Vital Skills: Upskilling Genie Code with Healthcare and Life Sciences skills

Genie Code, Databricks’ coding agent, enables users to perform data and AI tasks through natural language prompts while adhering to enterprise security and governance rules via Unity Catalog integration. The agent outperforms a leading coding assistant by over twice on real-world data tasks such as machine learning, dashboard creation, and job scheduling. However, its generic approach lacks the domain-specific knowledge required for healthcare and life sciences (HLS) workflows, where clinical standards and methodologies differ significantly from general data science practices.

To address this gap, Databricks has developed Vital Skills, a set of packaged playbooks built on the open Agent Skills standard. These skills embed industry-specific expertise into Genie Code, guiding users on critical steps, thresholds, and tools required for HLS tasks. The skills ensure generated code remains aligned with governed data and Databricks’ ecosystem, avoiding the need for manual setup of serving, GPUs, or scheduling. Vital Skills was tested on BiomniBench, a biomedical benchmark, where Genie Code alone struggled with domain-specific methods but improved markedly when paired with the skills.

Vital Skills was developed by Databricks’ Healthcare and Life Sciences field team, incorporating input from biostatisticians, privacy officers, bioinformaticians, and actuaries. The initial set includes skills for data processing, model lifecycle management, evaluation, and report generation, with plans to expand based on user needs. Skills like semantic-builder automate metric calculations from claims or electronic health records, while cohort-building skills translate plain-language descriptions into reproducible patient groups for research and clinical studies.

The Vital Skills framework also supports deployment of open-source biological foundation models, such as single-cell transcriptomics or protein structure models, by automating infrastructure setup and governance. Testing on a TEDDY single-cell model deployment task showed Genie Code without the skill completed only 19 of 27 readiness checks, while with the skill, all 27 were met. Skills include customizable evaluation harnesses to ensure reliability, and users can chain multiple skills to perform complex workflows like batch inferencing or vector search index creation.