# Recipes

Chains that combine MedSeek tools into the daily jobs of a clinical unit.
Every chain ends with a human review step; none of these outputs is final
until a named clinician signs it.

## Rounding recipe (hospitalist or resident)

Job: turn overnight events and the chart into a defensible daily note.

1. Paste the overnight events, latest vitals, and relevant labs into the
   conversation.
2. Call `progress_note_draft` without `filled_fields` for the SOAP scaffold.
3. Draft each field from what is actually in context. Leave a field empty
   instead of inventing content.
4. Call `progress_note_draft` again with `filled_fields`. Fix the gaps it
   reports; it checks structure only.
5. Run `note_completeness_check` with `checklist: soap` as a second pair of
   eyes on abbreviations and unresolved markers.

For assessment-first ambulatory notes, use `format: apso`. The tool never
assigns an E/M code level; it structures problems addressed, data reviewed,
and risk so a clinician can make that call themselves.

## Admission recipe

Job: assemble the 24-hour H&P without hunting through tabs.

1. Call `admission_history_physical_draft` for the field set.
2. Fill fields from triage notes, outside records, and pharmacy verification.
3. The validator insists on explicit allergies ("no known drug allergies"
   counts) and on stated medication reconciliation status, because those are
   the two silent killers of admission documentation.
4. `data_reviewed` names each external source and result so the E/M data
   element documents itself descriptively.

## Discharge recipe

Job: produce the summary and the patient's plain-language sheet together.

1. `discharge_summary_draft`, two-call loop as above. `pending_results` must
   name each test and its owner, or say "none".
2. `patient_instructions_draft` with `action=scaffold`, draft in plain
   language, then `action=score` until Flesch-Kincaid grade meets target with
   no jargon left.
3. Before asking anything outside the deployment (a consult service, a
   registry search), run the text through `deidentify_text`.

## Evidence recipe (any prescriber)

Job: answer "what does the literature and the label actually say" without
leaving citations behind.

1. `drug_label_lookup` for verbatim label sections with SPL set id and
   DailyMed link.
2. `pubmed_evidence_search` filtered to guideline / systematic review /
   randomized controlled trial as fits the question; keep the reproducible
   query string in the final answer.
3. `trial_search` when enrollment options matter, filtered by recruiting
   status; eligibility criteria are checked on the study page, not by the tool.
4. Compose the answer from those results with source, identifier, and
   retrieval date attached to every claim. A lookup returning nothing means
   "no results found", never "assume from memory".

## What these recipes deliberately avoid

- No diagnosis ranking, severity scoring, or code-level suggestion.
- No EHR access: everything arrives as pasted text the operator chose.
- No identified text leaves the deployment: lookups receive terms and codes,
  and the PHI guard blocks arguments that look like identifiers.