step 3 policies

Define hard rules that constrain your agent's behavior. Policies act as non-negotiable boundaries the agent cannot cross.

---

Default Policies

ClawEngine ships four default policies:

Policy	Type	Default State
No Harmful Content	Safety	✅ Enabled
Privacy Protection	Safety	✅ Enabled
Factual Accuracy	Content	✅ Enabled
Stay On Topic	Behavior	⬜ Disabled

---

Policy Types

• Safety — Prevent harmful, dangerous, illegal, or unethical outputs

• Content — Rules about accuracy, sourcing, citations, and quality standards

• Behavior — Operational rules like staying on topic, respecting rate limits, or escalation procedures

• Custom — Your own rules (add via API or define in the persona step as behavioral guidelines)

---

Toggling Policies

Each policy has a switch. Enabled policies are compiled into the "Policies & Guardrails" section of the system prompt. Disabled policies are excluded entirely.

---

How It Works at Runtime

Enabled policies are presented to the LLM as hard constraints in the system prompt. They take priority over the agent's goal and persona — if a user request conflicts with a policy, the agent refuses or redirects.

---

Best Practices

• Always keep safety policies enabled unless you have a specific reason to disable them

• Add content policies for agents that handle financial data, medical information, or legal advice

• Use behavior policies to keep agents focused in customer-facing deployments