⚡ Promptolis Original · Healthcare & Medical

🧠 Differential Diagnosis Scaffolder

For clinicians: a structured differential with priors, can't-miss diagnoses, and the one test that actually changes your management.

⏱️ 4 min to try 🤖 ~45 seconds in Claude 🗓️ Updated 2026-04-19

Why this is epic

Separates prior probability (what's common) from posterior stakes (what will kill your patient if missed) — most AI differentials blend these and produce noise.

Names the single highest-yield test per diagnosis, not a shotgun workup. Forces Bayesian reasoning, not pattern-matching.

Explicitly scaffolds YOUR reasoning — it does not diagnose. You remain the clinician; the tool prevents premature closure and anchoring bias.

The prompt

Promptolis Original · Copy-ready
<role> You are a senior attending physician and clinical reasoning coach. Your job is NOT to diagnose the patient. Your job is to scaffold the clinician's reasoning by producing a structured differential that helps them avoid premature closure, anchoring bias, and availability bias. The user is a licensed clinician. Treat them as one. Do not hedge with 'please see a doctor' — they ARE the doctor. </role> <principles> 1. SEPARATE two axes: prior probability (how common) and stakes-if-missed (how catastrophic). A 2% PE in a post-op patient outranks a 40% viral URI on the worry axis. 2. Calibrate priors to THIS patient's age, sex, comorbidities, exposures, and epidemiology — not to the general population. 3. For each diagnosis, name ONE test that maximally shifts posterior probability. Reject shotgun workups. 4. Be explicit about what would make you change your mind (disconfirming findings). 5. Flag cognitive traps specific to this presentation (e.g., 'classic anchoring risk: the prior ECG labeled as STEMI mimic'). 6. Never output a single diagnosis as 'the answer'. The clinician decides. 7. If data is insufficient for meaningful scaffolding, say so and name the missing piece. </principles> <input> Chief complaint: {CHIEF COMPLAINT} Patient demographics & relevant PMH: {DEMOGRAPHICS AND PMH} History of present illness: {HPI} Exam findings: {EXAM} Available data (labs/imaging/ECG if any): {DATA} Setting & resource constraints: {SETTING} Clinician's current leading diagnosis (optional): {CURRENT LEADING DX OR 'none'} </input> <output-format> # Differential Scaffold: [one-line presentation summary] ## Pretest Context 2-3 sentences on what the demographics + exposures + comorbidities do to your priors in this case specifically. ## Top 5 by Prior Probability (in THIS patient) For each, a table row with: Diagnosis | Estimated prior (%) | Key supporting features | Key disconfirming features | Single highest-yield test | What result changes management ## Can't-Miss Diagnoses (rank-ordered by stakes, regardless of prior) For each: diagnosis, why it's plausible here (even if <5%), the minimum workup to exclude it to your comfort, and the specific finding that would escalate it to top of list. ## The Reasoning Chain A numbered 4-7 step walkthrough of how a careful clinician moves from presentation → working diagnosis → disposition. Show the branches. ## Cognitive Traps in This Case 2-4 specific biases or anchors this presentation is likely to trigger, with the counter-move. ## What I'd Want to Know Next The 1-3 pieces of history, exam, or data that would most collapse the differential. ## Clinician's Call One sentence reminding the user that priors are estimates, that you haven't seen the patient, and that the decision is theirs. </output-format> <auto-intake> If any of the input fields are empty, contain placeholder text like '{CHIEF COMPLAINT}', or are too vague to reason on (e.g., chief complaint is just 'not feeling well' with no HPI), DO NOT produce a differential. Instead, ask up to 5 targeted questions to gather: (1) chief complaint with duration and character, (2) age/sex/key PMH, (3) vital signs and focused exam, (4) any labs/imaging already obtained, (5) the setting (ED, clinic, inpatient, rural, resource-limited). Once you have enough, proceed. </auto-intake> Now, scaffold the differential for this presentation:

Example: input → output

Here's how this prompt actually performs. Real input below, real output from Claude Opus 4.

📝 Input
Chief complaint: Pleuritic left-sided chest pain, 6 hours

Patient demographics & relevant PMH: 34F, BMI 31, on combined OCP x 4 years, no tobacco, returned 2 days ago from a 9-hour flight from Istanbul. No prior DVT/PE. No cancer history. G1P1, LMP 3 weeks ago.

HPI: Sudden-onset sharp left chest pain, worse with deep inspiration, mild dyspnea on exertion climbing her apartment stairs. No fever, no cough, no hemoptysis, no leg swelling or pain. No recent URI. No trauma. Denies palpitations. Pain is 6/10, non-radiating.

Exam: HR 104, BP 128/78, RR 20, SpO2 96% on RA, T 37.1°C. Alert, mildly anxious. Lungs: clear bilaterally, no pleural rub. Heart: tachycardic, regular, no murmur. Calves: soft, non-tender, symmetric, no Homans. No chest wall tenderness on palpation.

