A skill names the job, binds the tool, sets the data the agent may see, and shows what good output looks like. Take one from the catalog and adapt it to how you work, or add your own.
Residential surcharge billed above contract on 142 lines → dispute, with the contract clause cited.
Trace what an agent does back to a skill.
Set what the skill does
Name the skill, say when to use it and when not to, then set an output schema and add worked examples.
Bind the skill to a tool
Point the skill at a built-in tool, a data-scoped tool, or a webhook. A skill with no tool bound guides the agent but calls nothing.
Scope the data the skill may see
Set allow and deny lists on the tables and fields, plus row filters. The skill cannot read past that scope.
Test the skill, then switch it on
Dry-run the skill from a dialog and regression-test it against a stored library of cases, then switch it on per agent.
Two screens from the support agent and the decision timeline.
A skill sets which records the agent may read for a question and what good output looks like. The output cites those records, and the output schema decides whether it passes, warns, or fails.
A skill records its validation result, latency, and cost on the decision timeline. A skill that starts scoring lower shows up there before the output suffers.
A new surcharge format the examples do not cover. Add a worked example and the score recovers.
Skills ship for the six Operations modules. Adapt one to how you work, or add your own.
We'll pick the catalog skills that cover the job, bind the tools, set the data scope, and dry-run each skill against your own records.
