← ClaudeAtlas

human-approval-gateslisted

Design human-in-the-loop approval for high-risk agent tools. Use when tools delete data, spend money, send messages externally, change prod, or user says "human approval", "confirm before run", or Chinese "人审", "二次确认", "危险操作确认".
Wanbinyu/ai-surface-skills · ★ 1 · AI & Automation · score 70
Install: claude install-skill Wanbinyu/ai-surface-skills
# Human Approval Gates > If undo is hard or money moves, a human must say yes. ## Overview Require human confirmation for high-risk tool calls. ## Steps 1. List candidate tools (from permission matrix or inventory). 2. For each: risk reason, approval prompt text (what/why/blast radius), timeout behavior (deny by default). 3. Define who can approve (user vs role). 4. Logging: store decision for audit. 5. Never allow silent re-approval loops without new context. ## Exit criteria - [ ] HITL tool list - [ ] Prompt template per tool or class - [ ] Default on timeout: deny - [ ] Audit log fields - [ ] Bypass policy (if any) documented and narrow ## Anti-patterns - Auto-approve after N seconds - Vague "are you sure?" without naming the action - Burying destructive tools without gates ## Output template ```markdown ## Human approval design | tool | risk | approval prompt summary | timeout | auditor fields | |------|------|-------------------------|---------|----------------| ```