A useful skill router must be allowed to return no skill when evidence does not predict a net gain.
Recent studies disagree mainly because they test different exposure regimes, making unconditional semantic matching an unsafe default. Demystifying Agent Skills reports a 6.06 percentage-point advantage for skill artifacts over workflow memory built from the same experience. Its trajectory analysis attributes 65.7 percent of observed skill mechanisms to procedural anchoring and only 4.5 percent to explicit knowledge injection.
The router should compare task evidence with a no-skill baseline, load a narrow procedure only for a diagnosed execution gap, and measure the resulting failure type and token cost. Treat each skill as a versioned policy patch with a trigger, an exit condition, and task-level evaluation against the agent's native behavior. This extends the related New Runtime pattern: This note turns the existing capability-layer thesis into a conditional loading rule.
These results come from bounded benchmark harnesses and do not establish one universal skill policy across models, repositories, or production environments. Revise the router when a skill's measured gain disappears after a model, harness, dependency, or task distribution changes.