RiskBlind Radar 风险盲区雷达

AI 对话中的风险雷达——主动追问你没想清楚但做决策前必须想清楚的风险问题。

Security Scan

The skill's requested resources and runtime instructions are coherent with its stated purpose (proactively asking risk questions during conversations); no unexpected credentials, installs, or external endpoints are requested.

Purpose & Capability

Name/description (主动追问风险问题) match the SKILL.md, templates, and rules. The skill is instruction-only, asks to scan messages for trigger words and prompt up to two risk questions — all consistent with a 'risk radar' assistant.

Instruction Scope

Instructions confine behavior to scanning user messages for trigger words and asking short follow-ups. One notable design detail: a 'Risk Watch' background mode and time‑decay scoring are described, which implies persistent tracking of commitments and timestamps. The skill references an optional memory file (memory/riskblind/OBSERVATIONS.md) but does not strictly define when or whether that file will be written. If enabled, that would mean the agent stores conversational state across sessions; otherwise the background tracking cannot function fully. Otherwise there are no instructions to read unrelated system files or send data to external endpoints.

Install Mechanism

Instruction-only skill with no install spec and no code to download or run — lowest-risk installation. The README suggests a local placement under the OpenClaw workspace and editing openclaw.json, which is normal for skills.

Credentials

The skill requires no environment variables, no credentials, and no external binaries. That is proportional to its function.

Persistence & Privilege

always:false and normal autonomous invocation are used (no special privileges). However, the documented 'Risk Watch' and scoring (time-decay, commitments) imply the need for persistent logs if enabled; the skill references a memory path but marks logging as optional. Confirm whether you want persistent observation logs and where they will be stored and with what retention/permissions.

Assessment

This skill appears to do what it says: scan conversation text for trigger words and ask brief risk questions. Before installing, decide whether you want the optional persistent tracking: if you enable the memory/log file (memory/riskblind/OBSERVATIONS.md) the skill will keep state about commitments and timings — review where that file will live, its filesystem permissions, and retention policy. If you do not want any state saved, keep the memory feature disabled. Also test the skill in a private agent first to confirm trigger words and frequency are acceptable (it scans messages every turn and can ask up to two questions), and review any modifications to openclaw.json before applying them. No credentials, external downloads, or hidden endpoints were found.

核心目标

每次对话结束前，主动问一个用户没回答过、但必须回答的风险问题。

不是"静静观察"，而是帮你探测那些你自己都看不到的风险盲区。

---

一句话说明

RiskBlind Radar 是 AI 对话中的"风险雷达"——每次对话时，主动追问那些你没想清楚、但做决策前必须想清楚的风险问题。

---

两大工作模式

模式A：主动追问（Proactive Risk Questioner）— 主模式

当用户提到特定主题时，自动触发对应问题模板。

工作流程：

1. 用户消息 → 扫描触发词
2. 命中模板 → 激活风险问题
3. 在对话结束前提出（不延时）

核心原则：

• 问"你没问自己但必须想清楚"的风险问题
• 不问可以马上回答的问题
• 不问太多，一次最多2个
• 不在闲聊时追问，等有实质内容时

---

模式B：风险追踪（Risk Watch）— 后台层

在主动追问之外，持续追踪风险评分。

触发条件：

• 用户做出承诺（模板T7触发）
• 承诺到期未处理
• 时间衰减评分达到阈值

---

问题模板库（10个核心模板）

详见 QUESTION_TEMPLATES.md

---

风险评分

评分 = 重要性系数 × 时间衰减系数

重要性系数：
- 资金断裂：3
- 承诺未兑现：3
- 方向错误：3
- 执行拖延：2
- 技术细节：1

时间衰减系数：
- 1天内：×1.0
- 3天内：×1.5
- 7天内：×2.0

触发阈值：评分 ≥ 5 → 主动提醒

---

核心原则

1. 问，不记录 — 主动追问是主业，记录是辅业
2. 宁缺毋滥 — 没有好问题就不问，不为问而问
3. 不叠加压力 — 一次最多2个问题，不追问正在回答的问题
4. 只在实质内容时触发 — 闲聊时不触发模板

---

记忆文件

• memory/riskblind/OBSERVATIONS.md — 风险追踪日志（可选）
• skills/riskblind-radar/QUESTION_TEMPLATES.md — 问题模板库
• skills/riskblind-radar/RULES.md — 认知偏差参考 & 机会识别

---

文件结构

riskblind-radar/
├── SKILL.md                  # 技能定义
├── QUESTION_TEMPLATES.md     # 十大追问模板
├── RULES.md                  # 认知偏差参考 & 机会识别
├── memory/riskblind/         # 观察日志（可选，不含在分发包里）
└── README.md                 # 使用说明

---

最后更新：2026-04-10