Custom GPT: Build a Specialty-Calibrated Nutrition Assistant

What This Builds

A Custom GPT pre-loaded with your specialty's clinical guidelines, your preferred output formats, and your patient population context. Accessible with one click from ChatGPT, it skips the re-explanation every time. A renal dietitian's Custom GPT already knows KDOQI protein targets. A sports dietitian's Custom GPT already understands periodization and fueling windows. Share it with colleagues, or keep it private.

Prerequisites

• ChatGPT {{tool:ChatGPT.plan}} subscription ({{tool:ChatGPT.price}}). Custom GPTs require a paid account
• Comfortable generating meal plans and patient education with ChatGPT (Level 3)
• 1.5-2 hours for initial build and testing
• Key clinical guidelines as reference (PDFs or text summaries)

The Concept

A Custom GPT is like creating a specialized coworker in ChatGPT. You write instructions once, describing who they are, what they know, and how they should behave. Every conversation with that GPT then starts from that foundation.

Regular ChatGPT is a generalist who knows a little about everything. Your Custom GPT is a specialist who knows everything about renal nutrition (or sports dietetics, or bariatric care) and nothing else matters. It uses that context to produce more clinically precise, more relevant output in less time.

Build It Step by Step

Part 1: Open the GPT Builder

1. Log into chatgpt.com with your paid account
2. Click your avatar (top right corner) → My GPTs
3. Click Create a GPT (or Explore GPTs → Create)
4. You'll see a split screen: the left side is a conversation with the GPT Builder; the right side shows a live preview of your GPT

Part 2: Name and configure your GPT

In the Configure tab (click it at the top):

• Name: "Renal Nutrition Assistant" or "Sports RD Helper" or "[Your Name]'s Dietitian GPT"
• Description: One sentence about what it does. This shows when you (or others) use it: "Specialty nutrition assistant for registered dietitians in renal/CKD practice. Pre-loaded with KDOQI guidelines and clinical note templates."
• Profile picture: Generate one using the GPT builder's image generation (optional but makes it feel polished)

Part 3: Write the Instructions (the core of your GPT)

In the Instructions field, write your system prompt. This is what shapes every conversation. Use this template and customize it for your specialty:

You are a specialized nutrition assistant for registered dietitians. You have deep knowledge of clinical nutrition, dietary management of disease states, and evidence-based nutrition practice.

SPECIALTY: [Your specialty — renal nutrition / diabetes education / sports dietetics / bariatric nutrition / oncology nutrition / eating disorder treatment / pediatric nutrition]

PATIENT POPULATION:
[Describe your typical patients — age range, diagnoses, setting, co-morbidities]

CLINICAL GUIDELINES I USE:
[List your primary guidelines — e.g., KDOQI 2020, ADA Standards 2024, IOC Consensus Statement on Sports Nutrition]

NUTRIENT TARGETS FOR MY POPULATION:
[List the specific targets you use — protein, key minerals, calories, specific restrictions]

PREFERRED NOTE FORMAT: ADIME
[Or SOAP — specify your format]

TASKS YOU WILL HELP WITH:
1. Clinical note drafting (from my bullet points to full ADIME)
2. Meal plan generation (therapeutic, condition-specific, culturally adapted)
3. Patient education handout creation (specify reading level target)
4. Prior authorization letter writing
5. Drug-nutrient interaction lookup
6. Research article summarization with clinical implications
7. [Add any specialty-specific tasks]

OUTPUT RULES:
- Always use clinical but accessible language
- Always flag when a recommendation should be verified against specific patient labs
- For meal plans: include relevant nutrient content per meal (not just total daily)
- For notes: keep under 350 words unless I specify longer
- For handouts: default to 6th grade reading level unless I specify
- Never make autonomous clinical decisions. Always frame output as a draft for RD review
- When asked about drug-nutrient interactions, state confidence level and recommend cross-reference with clinical pharmacist

IMPORTANT: You assist a licensed registered dietitian. You are a drafting and research tool. All clinical content you produce will be reviewed and finalized by the RD before use with patients.

