Who maintains the skill layer?

As LLM-powered coding lowers the barrier to building software, a gap is widening between generating code that 'runs' and maintaining it properly in production. The post argues that agentic 'skills' — instructions written for LLM runtimes — are software artifacts with real dependencies and lifecycle costs, not self-maintaining magic. Skills that wrap a library inherit that library's maintenance burden, and the same logic that discourages forking a library applies to forking its skill layer. The recommended split: vendors maintain skills covering their own product surface, while users own the integration layer specific to their environment. Senior engineering craft — handling credentials, sampling before full loads, iterating rather than shipping first drafts — must now be encoded into skills so agents can follow it reliably.