AWS open-sourced 38 healthcare agent skills and measured a 51-62% drop in response variance.

awslabs/hcls-agent-skills covers 11 healthcare and life sciences domains under MIT-0. Across 410 domain prompts, skilled agents beat baseline with a 69.5-85.9% overall win rate, 78-85% on critical thinking (Cohen's d 0.65-1.03), 69.3-86.2% on scientific accuracy, and 51-61.9% lower variance in response scores. The skills are structured markdown with YAML frontmatter encoding the decision procedure itself, not retrieval or fine-tuning, and AWS reports they work across 20+ services without customization. The variance reduction is the number I'd sell internally, because it's the one that makes agent output reviewable.