360 Secure SkillHub
360 安全技能中心

human-gates

用于设计和审查自动化工作流中哪些步骤需要人类批准、判断或承担责任,按权威、问责、实质歧义和风险控制分类,并输出门禁触发条件、负责人、所需证据、可选结果与超时路径。

raguets
raguets
热度 18.13
最近更新 2026-09-04 00:50:58
分类:自动化依赖:纯提示词应用场景:工作流 / 任务管理

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add human-gates
Metadata
name
human-gates
description
Part of the Overpowered skill suite. Design or review where human approval, judgment, or accountability must remain in an automated workflow. Use when automation crosses authority, legal, financial, safety, irreversible, or genuinely ambiguous decisions; do not preserve manual steps merely because they exist today.
metadata
{"suite":"overpowered","suite_url":"https://github.com/raguets/overpowered","level":"0-primitive","version":"0.1.0"}

Overpowered / Human Gates

Core rule

Keep humans for authority, accountability, material ambiguity, and unacceptable risk — not for work a deterministic or agentic system can reliably perform.

Procedure

For each candidate human step:

  1. Identify why the human is currently involved.
  2. Classify the reason:
    • authority — only an authorized person may decide;
    • accountability — an accountable owner must accept the outcome;
    • material ambiguity — available evidence cannot reliably determine the case;
    • risk control — error cost exceeds the accepted automation boundary;
    • legacy/manual habit — no durable reason remains.
  3. Automate legacy/manual steps when controls are sufficient.
  4. Define each retained gate with:
    • trigger condition;
    • decision owner or role;
    • evidence presented to the human;
    • allowed decisions;
    • timeout/delegation path;
    • downstream effect.

Output contract

Gate | Trigger | Why human | Required evidence | Owner | Outcomes

Gotchas

  • “Human in the loop” is not a design; define the exact gate.
  • Do not make a human re-check deterministic calculations unless policy requires it.
  • Historical manual ownership is not proof that human judgment is needed.
  • Never infer delegated authority from precedent alone.

Stop condition

Stop when every retained human step has a concrete reason and every removed human step has an adequate automated control.