Safe agent specification
Defines tools, permissions and boundaries before building an agent.
Complete prompt
Write the specification for an agent whose goal is [GOAL]. Include scope, exclusions, allowed tools, input schemas, permissions, memory, prohibited data, confirmations and human escalation. Define success, evaluation, cost and time limits, and behavior for ambiguous results. Add foreseeable threats and defenses. The agent must never expand its permissions or perform unauthorized external actions.
Goal: answer internal incidents without changing settings.