Your AI agent skills are dependencies. Start treating them that way.

Agent skills are simple in structure—just markdown files with YAML headers—but their simplicity hides real engineering risk. The most overlooked failure is triggering: the agent reads only the name and description at startup. If those are vague, the skill never runs.

The harder lesson is about content. Most teams let the LLM write skills, producing generic mush. Real expertise comes from documenting gotchas—environment-specific fixes you discover by hand. Simon Willison's maxim applies: keep the domain expertise, let the model do the typing.

Context cost is the silent budget killer. Skill bodies compete for tokens with everything else in the window. The fix is progressive disclosure: split large skills, use a references folder, and keep the main body under 500 lines. For fragile steps, don't let the model guess—write deterministic scripts instead.

The most sobering finding is the security audit: 35% of public skills had flaws, 13% critical. Agent skills are dependencies and should be vetted like any other open-source package. The open standard doesn't guarantee safety.