菜单
Healthcare data science rules: encounter-based grain, clinical coding hierarchies, cost allocation, NULL semantics in clinical data.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Load when working with Xata Postgres branches: forking a branch, building or testing dbt models on a branch, wiring dbt to a branch's credentials, diffing two branches, or the pre-merge impact report. Covers create_xata_branch, delete_xata_branch, get_dbt_profile, xata_branch_diff, schema_diff_branches, and pgroll migrations.
Load at Step 1 before exploring the project. Covers output shape inference, incremental model handling, and what to trust in YML.
Use this skill before writing any SQL query. Covers: output shape inference (cardinality clues from the question), efficient schema exploration, iterative CTE-based query building, structured verification loop (row count, NULL audit, fan-out check, sample inspection), error recovery protocol, saving output to result.sql and result.csv, turn budget management, and common benchmark traps.
Populate the knowledge base from dbt project research. Proposes entries across all 6 categories at org, project, and connection scopes.
Load FIRST before any dbt project work. Covers the full 8-step dbt workflow: project scanning, skill loading, validation, macro discovery, research, technical spec, SQL writing, and verification. Also covers output shape inference, incremental model handling, and what to trust in YML.
Loaded at Step 2 for the full workflow. Covers column naming, type preservation, JOIN defaults, lookup joins, sibling models, materialization, packages, and filtering rules.
| name | domain-healthcare |
| description | Healthcare data science rules: encounter-based grain, clinical coding hierarchies, cost allocation, NULL semantics in clinical data. |
When aggregating clinical metrics, drive FROM the fact/event table (encounters, procedures, diagnoses), not the dimension table (patients, providers). Patients with zero encounters MUST NOT appear in utilization reports - they have no data to aggregate.
One patient encounter generates multiple diagnoses, procedures, medications, and cost entries. The grain of clinical fact tables is the EVENT (diagnosis, procedure), not the encounter or patient.
Do NOT aggregate to patient level without first verifying whether the task asks for patient-level or encounter-level output. Premature patient-level aggregation destroys per-encounter detail.
Healthcare coding systems are hierarchical - a parent code rolls up child codes:
When the task asks for a "condition" or "procedure category," check whether the project maps to a specific hierarchy level. Use the mapping table's grain, not a substring of the code.
When multiple codes map to one encounter, the encounter appears multiple times in the fact table - this is correct grain, not a duplicate.
When building a mapping between a complete taxonomy (all codes) and a partial crosswalk (only some codes have mappings), drive FROM the complete taxonomy table. LEFT JOIN the crosswalk onto it. Codes without a crosswalk mapping get NULL - that is correct, they are unmapped codes.
Driving from the crosswalk drops all codes that have no mapping, producing an incomplete taxonomy.
Healthcare costs are allocated across diagnoses and procedures within an encounter. SUM(cost) across all rows for a patient double-counts shared encounter costs.
When the task asks for "total cost," verify whether costs are pre-allocated (sum is correct) or shared (need to deduplicate by encounter first).
In clinical data, NULL means "not recorded," which is clinically different from "not present." A NULL diagnosis does NOT mean the patient is healthy - it means the data is incomplete.
NEVER filter NULL clinical columns unless the task explicitly excludes incomplete records. Filtering NULLs in clinical data silently drops patients with missing documentation.
