Clinical Medication Reconciliation Extraction Prompt

This prompt is designed for clinical pharmacists, care transition teams, and clinical data engineers who need to extract a structured Best Possible Medication History (BPMH) from unstructured clinical notes. It targets admission notes, discharge summaries, transfer documents, and medication administration records. The prompt produces normalized JSON records with medication name, dose, route, frequency, and a reconciliation status flag. It is built to handle PRN medications, tapered regimens, and intentional discontinuations.

This is not a general-purpose clinical extraction prompt. It assumes the input text contains medication information and that the output will be reviewed by a qualified clinician before any clinical decision. The prompt does not perform drug-drug interaction checks, allergy cross-referencing, or therapeutic duplication analysis. It will not catch clinical errors in the source documentation, nor will it flag medications that are contraindicated for the patient. The prompt's job is to faithfully structure what is stated in the text, not to interpret what should have been prescribed.

Use this prompt when you have a defined medication reconciliation workflow and need to convert free-text medication lists into a machine-readable format for downstream systems. Do not use this prompt on text that does not contain medication information, such as radiology reports or progress notes focused solely on physical exam findings. The prompt expects that each input document has already been identified as medication-relevant. If you need to first classify whether a document contains a medication list, pair this with a classification routing prompt before extraction.

The output schema is designed for direct ingestion into medication reconciliation modules, EHR data pipelines, or clinical data warehouses. Each extracted record includes a reconciliation_status field that flags whether the medication is continuing, discontinued, modified, or newly identified. This status is critical for care transition workflows where the goal is to identify discrepancies between what the patient was taking before admission and what is ordered at discharge. Always validate the extracted records against the source text and ensure a clinical pharmacist reviews the output before it is committed to any system of record.

This prompt template is designed to be dropped into an AI application that processes admission or discharge documentation. It forces the model to operate as a strict extraction engine, not a conversational assistant. The template uses square-bracket placeholders for all variable inputs—clinical note text, an optional pre-admission medication list, and a JSON output schema—so you can wire it directly into your data pipeline. The core instruction prohibits the model from inventing medications not present in the source text, which is the most common and dangerous failure mode in clinical extraction.

textCopy
Extract a structured medication list from the clinical note below. Return only valid JSON matching the output schema. Do not invent medications not present in the text. For each medication, determine the reconciliation status by comparing the documented medication against the provided pre-admission medication list if available. Flag any discrepancies.

[INPUT]

[PRE_ADMISSION_MEDICATIONS]

[OUTPUT_SCHEMA]

[CONSTRAINTS]
- Do not add, remove, or modify medications.
- If a field is not explicitly stated in the text, set its value to null.
- For reconciliation status, use only the values: "continued", "discontinued", "dose_changed", "new", "held", or "unable_to_determine".
- If a pre-admission list is not provided, set reconciliation status to "unable_to_determine" for all medications.
- For PRN medications, include the indication in the frequency field if documented.
- For tapered regimens, capture the full taper schedule as a structured array in the dosage_instructions field.
- Flag any intentional discontinuations explicitly mentioned in the note.

To adapt this template, replace each bracketed placeholder with your application's runtime data. [INPUT] should contain the raw clinical note text. [PRE_ADMISSION_MEDICATIONS] should contain a structured list of the patient's medications before admission, or an empty string if unavailable. [OUTPUT_SCHEMA] should be your exact JSON schema definition, including field names, types, and enum values. The [CONSTRAINTS] section is critical for high-risk clinical workflows: it enforces null handling for missing data, restricts reconciliation statuses to a controlled vocabulary, and explicitly requires handling for PRN medications and tapered regimens—two common sources of extraction errors. Before deploying, validate that your schema includes fields for medication name, dose, route, frequency, PRN status, taper schedule, reconciliation status, discrepancy flags, and source text spans for auditability. Always pair this prompt with a JSON schema validator in your application layer and route any output that fails validation to a human review queue.

This prompt is designed to be called from an application layer that manages the full lifecycle: receiving admission or discharge documentation, assembling the prompt with the correct [INPUT_TEXT] and [FORMULARY_CONTEXT], calling the LLM, validating the structured output, and routing low-confidence or unreconciled records for pharmacist review. The prompt itself is stateless—it expects a single clinical document per call and returns a complete medication list. The application is responsible for batching multi-page documents, deduplicating medications across multiple calls, and merging results into the patient's medication history.

Validation and retry logic is critical because missing a discontinued medication or misinterpreting a tapered regimen can cause patient harm. After receiving the JSON output, validate that every medication object contains the required fields (medication_name, status, source_span). Check that status values are strictly from the allowed enum: active, discontinued, held, completed, PRN. For discontinued medications, verify that a discontinuation_reason is present or explicitly null with a confidence note. If validation fails, retry the prompt once with the validation errors appended to [CONSTRAINTS]. If the retry also fails, route the record to a human review queue with the original text, the failed output, and the specific validation errors highlighted. Do not silently drop failed extractions.

Confidence scoring and human review gates should be implemented in the application layer. Parse the confidence field from each medication record. Define a threshold (e.g., confidence < 0.85) below which the medication is flagged for pharmacist review. Additionally, any medication with reconciliation_status: unreconciled or reconciliation_status: conflict must be routed to a human reviewer, regardless of confidence. The application should present the original source text, the extracted medication record, and the specific ambiguity or conflict in a review UI that allows the pharmacist to confirm, edit, or reject the extraction. Log all human corrections to improve future prompt iterations and fine-tuning datasets.

Model choice and context window considerations: This prompt works best with models that have strong instruction-following and structured output capabilities (e.g., GPT-4o, Claude 3.5 Sonnet, or fine-tuned clinical models). The input document plus the prompt template typically consumes 2,000–8,000 tokens. If admission notes exceed the context window, chunk the document by section (e.g., 'Home Medications', 'Hospital Medications', 'Discharge Instructions') and run the prompt on each section independently, then merge and deduplicate by medication_name + source_span. For RAG-enhanced workflows, retrieve the patient's prior medication list and the hospital formulary as [FORMULARY_CONTEXT] to help the model normalize drug names and detect discrepancies. Always log the prompt version, model version, input document hash, and output JSON for audit trails—clinical systems require full traceability for every automated extraction.