Databricks Adds Open-Source Vital Skills to Genie Code for Healthcare

What Vital Skills are and the problem they address

Vital Skills are packaged playbooks built on the open Agent Skills standard. Databricks says they equip AI coding agents with industry knowledge, so the agents know which questions to ask, which steps are non-negotiable and which tested tools to use. They are designed for Genie Code, the company's coding agent optimized for the Databricks platform, which lets users chat instead of code to perform data and AI tasks.

Databricks argues that general agents write plausible code but lack domain discipline. According to the company, they pick clinical thresholds silently and choose generic statistical methods instead of the specialized ones that are standard in the field. Genie Code is integrated with Unity Catalog, so the agent follows the same security and governance rules as the rest of the platform. Databricks says Genie Code outperforms a leading coding agent by more than twice on real-world data tasks, but concedes that it and other general assistants assume generic methods and thresholds that do not meet HLS standards.

Benchmark results and design principles

Databricks tested Genie Code with and without Vital Skills on BiomniBench, a public benchmark of biomedical data-analysis tasks graded against expert-written rubrics. Each task was run three times and graded blind by an LLM judge. According to the company, Genie Code alone rarely failed on obvious parts such as parsing documents and generating charts, but sometimes fell short on domain-specific methodology and was prone to overassumptions.

With Vital Skills, Genie Code scored higher on the relevant tasks and had no negative impact on tasks outside the skills' scope. Databricks reports improved consistency, accuracy and rigor, with fewer missed steps and checks and fewer retries.

Each skill began with a field engineer who had solved the problem several times with customers. Three principles guided the work:

• Ask before assuming. When a decision belongs to the analyst, such as a clinical threshold, the skill pauses, offers options and lets the user choose.
• Trust-tested tools, not improvised code. For sensitive work, Genie Code runs tested tools and code rather than writing new logic each time.
• Prove it. Every skill was run side by side against Genie Code without it, on the same tasks.

The skill catalog

The starter set spans bioinformatics, clinical work and general tasks, and Databricks expects it to grow. It welcomes contributions to the open-source repository. The skills can be used directly or customized.

• Bioinformatics: bulk-rnaseq, open-weight-models, pathway-enrichment-analysis (ORA or GSEA) and sc-rnaseq for single-cell data from .h5ad files.
• Clinical: cohort-builder, rwe-cohortstudy (comparative effectiveness with matching or IPW), semantic-builder and phi-deidentifier (HIPAA Safe-Harbor de-identification of a Unity Catalog table).
• Others: presentation-builder, report-writer, accelerators and skill-eval, which benchmarks a skill with paired runs with and without it.

The semantic-builder skill builds a semantic layer such as Metric Views using 64 common measures from EHR or claims data. Once defined in Unity Catalog, metrics such as length of stay can be reused across dashboards, Genie Agents, external BI tools and spreadsheets.

Cohort builder and open-weight models

With cohort-builder, an analyst describes a population in plain language, for example "patients with type 2 diabetes and uncontrolled blood sugar", and the skill turns it into a reproducible cohort. It works with coded records and, optionally, free-text notes and scanned PDFs. It checks feasibility first by showing how many patients qualify, and it stops to ask which lab value counts as "uncontrolled". Each patient comes with the reason for inclusion, without exposing raw clinical notes in the shareable result.

Databricks says the skill produced the same cohort on every test run, while Genie Code without it picked a different definition each time. In paired evaluations, free-text cohort F1 rose from 0.62 to 1.00.

The open-weight-models skill targets biological foundation models such as Geneformer, scGPT, Scimilarity and TEDDY, and protein structure models such as AlphaFold/OpenFold and Boltz2. Databricks notes these are difficult to run because they need specialized infrastructure. The skill pulls in the appropriate code, libraries and GPU serving.

What it means for users

Databricks says Vital Skills put the know-how of a biostatistician, privacy officer, bioinformatician or actuary in the hands of every analyst, helping teams move faster before engaging additional expertise. Because the skills run inside the Databricks platform, work prototyped with them can be promoted into scheduled Lakeflow pipelines and jobs, according to the company. The benchmark figures come from Databricks' own testing.