AI Agent Skills - What Building 170 Skills Across 9 Domains Teached Me About Portability

I built 170 AI agent skills across 9 domains over three months. Not because I planned to. Because my team kept needing the same patterns in different tools. The biggest lesson was not about skills. It was about portability. The SKILL.md open standard exists. Adoption is real — Claude Code, Codex CLI, Gemini CLI, Cursor, and others all support it. But "compatible" means different things to different tools: → Auto-triggering works in Claude Code, barely exists elsewhere → Progressive disclosure loads correctly in some tools, not others → Token budgets vary wildly — install too many skills and some silently disappear The engineering decision that paid off most: every Python tool uses only the standard library. No pip install. No dependencies. It runs on any machine with Python 3.8+. The decision cost: some tools are more fragile than their library-dependent alternatives. Honest trade-off. Full practical account — architecture lessons, portability gaps, and what I would do differently: Rea