Run an audit.

Paste the instruction that feels too vague, too brittle, or too easy for an agent to drift from. Rubrkit will grade it, mark weak dimensions, and turn it into a testable artifact.