360 Secure SkillHub
360 安全技能中心

Rejection Logger

该 Skill 指导智能体在评估多种方案后做出取舍、或判定用户请求不安全/超出范围时,按固定模板把被拒绝的目标、原因与替代方案追加到 .learnings/REJECTIONS.md,形成可供审计与自我校正的决策日志;随附一个接收目标、原因、替代方案的 shell 记录脚本。

balkanblbn
balkanblbn
热度 56.78
最近更新 2026-05-17 12:56:41
分类:知识依赖:读写文件 / 代码执行应用场景:笔记整理 / 工作流 / 模板

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 rejection-logger
Metadata
name
rejection-logger
description
Captures and logs choices, options, or prompts that the agent evaluated and decided NOT to execute. Use whenever you skip a task, reject an approach, or choose one method over another to provide transparency into your reasoning.

Rejection Logger

Transparency isn't just about showing what you did; it's about explaining what you didn't do. This skill helps you document rejected paths.

Core Workflow

1. Identify Rejection

When you evaluate multiple ways to solve a problem and pick one, or when you decide a user request is unsafe/out of scope, log it.

2. Log Entry

Append to .learnings/REJECTIONS.md (create if missing):

## [REJ-YYYYMMDD-XXX] <short_title>

**Timestamp**: ISO-8601
**Target**: <What was requested or considered>
**Decision**: REJECTED
**Reason**: <Why it was rejected (e.g., safety, complexity, better alternative)>
**Alternative**: <What was done instead>

When to Use

  • When a user asks for something and you say "No" or "I can't".
  • When you consider two tools and pick one.
  • When you refactor code and decide against a specific library.

Benefits

  • Audit Trail: Humans can see your internal deliberation.
  • Trust: Showing rejections proves you are thinking, not just guessing.
  • Self-Correction: Reviewing rejections helps you refine your decision boundaries.