Build an evaluation from a chat

You don’t have to build an evaluation by hand. In a task, describe what you want to measure, for example:

Set up a benchmark comparing two models on my “Refund requests” dataset, with a pass mark of 75%.

Watch the Juniper Research Desk agent and alert me if its answers start citing no source.

The Evaluation skill decides whether you need a benchmark or a monitor, finds the agent, skill, applet or workflow you named, reuses or drafts criteria, creates the evaluation and gives you a link to it. It tells you the defaults it used, so you can change them. It asks at most two questions, and it shows you its plan and waits for your go-ahead before it creates anything.

To draft or fix criteria on their own, ask for a rubric. That uses the Rubric Builder skill. To make a skill or agent score better against a benchmark, ask to improve it. That uses the Improve skill.