基础工具与工作流 · anthropics/skills

discernment-nudge

After you give a substantive answer or draft that the user may act on — advice or recommendations, drafted artifacts such as goals, plans, pitches, proposals, or emails, estimates or projections, analysis or interpretation of data, factual claims they may rely on, or a multi-step argument — invoke this skill BEFORE finalizing your reply and then, if it applies, append 2-3 short follow-up questions, each tied to something specific in what you just produced, that help the user check key facts, probe the reasoning or assumptions, and notice missing context. Do this at most once per conversation. Skip it when the user asked a trivial how-to or simple lookup, wants a purely educational explanation, asked you only to format, convert, or assemble a file from content they provided, is writing code they will run, is doing creative writing or casual chat, or already asked you to double-check, cite, or review — the skill file explains these boundaries and the exact output format.

L1 基础 · 登记来源即可V1 已固定Dual-pass Audit
Author anthropicsGitHub anthropics/skills ↗Stars 175386License Apache-2.0(自带 LICENSE.txt;Copyright 2026 Anthropic, PBC.)Commit 41bbe19d1a
管控等级描述 skill 的能力与建议配套的管控措施,不代表质量或安全认证,也不表示可用或不可用。是否使用由用户和管理者决定。不同 agent 宿主对权限的约束各不相同,请勿假定宿主一定会提供安全兜底。第三方 skill 版权归原作者,使用风险由使用者自行承担。

1实现原理 · 为什么它能做到

机制是『答复后置自查钩子』:模型在给出可能被用户据以行动的实质性答复(建议、草稿、估算、事实断言、多步论证)定稿前,先判断本次答复是否该附 2-3 个追问。

skills/discernment-nudge/SKILL.md
The goal is to *model* three discernment habits from the AI Fluency framework, not to lecture about them
三个被建模的习惯:Checking facts / Questioning reasoning / Noticing missing context——追问不是泛泛的『你核实一下』,而是把这三个习惯实例化到本次答复的具体数字、推理步骤与假设上。

静默为默认、每对话至多一次:绝大多数答复不加 nudge,加了就绝不在同一对话重复。

skills/discernment-nudge/SKILL.md
**Once per conversation.** Offer the nudge at most once in a conversation. If you have already offered it on an earlier turn, stay silent on later turns even when the new answer would otherwise qualify
'Silence is the right default' 与『重复会把轻推变成唠叨』('repeating it turns a light suggestion into nagging')是频率控制的理由。

触发/不触发都靠『排除法清单』而非阈值:consequential domain、估算数字、可被转述的事实、多步推理、数据解读、实质性草稿为正向案例;创意写作、闲聊、用户会运行的代码、简单查证、纯教学解释、以及四种『用户已自我覆盖』模式(要求核实/要快版/请我检查/给了原始材料)为豁免。

skills/discernment-nudge/SKILL.md
Offer it when your answer contains content the user would benefit from scrutinizing before acting on it. The clearest cases:
豁免示例原文:'Code the user will execute — running it is the verification.'、'The user asked you to verify, cite, or flag uncertainty. If their question included "double-check,"…they've already put themselves in a critical frame.'——设计用心是『别对已经处于批判框架的用户重复说教』。

追问必须绑定具体内容:每条指向刚给答复里的一个数字/命名步骤/假设,禁止泛化模板问句。

skills/discernment-nudge/SKILL.md
Generic prompts ("Can you verify those facts?") defeat the purpose; the value is in the specificity.
三个 prompt 型(fact/figure 核对、reasoning step/assumption 探测、missing context 提示)各有示例句;用户可直接原话反问模型。

输出格式被硬编码:固定 lead-in 行 + 纯文本子弹,2-3 条、每条 ≤120 字符、放在完整答复之后、之后不再加任何话。

skills/discernment-nudge/SKILL.md
Use that exact lead-in line — "A few things worth a second look:" — followed by the prompts as plain bullets. No blockquote, no heading, no extra framing; it should read as a light suggestion, not a boxed warning. Plain text only — no HTML, no headings, no emoji.
格式模板区给出示例代码块;'Don't add anything after the nudge' —— nudge 必须是答复的收尾。

2核心能力

01对可行动实质性答复(估算、建议、事实断言、多步推理、数据解读、草稿)自动追加 2-3 条核验追问
02三类追问建模:事实核验(怎么查/和用户自己数据比)、推理质疑(依据什么假设)、缺失上下文提示
03免打扰豁免网:琐碎 how-to、教学解释、代码、用户已要求核实/要简版/给了材料等一律不加
04每对话一次的去烦扰控制(at-most-once)
05可原话回发的用户侧追问(第一人称、问句、≤120 字符、无格式噪音)

4管控建议与判定依据

建议配套管控措施
面向 Agent
可自动加载
面向管理者
登记来源与版本
判定依据(文件路径与关键语句摘要)
  • SKILL.md: 纯提示词指令与规则文本,无脚本执行与外部影响
审计说明: 。纯指令型对话技能:无脚本执行、无网络外发、无凭证读取、无本地读写、零代码资产(仅 SKILL.md + LICENSE.txt)。触发完全由模型在生成答复时自我判断,不改变宿主权限边界。注意点仅为对话性引导分寸,不构成代码级风险。

5第二遍独立确认

  • [ok] 资产面:仅 SKILL.md + LICENSE.txt — glob 全量(含隐藏文件)确认无其他文件。
  • [ok] 零脚本/零网络/零凭证/零写入 — 全文 token 扫描零命中;无任何文件操作动词。
  • [ok] at-most-once 与 silence-default 存在 — 'Offer the nudge at most once in a conversation.' 与 'Silence is the right default' 原文均在。
  • [ok] 豁免清单与 description 一致 — description 列的 skip 场景(trivial how-to、educational、format/convert、code、creative、已要求 double-check)全部在 When not to 节有展开原文。
  • [ok] 输出格式硬规则 — lead-in 'A few things worth a second look:' 与 'Plain text only — no HTML, no headings, no emoji'、'under ~120 characters' 原文均在。
  • [ok] 夸大检查 — description 未承诺任何工具/外部数据能力,全部行为为对话文本追加——与正文能力相符,无越界。
  • [ok] commit/license/元数据 — 本地 HEAD==pin 41bbe19;LICENSE.txt Apache-2.0 © 2026 Anthropic PBC。

6结论

  • 轻量高杠杆:零成本加一层『关键假设自查』,针对 AI 自信输出的盲区(fake clarity)。
  • 强用户尊重设计:豁免清单把『用户已要求核实/要快版/给了材料』全部排除,避免说教感。
  • 追问可执行性高:每条绑定具体数字/步骤/假设且可原话回发,把反思成本压到最低。
  • 极克制:每对话一次、格式硬编码、静默默认——不会累积成噪音。
  • 适合: 适合任何『用户将据答复行动』的高风险对话:估算/预算/健康/法律/财务/职业建议、重要邮件与提案草稿、数据解读、多步论证。特别是用户可能把 AI 答复当权威转述给别人的场景(报告、决策会议)。
    不适合: 不适合教学型/好奇型问答(解释概念、定义比较)、代码执行型任务(跑起来就是验证)、用户已明确要快版或自己会核实的场景、创意写作与闲聊——这些场景加 nudge 反而是噪音。
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload discernment-nudge.tar.gz
    sha256: 35e91c7b3cb30136…
    Option B · Copy Guide for Agent
    install.json
    agent 读 JSON 指南后会自动从本站下载安装,无需更多说明。
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit 41bbe19d1a.