AWS publishes 38 open-source skills for more reliable healthcare AI agents

Decision procedures packaged as readable skills

Each skill is a structured SKILL.md document with triggers, dependencies and metadata in YAML frontmatter, followed by material such as decision frameworks, parameter tables, code patterns and validation criteria. The files follow the Agent Skills open standard and use progressive disclosure so an agent can load relevant material at inference time.

AWS divides the collection into reasoning and pipeline skills. Reasoning skills specify how to make judgments: the genomic-variant-interpretation skill, for example, incorporates ACMG/AMP evidence categories, population thresholds and computational-predictor cutoffs. Pipeline skills instead supply validated commands, parameters and templates; the variant-calling skill includes GATK4 HaplotypeCaller, VQSR and Mutect2 configurations.

This makes the approach distinct from both retrieval-augmented generation and fine-tuning. According to AWS, retrieval adds selected source passages, whereas a skill encodes a procedure and its error conditions. The skills are structured prompts activated from query patterns, not knowledge embedded through model training. Because their criteria remain human-readable, organizations can inspect and edit thresholds or protocols without retraining a model.

Deployment patterns and the context tradeoff

The repository can be cloned directly, or the skills alone can be installed with the universal skills CLI. AWS documents integrations for Kiro, Amazon Quick Desktop, the Strands Agents SDK and Amazon Bedrock AgentCore, as well as use with other coding-agent harnesses. The walkthrough lists Python 3.10 or later with uv and Git among its prerequisites.

Quick Desktop selectively activates a skill from the user’s question. AWS illustrates this with a risk-adjustment query in which the agent loads the relevant skill to reason about HCC mappings, hierarchy resolution and RAF differences. Strands developers can attach a directory of skills to an agent in Python, while AgentCore can make skills available to hosted agents at the environment level alongside its managed hosting, scaling, security-boundary and observability capabilities.

Loading all 38 files into one context consumes about 80,000 tokens, according to the post. AWS’s Kiro configuration therefore uses a skill-free coordinator to route requests to eight domain specialists, each carrying roughly 15,000 tokens of relevant skills. This reduces competition from unrelated instructions without requiring users to know which skill to invoke themselves.

Three examples of procedural guidance

In the drug-repurposing example, skills guide an agent evaluating TGFBR1 as a target for idiopathic pulmonary fibrosis. The procedure prioritizes interaction types and evidence sources, ranks direct target engagement above pathway and phenotypic evidence, maps the mechanism to fibrosis processes, and assesses translation using existing safety information, therapeutic-window compatibility and agreement between preclinical models and human disease.

A claims-operations example starts with a Medicare Advantage plan covering 12,000 members. The skilled workflow generates SQL that joins ICD-10 diagnoses to the HCC crosswalk, deduplicates HCCs within the measurement year and resolves Model V28 hierarchies before calculating scores. It also applies demographic segmentation and explains why summing before hierarchy resolution can double-count conditions, inflate RAF scores and create CMS RADV audit exposure.

For a study of 45 healthy adults, the imaging skills specify a T1-weighted MRI preprocessing order: reorientation, bias-field correction, skull stripping and registration to MNI152 space. The ordering matters because uncorrected intensity variation at brain boundaries can distort extraction, especially in temporal and frontal regions. The described output also includes stage-level error checks, visual quality-control artifacts and warnings about orientation metadata, residual shading, included neck tissue and registration failures.

What the 410-prompt evaluation found

AWS evaluated 380 single-skill and 30 cross-skill prompts in two configurations. Kiro CLI used its Auto model selection and had thinking and file-reading tools. The Strands configuration pinned the agent to Claude Sonnet 4.6; both its baseline and skilled conditions received a think tool, while only the latter loaded all 38 skills progressively. Claude Opus 4.7 served as the large-language-model judge.

The judge scored scientific accuracy, coherence, relevance, critical thinking and actionability on a 0–100 scale. Because LLM-judge scores can cluster in a narrow range, AWS emphasized win rate—the share of prompts where the skilled response scored higher—and Cohen’s d, which expresses the mean difference relative to pooled variation.

The strongest signal was critical thinking, supporting AWS’s interpretation that the files primarily contribute methodology: which framework to use, which assumptions to test and which limitations to surface. The different overall results—69.5% in Kiro and 85.9% in Strands—also show that the measured effect is not independent of the surrounding harness.

Benefits were negatively correlated with baseline quality: Pearson’s r was −0.59 for Kiro and −0.61 for Strands. On weak-baseline prompts, the skilled response won 87% of Kiro comparisons and 96% of Strands comparisons, with mean-score gains of 8.7 and 7.9 points. By contrast, strong-baseline prompts produced nearly flat averages: Kiro moved from 91.3 to 91.0 and Strands from 91.0 to 91.1, with win rates of 55% and 54%. One qualification was cross-domain reasoning, which recorded an 80% win rate even against a strong baseline score of 90.2.

AWS also reports lower score dispersion within domains. For clinical-data prompts in Kiro, the judge-score standard deviation fell from 6.8 to 3.3, a 51% reduction. That suggests greater consistency in that slice, but the figures remain results from a pairwise evaluation judged by another language model rather than evidence of clinical outcomes or independently validated patient-safety improvements.

Customization and testing

The repository includes guides for customizing existing files, designing new skills and checking quality before merging. Teams can copy a skill, alter decision thresholds, add internal protocols or remove irrelevant sections. AWS recommends evaluating such changes against prompts drawn from the organization’s actual workflows; the included framework generates pairwise baseline-versus-skill comparisons with the same per-dimension breakdown used in the published evaluation.

Skills are local files and create no AWS resources by themselves. Deployments made through the Amazon Bedrock AgentCore harness do create hosted resources that must be removed separately when no longer needed.

Source: AWS Machine Learning Blog

Definition. AWS healthcare AI agent skills are human-readable SKILL.md files that encode decision procedures, parameters, validation criteria and error checks for specialized workflows.

Key takeaways

FAQ

Sources