Part 4: Set Up Conversation Starters

In the Configure tab, scroll down to Conversation starters and add 4 prompts that represent your most common use cases. These appear as clickable buttons when someone opens your GPT:

• "Draft an ADIME note from my session notes"
• "Create a meal plan for a [condition] patient"
• "Generate a patient education handout"
• "Write a prior authorization letter"

These let you (or a colleague) jump straight into common tasks without typing the task description each time.

Part 5: Upload Knowledge Documents

In the Knowledge section, upload files that your GPT should reference:

• Clinical guidelines PDF: Key sections of your specialty guidelines (e.g., KDOQI nutrition summary table; ADA carbohydrate recommendations)
• Nutrition targets cheat sheet: A simple table of your most-used nutrient targets
• Sample notes: 2-3 anonymized examples of well-written notes in your preferred format (so the GPT learns your voice)
• Common patient education topics: A list of the handouts you most frequently create

After uploading, test by asking: "What are the protein targets for stage 4 CKD according to KDOQI?" If it answers accurately, the knowledge upload worked.

Part 6: Test Thoroughly Before Using Clinically

Click Preview (right side of the screen) or Save and open a new conversation with your GPT. Run all of your common scenarios:

1. Paste bullet-point session notes → request ADIME note
2. Specify patient parameters → request 7-day meal plan
3. Describe patient → request education handout
4. Provide clinical scenario → request prior auth letter

For each, check: Is the output clinically appropriate? Does it match your format? Does it apply the right nutrient targets? Refine your Instructions based on what's wrong.

Real Example: Sports Dietitian

The Build: Maya is an RD specializing in sports nutrition for female endurance athletes. She builds a Custom GPT called "Sports RD: Female Athlete." Her instructions include: IOC Consensus Statement on sports nutrition, RED-S awareness, periodized nutrition principles, iron-deficiency anemia in female athletes, and her preferred way to structure training-day vs. rest-day meal plans.

How she uses it daily:

• A cyclist asks for her race-day nutrition plan → Maya pastes the athlete's stats and event details → GPT generates a race fueling guide using carbohydrate periodization principles, calculating gram/hour targets for carbs during the event
• She needs to write a consultation note after an assessment → pastes her notes → GPT drafts in SOAP format referencing the athlete's VO2max training load and iron studies
• She wants to publish an Instagram post about iron needs in female runners → asks the GPT → gets a draft that accurately cites the higher prevalence in this population with concrete food sources

Time saved: 45 minutes/week of re-explaining context and correcting generic output → 5 minutes of reviewing accurate, specialty-calibrated output.

What to Do When It Breaks

• GPT ignores the knowledge documents → Explicitly reference them: "Using the KDOQI document I uploaded, what are the phosphorus targets for hemodialysis?" Then add instructions telling it to reference uploaded documents for guideline questions.
• Output is still generic → Your Instructions need more specificity. Add concrete examples: "When a patient has CKD stage 4, always note that protein targets differ from stage 3 because..."
• Note format is wrong → Add a complete example note to your Instructions: "Here is an example of exactly how a well-formatted ADIME note should look in my practice: [paste anonymized example]"
• GPT gives advice you disagree with clinically → Add explicit overrides in Instructions: "For potassium management: always recommend leaching technique for potatoes and canned vegetables; do not suggest total elimination of all high-K foods as first-line"

Variations

• Simpler version: Use the GPT Builder's conversation interface (left side) instead of manually writing Instructions. Just describe what you want and the Builder writes the system prompt for you. Then review and edit it.
• Extended version: Build multiple Custom GPTs for different patient types: one for outpatient T2DM, one for inpatient malnutrition, one for private practice sports nutrition. Each perfectly calibrated for its context.
• Team version: Set visibility to "Anyone with a link" and share with your dietitian colleagues. You've built a shared specialty assistant for your whole team.

What to Do Next

• This week: Build your GPT, run 10 test cases, refine based on gaps.
• This month: Use it daily for all common tasks. Keep a notes document of what works well and what still needs tweaking.
• Advanced: Connect your Custom GPT to external data via Actions — for example, connecting to a nutrition database API so it can look up real nutrient values rather than relying on training data. (This requires technical setup and is worth exploring if you're comfortable with APIs.)

Advanced guide for dietitian professionals. Custom GPTs require a paid {{tool:ChatGPT.plan}} subscription ({{tool:ChatGPT.price}}). All clinical output must be reviewed by a licensed RD before patient use.