Agent Skills are becoming an operational knowledge layer rather than a collection of clever prompts. Addy Osmani's repository packages engineering workflows across define, plan, build, verify, review, and ship. PostHog maintains a separate skills collection with official, team, community, and meta categories. The interesting question begins when these libraries contain hundreds of entries.
Scale creates failure modes: overlapping triggers, stale tool names, undocumented permissions, procedures without owners, references that no longer match the product, and giant files that defeat progressive disclosure. A skill needs a bounded job-to-be-done, a narrow description for activation, versioned scripts or references, tests for the fragile steps, and an explicit human gate where judgment remains necessary.
The library itself needs lifecycle metadata: owner, last verified date, supported agents, required tools, risk class, usage, failures, replacement skill, and deprecation state. Progressive disclosure reduces context cost, but it does not maintain correctness. A skill store without review and retirement becomes another dependency registry with executable instructions.