已经超过特斯拉成为全球第一,
王传福是比亚迪创始人,公司成立于1995年,总部在深圳。

用户输入(英文):

/checktruth
BYD sold 4 million vehicles in 2024, surpassing Tesla to become the global No.1.
Wang Chuanfu is the founder of BYD, founded in 1995, headquartered in Shenzhen.

输出:

【原子事实分解】
1. 比亚迪2024年销量400万辆
2. 比亚迪已经超过特斯拉成为全球第一
3. 王传福是比亚迪创始人
4. 比亚迪公司成立于1995年
5. 比亚迪总部在深圳

事实 1:比亚迪2024年销量400万辆
判定:✅ 正确
置信度:92%
依据:比亚迪2024年年报披露销量约395万辆,400万为近似值

事实 2:比亚迪已经超过特斯拉成为全球第一
判定:⚠️ 无法验证
置信度:—%
依据:需明确比较维度(销量/市值/净利润),不同维度结果不同

事实 3:王传福是比亚迪创始人
判定:✅ 正确
置信度:98%
依据:比亚迪官网及公开资料一致确认

事实 4:比亚迪公司成立于1995年
判定:✅ 正确
置信度:95%
依据:比亚迪官方公司简介

事实 5:比亚迪总部在深圳
判定:✅ 正确
置信度:99%
依据:公开资料一致确认

🔍 验证结果:部分正确(82分)

✅ 正确的事实:
  • 比亚迪2024年销量400万辆(置信度:92%)
  • 王传福是比亚迪创始人(置信度:98%)
  • 比亚迪公司成立于1995年(置信度:95%)
  • 比亚迪总部在深圳(置信度:99%)

⚠️ 无法验证:
  • 比亚迪已经超过特斯拉成为全球第一(原因:比较维度不明确)

📊 一致性检测:通过
📚 参考来源:比亚迪官网、2024年年报、新浪财经、雪球

文件结构

checktruth/
├── SKILL.md          # 本文件(核心)
├── prompts/          # 提示词模板(可选,供参考)
├── tests/            # 测试用例(可选)
├── docs/             # 设计文档(可选)
└── reference/        # 参考代码(可选,需外部API Key,非核心功能)

核心功能零配置:本 Skill 的所有核心验证逻辑均通过 SKILL.md 中的指令、 WorkBuddy 内置 LLM 能力和 WebSearch 完成,无需任何外部 API Key

reference/ 文件夹包含基于外部 LLM API 的参考实现(需自行配置 Key), 不是核心功能,仅供开发者参考。普通用户无需理会此文件夹。


版本:0.3.1 | 统一触发词:英文 /checktruth,中文 /别瞎说 | 自动判断模式A/B

Usage Guidance
Safe to consider for ordinary fact-checking. Review cited sources, avoid pasting confidential content, and ignore the reference Python scripts unless you intentionally want to configure external LLM API keys.
Capability Analysis
Type: OpenClaw Skill Name: checktruth Version: 1.0.1 The 'checktruth' skill is a fact-checking tool that decomposes text into atomic facts and verifies them using web searches and multi-source LLM comparisons. The bundle includes well-documented Python reference scripts in the 'reference/' directory and clear instructions in 'SKILL.md'. No evidence of data exfiltration, malicious execution, or harmful prompt injection was found; the use of environment variables for API keys in the reference scripts is standard practice and clearly disclosed in the security documentation.
Capability Tags
requires-sensitive-credentials
Capability Assessment
Purpose & Capability
The stated fact-checking purpose matches the core workflow: decompose claims, search for evidence, and score confidence. Users should notice that the package also includes optional reference Python code for external LLM providers.
Instruction Scope
The skill is user-invocable and instructs the agent to use WebSearch/WebFetch for evidence; this is purpose-aligned but means external search results can affect the answer.
Install Mechanism
There is no install spec and no evidence of automatic code execution. The Python files are presented as optional reference code, not core runtime.
Credentials
Registry requirements declare no required credentials, which fits the core skill. Optional reference scripts read OpenAI/Anthropic/Gemini-style API keys from environment variables if explicitly run.
Persistence & Privilege
No persistence, background workers, privileged file writes, account mutation, or destructive actions are shown. Optional scripts only read user-specified input files.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install checktruth
  3. After installation, invoke the skill by name or use /checktruth
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.1
**Summary:** Major update — Skill renamed, new unified triggers, file structure improved, and clear separation of core logic from optional code. - Skill renamed from "biexishuo" to "checktruth"; triggers are now `/checktruth` (English) and `/别瞎说` (Chinese). - Automatic detection of input structure to select Q&A validation (模式A) or article/statement verification (模式B). - All core functionality is driven by SKILL.md prompts, built-in LLM, and WebSearch; no external API keys required. - Optional Python reference code moved to `reference/` directory (not required for normal users, requires manual API key if used). - Security, data handling, and dependency statements now distinguish between core functions and optional scripts. - Documentation, examples, and usage instructions updated for clarity and multi-language support.
v1.0.0
- 首发版本,正式发布“别瞎说(check truth)”事实核查技能 - 支持两种验证模式:问答验证(验证AI回答)和文章/论点验证(核查段落、帖子、论点等内容) - 自动分解内容为原子事实,逐条对照权威来源验证,并标注置信度 - 内置一致性检测,发现内容内在矛盾 - 输出综合打分与详细依据,所有结果均用中文显示 - 无须外部API密钥,依赖内建模型与实时Web搜索
Metadata
Slug checktruth
Version 1.0.1
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 2
Frequently Asked Questions

What is 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker?

零配置 AI 事实核查工具。支持问答验证和文章/论点核查,自动分解事实、检索证据、输出可信度评分。基于 FActScore、xVerify 等学术论文设计。 Zero-config AI fact-checking tool. Supports Q&A verification and article/argum... It is an AI Agent Skill for Claude Code / OpenClaw, with 63 downloads so far.

How do I install 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker?

Run "/install checktruth" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker free?

Yes, 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker support?

别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created 别瞎说 - AI事实核查器/Checktruth - AI Fact-Checker?

It is built and maintained by NLMakes (@xuehengzhang10-hub); the current version is v1.0.1.

💬 Comments