AI glossary
Guardrails
Guardrails are the checks around a model call that stop unsafe or unwanted input and output, enforced in code rather than only asked for in the prompt.
They sit on both sides of the model: on the way in, checking length, topic and personal data; on the way out, checking format, banned content and whether a tool call is allowed.
A simple example is an allowlist: the assistant may call these three tools with these argument ranges, and anything else is refused by the code before it runs.
A guardrail written only as a sentence in the system prompt is a hope. Real guardrails are code that runs whatever the model says.
Learn it hands-on
These free lessons teach it with commands to run in a lab and checked tasks. Read them whole with no account.