SKILL.md Files Are Loader Specs, Not Prompts

Lax Meiyappan digs into the three-level execution architecture behind SKILL.md files and lands on a concrete insight most developers miss: these aren’t static instructions, they’re loader specifications with real cost implications. He restructured a 1,200-line monolithic skill into a 180-line spine with deferred reference files and cut context consumption from 20% to 7%. The piece also surfaces a subtler problem – skills tuned for one model can degrade on another. A writing skill optimized for Sonnet produced choppy output on Opus because the more capable model applied “short sentences” literally instead of judgmentally. With Claude Code, Kiro, Cursor, and Codex CLI all supporting the open SKILL.md standard, understanding the loading mechanics rather than just the prose is becoming a real differentiator in how teams get value from agent tooling.