Member-only story
AI Agent Skills at Scale: What Building 170 Skills Across 9 Domains Taught Me About Portability
The AI skills ecosystem is converging in theory and fragmenting in practice. A practical account of what works, what breaks, and what the zero-dependency constraint actually means.
The skill worked perfectly in Claude Code. A marketing content auditor that checked brand voice consistency, flagged cliche phrases, and suggested improvements — tested across three client projects over two weeks. Then one of my engineers tried loading it in Codex CLI.
Note: AI tools helped refine this piece. The skills, testing, architecture decisions, and limitations described are from my direct experience building and maintaining the claude-skills repository.
Nothing broke loudly. The skill loaded. It ran. But the auto-triggering based on file type did not work. The progressive disclosure — loading reference files only when needed — fell back to dumping everything into context at once. Token consumption tripled.
The output quality dropped because the model was drowning in irrelevant instructions meant for other scenarios.
