Guardrails and Safety

Safety isn't a wrapper, it's an architecture decision.

A "be safe" line in the system prompt is not a guardrail. A single output filter is not a guardrail. Safety in production is layered, observable, and tested against an explicit adversarial set — same rigour as functional evaluation.

When to use this skill

• The user is building an agent that takes free-form user input.
• The user is exposing tools with real side-effects (writes, payments, messages).
• The user operates in a regulated domain (finance, health, legal, education).
• The user reports a jailbreak, prompt-injection, or unintended action incident.

The five layers

Treat these as a stack. Skipping a layer is a deliberate choice with a cost.