Open-source agent skills for safer, more accurate HCLS AI reasoning

# What the release does

AWS published a collection called HCLS Agent Skills: 38 open-source agent skills that encode clinical and life-sciences decision procedures in structured markdown (SKILL.md) files. Each skill declares triggers, dependencies, and metadata in YAML frontmatter and contains explicit decision frameworks, parameter tables, code patterns, and validation criteria.

The aim is to close a common failure mode where foundation-model agents cite correct guidelines but apply them incorrectly. Examples cited include misclassifying genomic variants by misapplying ACMG/AMP evidence categories, skipping population frequency thresholds, or inventing computational predictor scores. The skills provide the procedural logic practitioners expect, in a machine-consumable format.

# How skills are structured

Skills follow the Agent Skills open standard and are split into two types:

• Reasoning skills: encode methodology and decision frameworks (for example, a genomic-variant-interpretation skill that captures ACMG/AMP evidence categories, frequency thresholds, and predictor cutoffs).
• Pipeline skills: provide tool-specific commands, validated parameters, and code templates (for example, variant-calling skill with GATK4 HaplotypeCaller commands and Mutect2 configurations).

This dual taxonomy supplies both judgment (how to think) and technical precision (how to run tools) so agents can produce correct decisions and runnable artifacts.

Skills are not retrieval-augmented snippets nor model fine-tuning. They are structured prompts that activate based on trigger patterns in a user query and progressively disclose the decision procedure and error conditions to the agent at inference time. The key properties called out are:

• Auditable: the procedures are human-readable in markdown, not hidden in model weights.
• Portable: the same skill files work across many agent runtimes without per-service customization.
• Maintainable: updating a skill is a text-edit, not a retrain.

# Installation and supported harnesses

The repository can be cloned and installed. For skills-only installation there is a universal skills CLI. For specific harnesses there are helper scripts and examples:

• Clone: git clone https://github.com/awslabs/hcls-agent-skills.git
• Skills CLI: npx skills add awslabs/hcls-agent-skills

• AWS Strands Agents SDK: load skills directly in Python with strands.Agent and strands.skills.AgentSkills.

Supported agent runtimes include Amazon Bedrock AgentCore, AWS Strands Agents SDK, Kiro, Amazon Quick Desktop, Claude Code, OpenAI Codex, and more.

# Evaluation and results

AWS reports a 410-prompt evaluation comparing the same agents with and without skills across drug discovery, healthcare operations, and medical imaging workflows. Agents equipped with skills won between 70% and 86% of head-to-head comparisons, with the largest improvement on critical-thinking tasks (78–85% win rate, effect sizes d = 0.65–1.03). The release also provides three worked use cases demonstrating installation and end-to-end behavior across platforms.

# Practical implications for teams

• Use skills when domain procedures matter more than factual recall. They codify thresholds, evidence categories, and error checks that practitioners expect.
• Keep skills as text artifacts to preserve audit trails and to speed policy or protocol updates.
• Combine reasoning and pipeline skills so agents both decide correctly and produce runnable outputs.

# Where to start