Guardrail

What is a Guardrail

A guardrail is a hard rule that blocks an entire category of action outright: never force-push to main, never delete without a lock, never commit a file matching a secrets pattern. It is not a judgment call the model makes case by case, it is a boundary that holds regardless of how reasonable the model's argument for crossing it might sound in the moment.

Guardrails are usually implemented as hooks, since a hook can actually stop an action with a real exit code, where a prompt instruction can only ask nicely. This distinction matters most for anything the model could talk itself into under specific circumstances, an instruction says please don't, a guardrail makes it structurally not happen.

Guardrails matter most for AFK or otherwise unsupervised work, where there is no human immediately present to catch a bad call before it executes. A task with solid guardrails around the genuinely dangerous actions can be run unattended with real confidence, a task relying only on the model's own judgment to avoid disaster generally cannot.