judge-criteria-designer

Describe the action boundary, domain, existing policies, and worst-case failure to the agent. It produces testable yes/no criteria across authorization, evidence, exposure and risk, and policy categories, plus decision rules for allow, block, revise, or escalate.

An action proposal is a structured object the actor agent must produce before executing, with fields for intended action, reason, supporting evidence, authorization basis, expected consequence, data exposed, reversibility, and risk flags. The judge inspects this proposal before execution.

Human-in-the-loop approval is needed when the action is irreversible, exposes sensitive data, has ambiguous authorization, or crosses legal, security, or compliance boundaries. The Skill flags these conditions and designs criteria that surface ambiguity rather than resolving it silently.

Common failures include acting on inferred rather than explicit authorization, citing stale or insufficient evidence, and exposing data to unintended recipients. The Skill ends each design with the three most likely mundane failures and maps each to the criteria that catch it.

The Skill does not invent policies the user has not described. When policy gaps exist, it flags them explicitly and suggests what the user needs to define, keeping criteria grounded in stated rules and norms.