A new open-source collection gives AI agents explicit procedures for 38 healthcare and life-science tasks, from genomic variant interpretation to medical imaging. The goal is to prevent a model from citing the correct guideline while applying its thresholds or evidence categories incorrectly.
Each skill is a readable markdown document with triggers, dependencies, parameter tables, decision rules and validation checks. Reasoning skills describe how to apply a professional framework, while pipeline skills provide tested commands and code templates. The collection spans 11 domains and is released under the permissive MIT-0 license.
In a 410-prompt evaluation reported by AWS, agents equipped with the documents won 70% to 86% of head-to-head comparisons with otherwise similar agents. Gains were strongest on critical-thinking criteria, with win rates from 78% to 85%. The skills can be used by more than 20 agent tools and are easier to audit or update than knowledge embedded in model weights.
These results are not clinical validation, and structured prompts cannot replace qualified review for decisions affecting patients. The approach differs from retrieval, which supplies relevant passages, by encoding the procedure and common errors directly. Its practical advantage is inspectability: a medical expert can review the instructions and revise them when a policy changes without retraining a model.