extract-health-datalisted

# Extract-Health-Data Turns raw materials into trustworthy, pattern-ready data in two phases: **Phase A** extracts each source faithfully and traceably; **Phase B** compiles cross-source views (timeline, lab-trends, and others) once, so downstream work can spot patterns without redoing the compilation each time. Both phases reorganise data only — they never interpret, conclude, or flag. That is downstream work. ## Use the highest-capability model available Extraction error rates scale inversely with model capability. The cost of one careful extraction with the best model is far smaller than the cost of an investigation built on bad data. Default to the top-tier model available (Opus over Sonnet over Haiku) for this work. Lower-tier models routinely make the silent transcription errors this skill is built to prevent. ## Principles — apply to every extracted item These rules are universal. They apply regardless of source type or output shape. Output shape itself is not prescribed — use whatever structure (table, list, sections) faithfully represents the source while following these principles. - **Verbatim originals are canonical.** The source file is the truth. The extracted markdown is an interpretation. When they disagree, the source wins. - **Loss-less to STRUCTURE, not just to values — the extraction is a representation, never a summary.** Preserve not only every individual value but the source's entire internal *structure*: its ordering, grouping, sectioning,