Skills aren’t policies
Procedure, authority, and enforcement are different concerns.
A skill can teach an agent how to review a rollout. It can describe the evidence to gather, useful tools, common failure modes, and the shape of a good report.
It should not be the only thing preventing that agent from bypassing an SLO.
Skills and policies answer different questions:
- A skill asks: How should capable work be done?
- A policy asks: What is permitted, required, or prohibited?
The distinction matters because instructions are interpreted by the same intelligence whose behavior they are trying to constrain. That is appropriate for judgment and procedure. It is weak protection for hard boundaries.
Put flexible expertise in skills. Put non-negotiable constraints in mechanically enforced policy, permissions, and interfaces. Then let the model exercise wide autonomy between those boundaries.
This is not an argument for rigid workflows. It is the opposite: when the boundary is dependable, the interior can remain dynamic.