Cedar Ridge Intake Audit

Cedar Ridge Intake Audit

Use this skill for any task that asks you to audit a batch/roster/program of records in the Cedar Ridge Intake Coordination Portal and return a single JSON object describing readiness, eligibility, risk, blockers, action plans, and cohort counts.

The portal serves several domains that share one procedure:

• New-patient access verification (a roster of patients → insurance / prescription / pharmacy / risk / registration status).
• Referral readiness audit (a referral batch → coding discrepancies, duplicates, shared-insurance anomalies, blockers, action plan, priority tiers).
• Dialysis transfer review (a transfer batch → packet completeness/freshness, capacity feasibility, intake decision, contact routing).
• Referral-to-chart activation (a referral batch → readiness, code discrepancies, chart-activation needs, correspondence queue, priority order).
• Chronic-care enrollment panel (a program's candidates → eligibility, enrollment disposition, reason codes, follow-up cadence, monitoring package).

The vocabulary and output shape differ per task and per domain. The exact allowed enum values, required keys, and ordering rules for the current task are defined only in that task's input/payloads/answer_template.json. Read it first and treat it as the binding contract — never rely on vocabulary remembered from other tasks.

What you must produce

A single JSON object that conforms exactly to the task's answer_template.json. JSON only — no prose, no markdown fences, no commentary. See references/output_assembly.md for the full contract rules and self-check.

Principles

1. The template is the contract. Emit only its required keys, only its allowed enum values, and follow its ordering rules. Inventing a value or omitting a required key is a failure.
2. Filter to scope. Portal list endpoints return all records across every batch/roster/program, including distractors from unrelated batches. Always restrict to the batch_id / roster_id / program_code named in the prompt.
3. Pull every related record. A referral needs its patient, chart, ICD metadata, documents, coverage, and authorization. A transfer needs its patient, packet documents, and facility capacity. A candidate needs its chart and clinical history. A missing related record is the top cause of wrong codes.
4. Dates and service lines come from the environment, not the prompt. A roster's requested_service_date / service_line, a panel's as_of_date, etc., live in portal tables (intake_rosters, etc.) — fetch them.
5. JSON only. The whole response must parse as one JSON object.

Procedure

Phase 0 — Parse the task

From the prompt, identify:

• The scope key and value: a roster_id (new-patient access), a batch_id (referral / transfer batches), or a program_code (enrollment panel).
• The in-scope entity set: an explicit id list (e.g., a set of patient ids), "all referrals in batch <batch_id>", or "all current candidates for <program_code>".
• The domain, which selects the classification logic in references/audit_logic.md.
• The deliverable dimensions the office asks for (these map to the template's sections).

Phase 1 — Read the contract

Read input/payloads/answer_template.json. Extract: required top-level keys; constant/required-value fields; per-list ordering rules; per-item required keys; enum allowed values; summary count keys and value types. Also read environment_access.md for the base URL and the allowed endpoint list. If a payloads/target_roster.json (or similar payload) is present, use it for the patient-id list and notes, but still confirm scope metadata against the portal.

Phase 2 — Connect and gather (filtered)

Use only endpoints from environment_access.md. See references/portal_endpoints.md for response schemas and the SQL endpoint.

• Referral / transfer batch: GET /referrals (or /transfers), then keep only rows whose batch_id equals the scope key. Re-check the filtered count.
• New-patient roster: fetch the intake_rosters row for the roster_id (via POST /query) for requested_service_date / service_line; use the patient-id list from the prompt/payload.
• Enrollment panel: GET /programs/{program_code}/candidates for the candidate list.
• Per entity, pull all related records (patient identity, chart, documents, ICD metadata, coverage, PBM, pharmacy, lifestyle, facility capacity as the domain requires). Use POST /query ({"sql": "..."}) for joins, duplicate detection, shared-insurance grouping, and capacity-on-date lookups.

Phase 3 — Classify and derive

Apply the shared evidence-checking patterns from references/audit_logic.md:

• Completeness (missing records / imaging / documents / chart artifacts).
• Freshness (document age vs. freshness_limit_days).
• Clinical coding consistency (ICD chapter/service_family/narrative/laterality vs. referral service line and reason).
• Duplicate detection (same patient, multiple referrals in one batch).
• Shared-insurance anomalies (same insurance_id across different patients).
• Authorization blockers (pending / denied / not_submitted when required).
• Capacity & feasibility (open chairs on the requested start date × packet readiness).
• Risk (lifestyle + clinical flags → overall risk).
• Eligibility & enrollment (target condition + active diagnosis + consent + chart → enroll / hold / reject; acuity → cadence + monitoring package).
• Readiness taxonomy and priority tiering (urgency × readiness → tier).
• Contact routing / correspondence queue / action plans.

Map the evidence to the exact enum members the current template lists.

Phase 4 — Assemble the JSON

Build one JSON object per references/output_assembly.md:

• All required top-level keys; constants (task_id, batch_id, program_code, …) at their required values.
• Lists ordered per the template (ascending by id unless it says otherwise; unordered sets emitted stably sorted).
• Only allowed enum values; ids uppercase exactly as the portal returns them.
• Summary counts as integers, keyed by the template's members, reconciling to the per-entity lists (include zero-valued keys).
• Empty sets as []; null only where the template permits.

Phase 5 — Self-check

Run the checklist in references/output_assembly.md: required keys present, no extras, all enums allowed, ordering correct, ids cased correctly, counts sum to totals, reference dates from the environment, JSON only. Then emit the JSON object as the entire response.

Files

• references/portal_endpoints.md — endpoint contract, response schemas, the SQL endpoint, gathering rules.
• references/audit_logic.md — the shared classification patterns per domain.
• references/output_assembly.md — the JSON contract rules and self-check.