A strong prompt cannot compensate for an unclear task or an uncontrolled workflow. When I think about an AI implementation, I start with the operational decision: what work should the system support, what information may it use, and what must remain under human control?
Design the surrounding system
- Inputs: known sources, ownership and data sensitivity.
- Instructions: purpose, constraints and an explicit stop condition.
- Tools: the minimum permissions required for the task.
- Evaluation: examples and criteria that reveal whether outputs are useful.
- Review: human approval where errors or side effects matter.
- Operations: monitoring, incident handling and a route to disable or revise the workflow.
NIST frames AI risk management as a continuous set of govern, map, measure and manage activities. That is a useful reminder: the model call is one component of an operational system.