Community编程与开发github.com

事实核查

在交付任何事实性声明之前自动核实其准确性,确保用户基于可靠信息做出决策。适用于规则、政策、法律、财务、日期或产品事实等场景。

事实核查 是什么?

该技能通过系统化的五步流程确保AI输出内容的准确性。它首先从草稿回答中提取所有事实性声明,然后根据其对用户决策的影响程度进行分类(结构性声明必须验证,背景声明仅需标注)。对于关键声明,它要求对照原始来源而非摘要进行核对,特别关注日期、金额、预测和负面声明等容易出错的元素。它支持通过可靠抓取工具或真实浏览器获取页面内容,并能处理来源不可达、模板过时和算术差异等情况。最终,它会在答案中用清晰标签(✅已验证、⚠️未验证、❓推测)标记每个声明,并在交付前进行最终检查。该技能适用于Claude Code、Cursor、Codex等AI辅助工具,特别适合需要高准确性输出的行业如法律、金融和产品开发。

兼容平台✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/josie2479-maker/claude-skills/tree/HEAD/skills/fact-verification

Installed? Explore more 编程与开发 skills: steipete/bluebubbles, steipete/eightctl, steipete/blucli · View all 6 →

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

事实核查 是做什么的?

许可协议: 根据 CC BY 4.0 发布——可自由使用、改编和分享,需注明出处。

1. 提取声明

列出草稿回答中的每一个事实性声明。跳过观点和推理。

2. 分类

为每个声明标记:

  • 结构性声明:用户将据此做出决定、采取行动或转发 → 必须验证
  • 背景声明:仅提供背景信息 → 标记,无需深究

3. 验证承重声明

  • 对照原始来源(发布商自己的网站、法规或标准本身、官方文档)进行核查,而不是对照摘要。

  • 如果可用,使用可靠抓取工具获取页面文本;否则使用真实浏览器工具拉取页面,避免使用摘要性抓取工具。若无法访问来源,请如实说明;不要凭记忆补充信息。

  • 日期、截止日期、金额:根据来源自身的措辞确认,并注明其截止日期。

  • 预测和警告:已过时的预测不是结果。为每个数字标注“已观测”或“已预测”。对于过往预测、峰值或警报级别,请查找观测记录(测量日志或系统记录数据),而非预测或警报文本,并报告实际数据。来自撤销预测的数字从未发生过。

  • 在来源中找到确切的数字或措辞。一个有效的链接、真实的期刊以及匹配的主题并不能证明来源确实说了那个内容。多个来源一致只有在它们没有相互复制的情况下才有效。

  • 模板或旧文档中引用的条款可能已过时。将每个被引用的规则、条款或章节与当前官方文本进行核对,而不是模板。

  • 负面声明(“无要求”、“未找到”)仅涵盖实际搜索过的来源和词语。在陈述此类声明之前:

    • 搜索可能产生该义务的每一层(法律、法规、官方指南、地方规则)以及项目本身的记录;
    • 使用同义词搜索概念,而不是仅使用一个关键词;
    • 将声明范围限定为所搜索的内容(“在所检查的来源中未发现相关法规”,而不是“无要求”)。

    将该范围保留在标题本身。标题下方的注意事项在被复制时不会随标题一同传播。

  • 区分规则允许的内容与用户自身案例中实际发生的情况。请求、信件、批准和决定来自用户的记录(他们的笔记、决策和事件日志、项目中的早期研究),而非来自规则本身。首先搜索这些记录,如果明显的位置为空,则询问记录存储在哪里。

  • 产品计划(“X 是否已计划?”):首先查看供应商帮助中心的“当前限制”部分,然后查看发布说明,最后查看博客和新闻。

  • 提供给您的算术差异:从来源重新加总每一级(行项目、小计、总计)。给定的解释可能只涵盖了几个差异中的一个。

4. 在答案中标注

  • ✅ 已验证:[来源,检查日期]
  • ⚠️ 未验证:来自训练数据,可能已过时
  • ❓ 推测/推断:说明其依据

5. 交付前

重新阅读步骤 2-4。任何承重声明如果没有 ✅ 或明确的 ⚠️ → 修复它或标记它。

Individual skills in this repo

This repo contains 3 individual skills — each has its own dedicated page.

josie2479-maker/claude-skills

Test a ready-made list of deliverables, documents or artifacts before anyone commits to tracking it. Use when someone is handed a list from a template, vendor, consultant, PMO framework or AI (\"standard project documents\", \"required artifacts\", \"deliverables checklist\") and asks whether to adopt it, what is really required, what to drop, or what is missing. Tags each item by who requires it (funder, law or regulation with citation, a named person, or standard practice that nobody requires), reads the funder's or agency's own template section by section to find required items the list leaves out, and returns a table plus a drop / keep / add list. Not for building a list from scratch with no list in hand, and not for checking whether finished documents are good.

josie2479-maker/claude-skills

Interview the user about something they are ABOUT TO BUILD, before any of it is built — a feature, a system, an automation, a workflow, a document structure. Use when a plan exists only in their head or as a rough idea and the design decisions are still open: \"grill me\", \"interview me about this\", \"help me think this through before I build it\", \"what am I not deciding\", \"ask me what you need to know\". Works the design as a dependency tree, asking one answerable question per round (the one that unblocks the most decisions first) with a recommended answer attached, and never asking anything it could look up itself. BOUNDARY — this is front-loaded, for decisions not yet made. Reviewing a plan that already exists and stress-testing it for holes is the plc skill. Interrogating information someone else handed you is critical-thinking-walkthrough. If the thing is already written down and the question is \"is this any good\", it is not this skill.

josie2479-maker/claude-skills

Strip AI-essay voice from a document that will be shared outside this conversation — a memo, spec, brief, proposal, SOW, report, or any deliverable going to a supervisor, agency, client, or contractor. Use when asked to \"clean this up\", \"make this professional\", \"take the Claude out of it\", \"de-AI this\", \"make it ready to send\", or when a draft is about to be handed off and has not had a register pass. Also run it unprompted as the last step when drafting any external-facing document. Not for chat replies, personal notes, or internal scratch files — conversational voice is fine there and flattening it is a loss. Not for fiction or other creative writing, where the voice is the point; see the fiction note in the body.

相关技能