Clinical Prompting Flowchart

Clinical Prompting Flowchart

Step-by-step guide for building effective LLM prompts in clinical practice — Eduardo Mayorga, MD, 2026

You Have a Patient to Work Up Using an LLM
Follow this sequence: De-identify → Frame Your Prompt (RTF or Domain-Specific) → Context → Differential → Workup → Treatment → Verify
1
De-identify All Patient Data
Before typing anything, strip: names, DOB (use age), MRN/SSN, facility names, room numbers, admission dates (use “Day 1, Day 3”).
Instead of
"John Smith, DOB 03/15/1974, MRN 123456, admitted to Bascom Palmer on 3/28/2026…"
Write
"A 52-year-old male, admitted Day 1 to an academic eye center…"
Decision: How complex is your clinical question?
Quick, focused question → RTF   |   Complex case with many variables → Domain-Specific Framework
Simpler RTF Framework
Use when you have a focused clinical question: a single differential, a quick drug check, a guideline lookup. Three elements are all you need.
ElementYou Write…
RoleSpecialty + expertise level
TaskExactly what you need
FormatHow you want the output
RTF Example
"You are a board-certified retina specialist [Role]. Generate a ranked differential for a 68-year-old with acute painless vision loss and a cherry-red spot on fundoscopy [Task]. Present as a table: Diagnosis | Probability | Supporting Evidence | Must-Not-Miss [Format]."
Best for: Straightforward differentials, quick lookups, pattern-recognition questions, single-system problems.
Detailed Domain-Specific Framework (BRAIN)
Use when the case is complex, multi-system, or high-stakes. BRAIN builds in evidence standards, safety constraints, and patient-specific nuance from the very first prompt.
ElementYou Write…
BackgroundClinical context, reasoning style, and what stage you’re at
RoleSpecific persona (attending, consultant, subspecialist…)
ApproachEvidence standards and CoT method (guidelines to cite, reasoning style)
InstructionsTask details, constraints, safety rules, output requirements
NuanceComorbidities, allergies, organ function, patient preferences
BRAIN Example — Diagnostic Phase
[B] Neuro-ophthalmology consult. 52-yo F with sudden monocular vision loss, APD, and a swollen disc. I need a structured differential before ordering workup. [R] You are a fellowship-trained neuro-ophthalmologist at an academic center. [A] Use Bayesian reasoning. Cite current AAO PPPs and NANOS guidelines. Show probability updates as you weigh each finding. [I] Generate a top-5 differential with: • Estimated probability for each • Key supporting and refuting evidence • Must-not-miss flag • Recommended first-line diagnostic test for each [N] PMH: type 2 diabetes (A1c 8.2), HTN, obesity. Allergy: sulfa. eGFR 45. Current meds: metformin, lisinopril, atorvastatin. Patient is anxious about vision prognosis.
Best for: Multi-system cases, patients with many comorbidities, high-stakes decisions, cases requiring specific guideline citations, unfamiliar subspecialty territory.
↓ Both paths converge here — now load your clinical data ↓
+
Load Structured Clinical Context
Present data the way you would hand off to a senior colleague. Include all that is relevant:
History & Exam Diagnostics Patient Factors
Chief complaint & HPI
Physical exam findings
Relevant PMH, FH, SH
Lab results
Imaging results
Special studies / pathology
Medications & allergies
Renal / hepatic function
Patient preferences
RTF users: Paste your clinical data after the RTF prompt. BRAIN users: Much of this is already in your [B] and [N] elements — add any remaining data here.
3
Generate a Differential Diagnosis
Use Two-Step Prompting to separate analysis from ranking — this reduces premature anchoring:
First Prompt — Analyze
"Analyze the following clinical data. Weigh each finding, identify patterns, and note red flags. Do NOT generate a differential yet. [Paste your de-identified case data here]"
Second Prompt — Rank
"Now generate a ranked differential (top 5–7). For each diagnosis, provide: • Estimated probability • Key supporting evidence • Key refuting evidence • Must-not-miss flag (Y/N) • One atypical presentation to watch for"
Choose a Chain-of-Thought (CoT) reasoning style to include in your prompt:
CoT StyleAdd This to Your PromptBest For
Bayesian “Assign a prior probability, then update it as each piece of evidence is considered.” Differential with labs/imaging
Hierarchical “Reason from broad organ systems down to specific diagnoses.” Undifferentiated presentations
Causal Abduction “Generate hypotheses, then seek confirming AND disconfirming evidence.” Complex / atypical cases (best overall)
Skip CoT Don’t add a reasoning instruction. Simple pattern-recognition, quick look-ups
Tip: CoT can actually reduce accuracy on simple pattern-recognition tasks (NEJM AI, 2025). Match complexity to the question. BRAIN users: you already specified your CoT style in the [A] element.
