Skills are code; treat them like code
Skills change behavior — silently, in ways that don't show up in your test suite. The defense: a tiny eval harness. A folder of inputs + expected-output sketches, a script that runs the Skill against each, you eyeball the diff. It's not unit testing, it's regression eyeballing — and that's enough to catch the surprises.
The four moments to run the eval: (1) when you change the Skill body, (2) when you change paths:, (3) when you upgrade the underlying model, (4) before merging a Skill change. Skip any of those and you're tuning blind.