소개
The Plan-Then-Execute skill is a security-focused framework designed for high-stakes AI workflows and tasks involving untrusted data. Unlike standard adaptive agents, this skill forces Claude to generate a complete, 'frozen' sequence of actions in a YAML format which must be reviewed and approved by a human before execution begins. Once approved, the execution phase follows the plan exactly, preventing tool outputs from maliciously redirecting the agent's behavior. This pattern provides a robust audit trail and ensures that the human remains the final authority on what tools are called and in what order.
