🛡️ The Gatekeeper — Context-Aware Safety Persona

Keep your AI agent useful without being dangerous.

The Problem

AI agents drift outside their role, overshare private info, or take risky actions without confirmation. This is the #1 complaint in production agent deployments.

Your agent deletes files without asking. It echoes passwords in chat logs. It force-pushes to main. It emails the wrong person. And by the time you notice, the damage is done.

The Solution

The Gatekeeper is a complete safety persona that classifies every action into three tiers:

• 🟢 SAFE — Auto-proceed (reading files, searching, drafting)
• 🟡 CAUTION — Brief confirm (editing files, git commits, API calls)
• 🔴 DANGEROUS — Full stop + explain (deleting data, sending externally, purchases)

What You Get

SOUL.md — The Safety Persona

A battle-tested persona with:

• Tiered action classification (safe/caution/dangerous)
• Explicit "never do" lists (8 hard rules that can't be overridden)
• PII auto-detection and redaction
• Tone calibration (efficient for safe actions, careful for dangerous ones)
• Error recovery protocol

AGENTS.md — Confirmation Gate Rules

• Gate protocol with formatting templates
• Batch confirmation for related actions
• Session-level bypass rules (caution only, never dangerous)
• Destructive action protocol (prefer trash, create backups)
• PII handling workflow
• Sub-agent safety inheritance

safety-rules.md — Configurable Risk Matrix

A complete, editable matrix covering:

• File operations (7 action types)
• Shell commands (10 action types)
• Communication (6 action types)
• Data & privacy (7 action types)
• Database operations (7 action types)
• Financial & purchasing (5 action types)

Each with default tier, notes, and a customization guide.

examples/ — 10 Real-World Edge Cases

1. User accidentally shares a password
2. Bulk delete request
3. User says "skip confirmations"
4. Email to unknown person
5. Force push to main branch
6. SQL injection in user input

Unzip and copy SOUL.md, AGENTS.md, safety-rules.md, and examples/ into your OpenClaw workspace. Customize safety-rules.md to match your risk tolerance.