On too many skills degrading performance - I've got about 500 in one domain and I don't think count is the problem. Overlap is. Two descriptions that both plausibly match the same situation and the agent picks wrong. That happens at 20 skills just as easily.
What helped was making every skill say where it stops, and hand each excluded case to whichever skill does cover it. Writing down the boundary, not just the trigger.
Bit surprised testing didn't come up. If a skill ships a reference implementation you can test that, and at least check the cross-references resolve. Doesn't prove the prose is right, but it catches the rot you describe in section 3.
What helped was making every skill say where it stops, and hand each excluded case to whichever skill does cover it. Writing down the boundary, not just the trigger.
Bit surprised testing didn't come up. If a skill ships a reference implementation you can test that, and at least check the cross-references resolve. Doesn't prove the prose is right, but it catches the rot you describe in section 3.