This is a meta-skill for improving ToolUniverse research skills themselves. It codifies 15 patterns learned from building scientific research agents: verify tool schemas before calling them, query foundation databases first, grade evidence by strength (mechanistic beats associational), and structure reports with quantified completeness targets. The critical insight is pattern 14: skills that just list API calls score 3-5/10 in usefulness, while skills that include interpretation tables and synthesis frameworks score 7-9/10. Use this when you're writing a new research skill or debugging why an existing one feels like a data dump instead of an analysis. It also covers computational procedures for when APIs don't exist and you need pandas or scipy to fill the gap.
npx skills add https://github.com/mims-harvard/tooluniverse --skill devtu-optimize-skills