Make a clear, safer operating decision.
You will be able to write agent instructions that a colleague can review, test, and update without guessing what the agent was meant to do.
Good agent work is useful before it is impressive.
Good instructions are operating documentation. They state the task, trusted source, required checks, allowed language, output shape, and stop conditions. A long prompt cannot compensate for missing policy or unclear authority.
Make the relationship visible.
The language that keeps the work clear.
Work through the decision in order.
- 01
State role and outcome
Open with the job, user, and useful result; omit marketing language.
- 02
Name trusted sources
List the approved knowledge and data sources, then say what to do if they disagree or are unavailable.
- 03
Write decision rules
Use short, testable rules for questions, recommendations, policy limits, and escalation.
- 04
Require a usable output
Specify the fields a human needs to review or continue the work, including why the agent chose its action.
A realistic, bounded implementation.
A support agent assists with a subscription-cancellation request. Its instructions say it may explain the published cancellation process and retrieve the plan only after identity verification.
It must use the current policy source, never promise a refund, and produce a record with customer intent, verified status, policy reference, action requested, and human owner if an exception is needed.
A reviewer can now test the policy boundary and the output separately. The agent is not rewarded for sounding confident; it is rewarded for being accurate and appropriately limited.
Use this copyable working template.
Adapt it to the client’s evidence, policy, people, and tools. Do not treat placeholders as approved instructions.
Put the operating system around the agent.
Keep approved messaging, knowledge, contact context, and workflow conditions in the systems that operators already use; link every agent instruction to the current owner of the underlying policy.
Before you move on
- Rewrite one current prompt as auditable instructions.
- Ask a second person to find the source hierarchy and stop rule without explanation.
- Create two test cases that would expose an unclear rule.
- Instructions separate facts, rules, outputs, and limits.
- The agent knows what source is authoritative.
- A human can review why an action was proposed.
- Updates have an owner.
Put the learning to work
Build an AI service people can trust.
Create a free Spacebrain account and use the operating layer around your AI service.