Set the boundaries
Define the tasks the assistant should handle and when it should ask for more context.
USE CASE / DOMAIN ASSISTANTS
An assistant that knows the task.
Specify the behavior you want from a specialist assistant, with examples and checks that make quality easier to discuss.
Explore the workspaceTHE STARTING POINT
A real workflow.
A focused experiment.
An internal engineering assistant can focus on one workflow, such as data-import troubleshooting. Keep the task narrow and review failures before expanding its scope.
ILLUSTRATIVE EXAMPLE · NOT A LIVE MODEL OUTPUT
Help a teammate troubleshoot a failed data import without inventing missing information.
{
"response": "Check the JSONL record on the reported line. Each record needs an instruction and an output. Share the validation error if the format looks correct.",
"next_step": "inspect_validation_error"
}THE WORKFLOW
Make the task repeatable.
Curate useful responses
Include preferred structure, terminology, and examples of uncertainty handled well.
Review response quality
Use exact match for constrained outputs and human review for open-ended responses.
WHAT TO CHECK
Measure the details
that matter.
- Task completion
- Grounded terminology
- Useful clarification questions
- Consistent response format
KEEP EXPLORING