AWS open-sources 38 agent skills across 11 HCLS domains

The gap AWS is targeting is not a lack of domain facts. As the post puts it, models have seen the guidelines in training and in the system prompt, but still misapply evidence categories, skip population frequency thresholds, or hallucinate computational predictor scores. The failure mode is silent: outputs look correct but apply wrong criteria, which carries regulatory and patient safety consequences.

The approach AWS takes is deliberately lightweight. Each skill is a structured markdown document (SKILL.md) that declares triggers, dependencies, and metadata in YAML frontmatter, and encodes decision frameworks, parameter tables, code patterns, and validation criteria. Because the decision criteria are human-readable rather than hidden in model weights, annual policy changes or new experiment criteria can be reflected by editing a text file rather than retraining a model. The collection is sorted into reasoning skills, which guide how the agent thinks, and pipeline skills, which supply tool-specific commands and validated parameters.

The evaluation result points to where this kind of specialization pays off. Skills helped most when the base agent struggled, and the correlation between baseline quality and skill benefit was negative in both harness configurations. The open question is how much of the measured gain carries over to each team's own models and workflows — the win rate varied by harness setup, and prompts where the base agent already performed well saw only marginal improvement. For teams in regulated HCLS settings, the more consequential number may be the reduction in score variance, which speaks to consistency rather than average quality.