Append these anti-bias safety lines to every differential prompt:
Safety CheckCopy-Paste This Language
Must-Not-Miss “Always consider dangerous diagnoses that could be fatal if missed, regardless of probability.”
Atypical Presentations “For each top diagnosis, list at least one atypical presentation the clinician should watch for.”
Counter-Evidence “List the strongest evidence AGAINST each of your top 3 diagnoses.”
4
Plan & Iterate the Diagnostic Workup
Use Iterative Stepwise Prompting — request a workup, feed back results, and let the AI refine:
Workup Prompt
"Based on the differential above, recommend a stepwise diagnostic workup: 1. Labs, imaging, and special studies ordered by pre-test probability and cost-effectiveness 2. Account for these patient constraints: [eGFR, pregnancy, contrast allergy, etc.] 3. For each test, state what result would confirm or rule out each top diagnosis"
Iterative Loop — Repeat as Results Come In
Feed-Back Prompt (use each time new results arrive)
"Results are back: [paste new results here]. Update the differential: remove ruled-out diagnoses, adjust probabilities, and recommend the next diagnostic step."
Repeat this feed-back cycle until you reach a working diagnosis or the differential is sufficiently narrowed.
5
Develop the Treatment Plan
If you started with RTF:
Treatment decisions carry the highest implementation risk. Switch to BRAIN now to add the precision layer — evidence standards, safety constraints, and patient-specific nuance — before recommending action.
If you started with BRAIN:
Continue with your BRAIN structure. Update the [B]ackground with workup results and confirmed diagnosis, then refine your [I]nstructions for the treatment task.
BRAIN Framework for Treatment Prompts
ElementWhat to Write at This Stage
Background Summarize the case journey: working diagnosis, key findings, workup results
Role Specific treating persona: attending, subspecialist, clinical pharmacist, etc.
Approach Evidence standards to follow (e.g., “Cite current AAO PPP / AHA / IDSA guidelines”)
Instructions Treatment task + safety rules: dosing, monitoring, interaction checks, contraindications
Nuance Patient-specific factors: comorbidities, allergies, renal/hepatic function, preferences
BRAIN Treatment Prompt Example
[B] Working diagnosis: anterior ischemic optic neuropathy (arteritic) in a 72-yo F. ESR 88, CRP 54, temporal artery biopsy positive. eGFR 38. Current meds: metformin, lisinopril, atorvastatin. [R] You are a neuro-ophthalmologist managing this patient with rheumatology co-management. [A] Follow current AAO and ACR guidelines. Cite specific guideline recommendations. [I] Propose a treatment plan including: • 1st-line, 2nd-line, and alternative therapies with dosing • Renal-adjusted dosing given eGFR 38 • Drug interactions with current medications • Contraindication check against comorbidities • Monitoring parameters and follow-up timeline • Rate your confidence in each recommendation (high/moderate/low) [N] Patient is anxious about long-term steroid side effects. Has a history of GI bleed 2 years ago. Lives alone, limited mobility.
Evidence: Metacognitive prompting (asking the AI to evaluate its own confidence) reduced harmful recommendations by 45% (Esmaeilzadeh). Pharmacist + LLM co-pilot showed strongest medication safety outcomes across 16 specialties.
6
Verify — NEVER Skip This Step
Run these self-critique prompts on the AI output before taking any clinical action:
Ask the AIWhat It Catches
“What diagnoses might I be missing?”Blind spots, rare differentials
“What are the strongest arguments against this treatment plan?”Forces adversarial self-evaluation
“Identify any inconsistencies in the reasoning above.”Logical errors, contradictions
“Are any of the cited guidelines outdated? What is the most current recommendation?”Stale protocol recommendations
“Cross-check the full medication list for interactions I may have missed.”Incomplete drug-interaction screening
Watch for these failure modes: Hallucinated diagnostic criteria • Outdated guidelines • Missed drug interactions • Overconfident probability estimates • Fabricated references • Consistent ≠ Correct
Decision: Are you satisfied with the output?
Yes Output is clinically sound

Apply your clinical judgment. The AI output is a first draft — never a final order. You are always the decision-maker.

No Needs refinement

Feed back corrections, add missing context, or try a different CoT style. Return to the relevant step above.

Remember: Clinical reasoning is iterative. New data at any stage can send you back to update the differential, refine the workup, or adjust the treatment plan.
Safety / Privacy
Differential & Context
Workup & Iteration
Treatment Plan
Verification
Decision Point
Companion to: LLMs as Your Diagnostic and Treatment Assistant — Eduardo Mayorga, MD — 2026