| name | ehr-governance-packets |
| description | Produce normalized JSON for EHR quality-governance / care-coordination packet tasks against a read-only EHR "quality API": duplicate-chart merge-readiness packets, specialty referral coordination letters, care-transition/handoff packets, duplicate-review + service-request validation, and referral-batch audits. Use whenever a prompt hands you case objects (patient IDs, a duplicate candidate ID, a referral or referral-batch ID, a service-request ID, a recipient provider) plus an answer_template JSON schema and asks for a single normalized JSON object reconciling the chart, coding, evidence, provider contacts, and readiness/disposition. Triggers: "merge readiness", "referral coordination packet", "care transition packet", "duplicate review", "ServiceRequest quality", "referral audit", "normalized JSON conforming to answer_template". |
EHR governance / coordination packets
These tasks all share one shape: read a read-only EHR quality API, reconcile
the evidence, and emit one normalized JSON object that conforms exactly to the
provided answer_template. The template — not your prose — is the contract.
Getting these right is mostly about (a) obeying the template's keys/enums/
ordering literally and (b) applying a small set of reconciliation rules
consistently, while ignoring planted distractors.
Procedure
-
Read the answer template first. It lists every required top-level key,
the enum vocabulary for each field, ordering rules, and which arrays are
"sets". Build your output to match its keys and types one-for-one. If it
states a required_value (e.g. a task_id), emit it verbatim.
-
Identify the case objects named in the prompt: patient IDs, duplicate
candidate ID, referral ID / batch ID, service-request ID, the recipient/
specialist provider, service line.
-
Pull every relevant record from the environment's read-only endpoints
before deciding anything: patient demographics, active clinical lists
(conditions / medications / allergies), encounters, immunizations,
disclosures, documents, service-requests, referrals, duplicate candidates,
audit logs, the provider directory, the ICD-10 directory, and the
service-code directory. Fetch the reference directories (ICD-10,
service-codes, providers) so you can validate codes and fill contact
details rather than guess.
-
Trust structured records over free text. coordination_note,
care_plan_notes, a duplicate merge_preview, "confirm X before letter"
hints, and similar prose are non-authoritative and frequently
distractors. If the structured record is complete/valid, treat it as
ready even when a note nags for confirmation; if a preview disagrees with the
patient's own active-list endpoints, the patient endpoints win.
-
Apply the reconciliation rules in references/reconciliation-rules.md
for the packet type at hand (active-key unions, canonical/duplicate
resolution, evidence & document selection, provider selection, ICD-10 and
service-code validation, encounter selection, batch-audit logic, risk flags,
readiness/disposition).
-
Obey ordering and set semantics. Sort exactly as the template says
(alphabetical / ascending by id or code / newest-to-oldest). For
"set-semantics" arrays order is ignored but membership must be exact —
include every correct member and nothing extra.
-
Emit JSON only — no commentary. Correct enums, YYYY-MM-DD dates,
booleans as booleans, null where the schema allows it.
Core reconciliation principles (memorize these)
-
Active means active. Any union/key list of "active" conditions/meds/
allergies = every record whose status == "active", keyed by
normalized_key, unioned across all in-scope patients. Include all active
keys regardless of how generic the key name looks; exclude only non-active
records (inactive, entered-in-error) — those are your excluded/
distractor lists.
-
Structured data beats notes/previews (see step 4). This single rule
resolves most "is it ready?" ambiguities.
-
Report stored status faithfully for duplicates. A duplicate candidate's
status and each patient's canonical_status are the source of truth for
the disposition; don't override them with your own clinical theory.
-
Validate codes, don't eyeball them. Look each ICD-10 / service code up in
the reference directory and compare chapter, laterality, and expected terms to
the narrative and to the patient's own evidence.
-
Distractors are deliberate. Namesake patients, stale/off-topic encounters,
generic chart_summary/ehr_export documents, other patients' audit rows,
inactive/legacy records, and nagging notes are planted to be excluded.
See references/reconciliation-rules.md for the detailed, per-packet rules and
references/output-checklist.md for a pre-submission check.
