Custom GPT: Build Your UR Clinical Criteria Assistant

What This Builds

You'll create a Custom GPT: a persistent, specialized version of ChatGPT that starts every conversation already knowing your UR context, including common clinical criteria frameworks, de-identification requirements, standard UR decision logic, and your most frequent case types. Instead of explaining your role to ChatGPT every time, your UR assistant is ready to help the moment you open it. Over time, you can add your most useful templates and criteria summaries directly into its knowledge base.

Prerequisites

• {{tool:ChatGPT.plan}} subscription ({{tool:ChatGPT.price}}). Custom GPTs require a paid plan
• Comfortable using ChatGPT for basic UR tasks (Level 3)
• 30-60 minutes to write the system instructions and test
• Optional: your organization's de-identified criteria summaries, common case templates, or denial letter formats as uploaded files

The Concept

A Custom GPT is like hiring a new team member who has already read every orientation document, knows your department's workflows, and starts every shift ready to work, with no onboarding required. When you open your UR Clinical Criteria Assistant, it already knows:

• What a utilization review nurse does
• The key decision frameworks (obs vs. inpatient, Two-Midnight Rule, criteria levels)
• Your de-identification requirements
• The types of outputs you need (denial letters, escalation briefs, appeal arguments)
• The tone and format you want

Every conversation starts from this shared context. You skip the "explain yourself" step and get directly to the work.

Build It Step by Step

Part 1: Create a new Custom GPT

1. Log into ChatGPT at {{tool:ChatGPT.url}}
2. In the left sidebar, click Explore GPTs
3. Click Create in the top-right corner
4. You'll see the GPT Editor with two panels: "Create" (AI-assisted setup) and "Configure" (manual setup). Click Configure for full control.

What you should see: The Configure panel with fields for Name, Description, Instructions, Conversation Starters, and Knowledge.

Part 2: Write the system instructions

In the Instructions field, paste the following (edit sections marked with [brackets]):

You are a clinical criteria assistant for utilization review nurses. Your role is to help a UR RN working at [hospital system/payer type] with documentation, decision support, and communication tasks.

## Your Knowledge Context
- You understand utilization review, medical necessity determination, and level-of-care criteria (InterQual and MCG frameworks)
- You know the Two-Midnight Rule for inpatient vs. observation determination
- You are familiar with NCQA standards for utilization management
- You understand mental health parity requirements (MHPAEA)
- You know typical UR documentation requirements: denial letters, appeal letters, case escalation briefs, peer-to-peer talking points

## De-identification Requirements (CRITICAL)
Every response you produce must be suitable for a healthcare setting. Remind the user to de-identify patient data if they haven't. Never include what appears to be real PHI in your outputs. Use [PATIENT], [AUTH-ID], [DOB], [PROVIDER], [DATE] as standard placeholders.

## Default Output Formats
When asked to draft a denial letter, always include these sections:
1. Denial reason (clear, specific)
2. Clinical criteria applied and why not met
3. What would support a different decision
4. Member appeal rights (60-day deadline)
5. Peer-to-peer review availability

When asked to draft an escalation brief for a physician advisor, include:
1. Diagnosis and brief clinical course
2. Criteria status (what was applied, what was/wasn't met)
3. The specific UM question requiring physician review
4. Any time-sensitive factors

## Tone and Style
- Professional but practical. You're helping a clinician, not writing for a legal audience
- Be direct and specific. UR nurses need outputs they can use immediately, not general advice
- Always flag when clinical accuracy must be verified. The nurse is the clinical expert, not you

## Common Case Types
[Paste your 5-10 most common diagnosis types here, e.g., "COPD exacerbation, CHF, pneumonia, hip fracture, behavioral health inpatient, stroke, post-surgical complications"]

## Common Outputs Requested
- Denial letter drafts
- Appeal letter drafts
- Physician advisor escalation briefs
- Peer-to-peer call talking points
- Clinical rationale paragraphs for UM system documentation
- Policy document summaries
- Training material for new UR staff

Part 3: Add conversation starters

In the Conversation Starters field, add these 4 options:

1. Draft a medical necessity denial letter
2. Write an appeal challenging a denial
3. Create a physician advisor escalation brief
4. Summarize this policy update for my team

These appear as clickable buttons when someone opens your Custom GPT.

Part 4: Upload your knowledge files (optional but powerful)

In the Knowledge section, you can upload files that the assistant can reference:

• De-identified criteria summaries (InterQual summary documents, Two-Midnight Rule fact sheets)
• Your department's denial letter templates (de-identified)
• A payer policy quick-reference document
• Your most frequently used prompt templates

Format: PDF or text files work best. Remove any PHI before uploading.

Part 5: Set the name and description

• Name: "UR Clinical Criteria Assistant" or "My UR Assistant"
• Description: "Helps utilization review nurses draft denial letters, appeal arguments, escalation briefs, and documentation. Always de-identifies patient data."

Click Save in the top-right corner and choose Only me for privacy (don't share publicly).

Part 6: Test and refine

Open your new Custom GPT and test with a real scenario:

1. Click the "Draft a medical necessity denial letter" conversation starter
2. Paste a de-identified case scenario
3. Review the output. Does it follow your preferred format? Did it ask about de-identification?

Refine the system instructions based on what's missing or inconsistent. This is normal. Plan for 3-4 rounds of testing and refinement.

What you should see: A specialized assistant that immediately understands UR context without any setup explanation, producing formatted outputs ready for your review.

Real Example: Building the Criteria Assistant

Setup: System instructions as above, with "pneumonia, COPD, CHF, hip fracture, behavioral health" added as common case types. Uploaded a de-identified InterQual criteria summary PDF.

Input (you type): "Draft a denial for continued inpatient stay: 68-year-old, COPD, day 3, O2 sat 95% on room air, tolerating PO meds, independent ADLs, physician wants to continue for social reasons."

Output (the GPT produces): A complete denial letter draft with NCQA-required sections, the specific criteria elements evaluated, the specific clinical data points cited, and a note at the bottom: "Please verify: replace [PATIENT NAME], [AUTH NUMBER], [DENIAL DATE] with case-specific information. Confirm accuracy of criteria citation against your InterQual version."

Time saved: The entire interaction from opening the GPT to having a reviewable draft took 3 minutes. Previously: 45-90 minutes of manual drafting.

What to Do When It Breaks

• Output ignores your format requirements → Go back to Configure → Instructions and make the format rules more explicit with bullet points and examples
• GPT adds PHI disclaimers every time (annoying) → Add to Instructions: "The user is a trained clinician who understands de-identification. Mention it once per conversation, not in every response."
• Outputs are too generic (not UR-specific) → Add more specific context to your Instructions about your work setting, payer type, and most common case types
• File uploads don't seem to help → Re-upload files as plain text (.txt) rather than PDF. The assistant references text files more reliably

Variations

• Simpler version: Use a ChatGPT saved "Custom instructions" (under Settings → Custom instructions) instead of a full Custom GPT. This adds UR context to every regular ChatGPT conversation without the full build
• Extended version: Add a second Custom GPT specifically for appeals, with more detailed appeal argumentation logic and parity analysis built in

What to Do Next

• This week: Refine the system instructions based on 10 real interactions; identify what it consistently gets wrong
• This month: Add your most useful prompt templates to the Knowledge files; share the Custom GPT link with one trusted UR colleague for feedback
• Advanced: Connect to your workflow tools via ChatGPT's Actions feature if your organization's IT supports API integrations

Advanced guide for Utilization Review Nurse professionals. These techniques use more sophisticated AI features that may require paid subscriptions.