TEST THE SYSTEM PROMPT

Don’t evaluate an Instruction by reading it and thinking it “seems reasonable”.

You have to run it.

Normal Case

A request exactly as the Agent was designed for.

Check:

• whether it goes through the full workflow;
• whether the output is correct.

Missing Data Case

Deliberately leave data out.

Check:

• whether the Agent makes things up;
• whether it triggers the Failure Rule.

Conflict Case

Provide contradictory data.

Check whether the Agent:

• detects it;
• distinguishes between sources;
• avoids concluding too early.

Scope Test

Ask the Agent to do something outside its scope.

Check whether the Agent keeps its boundary.

Long Workflow Test

Use a task that is long enough.

Check whether the Agent:

• skips steps;
• shortens the workflow;
• forgets the final requirement.

This is a very common error with long workflows.