prism-shadow/gdpevo/fewshot-attempt-02

Northstar Structured Reviews

Overview

Use this skill to turn a Northstar task bundle into strict JSON that matches the supplied answer template. The bundle usually includes a prompt, task context, an answer template, and an environment access note.

Workflow

1. Read prompt.txt, task_context.json, and answer_template.json.
2. If environment_access.md is present, use it only to obtain the base URL and bearer token for the shared environment.
3. Use the shared environment endpoints or POST /sql/query to pull only the records needed to answer the template.
4. Treat answer_template.json as the contract. The prompt explains the business question; the template defines the output shape.
5. Return JSON only. Do not add prose, markdown, comments, or extra keys.

Identify the family

• authorization plus PT criteria and document lists -> UM nurse determination.
• appeal_id, drug, and assistance -> pharmacy appeal and manufacturer assistance intake.
• claim_id, benchmark_source, and lines -> claim repricing / payment integrity.
• p2p_id, PET criteria, and missing_pet_factors -> peer-to-peer review.
• period, rows, and a threshold ratio -> margin queue analysis.

Shared rules

• Match field names, nesting, and enum values exactly.
• Preserve required ordering rules from the template.
• Use ISO calendar dates and cents-rounded currency when required.
• Use null only when the schema allows it.
• Keep list order exactly as requested by the schema or source record order.
• If the template forbids extra fields, omit everything not named in the schema.
• Keep basis_audit records in precedence order: controlling records first, then exceptions or exclusions.

Family-specific handling

UM nurse review

• Fill authorization, criteria_results, evidence_documents, excluded_documents, determination_letter, and next_action.
• Prefer current clinical records over stale exports.
• Put the records that directly control the outcome into controlling_record_ids.

Pharmacy appeal and assistance

• Split failures into documented vs insufficient evidence.
• Keep appeal packet items before assistance items.
• Use the assistance status that matches the drug/program fit and the missing fields actually absent.

Claim repricing

• Use the effective benchmark source and version for the claim context.
• Recalculate each line in claim-line order, then the totals and recovery amount.
• Mark stale rate sources as rejected in stale_source_rejected.

Peer-to-peer review

• Resolve PET-IND and PET-FACTOR, then list only unresolved criteria.
• Record whether new patient-specific information materially changed the review.
• If adverse, include the internal appeal deadline.

Margin queue

• Keep rows in the queue row order from the context.
• Flag below-threshold and charge-sensitive rows according to the template rules.
• Identify the top issue and the gap to 120 percent of cost.

Audit trail

• Use actual record IDs from the environment.
• Put controlling IDs first in precedence_record_order, then exceptions and stale or excluded records.
• Do not invent audit IDs or reuse task labels as record IDs.