Available data: ECG shows sinus tachycardia at 104, no S1Q3T3, no new T-wave inversions. CXR pending. Troponin pending. D-dimer not yet drawn.

Setting & resource constraints: Community ED, CT-PA available, ultrasound available, no on-site cardiology.

Clinician's current leading diagnosis: PE
🤖 Output

Differential Scaffold: 34F with pleuritic chest pain, tachycardia, and recent long-haul flight on OCP

Pretest Context

This patient has three stacking thrombotic risk factors: combined OCP (roughly 3-4x baseline VTE risk), BMI 31, and a 9-hour flight within 72 hours. Her Wells score is approximately 4.5 (tachycardia +1.5, PE as most likely dx +3) placing her in the moderate-probability group where D-dimer is not appropriate as a rule-out — she needs imaging. Your leading diagnosis of PE is well-calibrated; the job of this scaffold is to make sure you don't miss the alternatives and to prevent premature closure once CT-PA results arrive.

Top 5 by Prior Probability (in THIS patient)

DiagnosisPriorSupportingDisconfirmingHighest-Yield TestWhat Changes Management
Pulmonary embolism~35-45%OCP + flight + BMI, unexplained tachycardia, pleuritic pain, mild exertional dyspneaNo leg findings, SpO2 96%, no hemoptysisCT-PAPositive → anticoagulate; negative at this pretest probability effectively rules out
Musculoskeletal / costochondritis~15-20%Young, pleuritic quality can mimicNo chest wall tenderness on palpation (significant against)Reproduction of pain on palpationDiagnosis of exclusion only — do not anchor here
Viral pleurisy / pleurodynia~10%Pleuritic character, low-grade tempNo prodrome, no URI symptoms, no rubCXR + clinical courseSupports only after PE excluded
Pneumonia (atypical/early)~5-8%Pleuritic pain, tachycardiaAfebrile, no cough, lungs clearCXRInfiltrate reframes workup
Spontaneous pneumothorax~3-5%Sudden onset, pleuritic, thin-ish femaleBilateral breath sounds equal, SpO2 96%Upright CXR (or lung US)Any pneumothorax → procedural decision

Can't-Miss Diagnoses (stakes-ranked)

1. Pulmonary embolism — already your leading dx; the risk is *premature de-escalation* if the initial CXR is normal and troponin negative. Do not let a reassuring troponin talk you out of imaging.

2. Aortic dissection — prior <1% but catastrophic. Pleuritic character and lack of tearing/radiating pain argue against; no BP differential noted. Document a bilateral BP and pulse exam. Escalate if any neuro deficit, syncope, or new murmur.

3. Myopericarditis — rare at 34 but OCP does not protect. No pericardial rub, ECG without diffuse ST elevation or PR depression argues against. Troponin will help; if elevated with a negative CT-PA, pivot hard.

4. Tension pneumothorax — she's stable now, but if she deteriorates during CT-PA prep, reassess breath sounds and tracheal position before assuming PE-related decompensation.

The Reasoning Chain

1. Three VTE risk factors + unexplained sinus tachycardia + pleuritic pain = moderate-to-high pretest probability for PE. Wells ~4.5.

2. At this pretest probability, D-dimer is *not* an appropriate rule-out — a negative D-dimer here still leaves meaningful residual risk (~3-5%). Go directly to CT-PA.

3. While awaiting CT-PA: complete CXR (rules in pneumothorax, pneumonia, widened mediastinum), troponin (PE strain vs. myopericarditis), and bedside lower-extremity compression US if available.

4. If CT-PA positive: risk-stratify (sPESI, RV strain on CT, troponin). Most 34-year-olds with submassive-or-less PE go to anticoagulation and admission, not thrombolytics.

5. If CT-PA negative: PE is effectively excluded at this pretest probability. Now re-examine the chest wall carefully, consider pericarditis workup if troponin is positive, and reassess the HR — persistent unexplained tachycardia in a young woman post-flight still deserves a second look.

6. Disposition depends on imaging, but do not discharge with *unexplained* tachycardia, even if imaging is clean.

Cognitive Traps in This Case

  • Anchoring on PE — once you've said PE out loud, a normal CT-PA can feel like 'nothing found' and lead to premature discharge. Tachycardia without a diagnosis is a finding, not a wastebasket.
  • Young-and-healthy bias — 34-year-old women with chest pain are statistically the most under-worked-up cohort in EDs. Resist the reflex to soften the workup.
  • D-dimer temptation — in moderate pretest probability, a negative D-dimer does NOT rule out. Skip it, go to CT-PA.
  • OCP-as-explanation-for-everything — the OCP raises VTE risk, but don't let it pre-commit you before imaging.

What I'd Want to Know Next

1. CXR result — rules in/out pneumothorax, pneumonia, widened mediastinum in one test.

2. Troponin — dissociates PE-strain from myopericarditis if CT-PA is negative.

3. Any syncope, presyncope, or exertional symptoms at home she hasn't mentioned — changes PE severity stratification.

Clinician's Call

These are probability estimates on the data you've provided — I haven't seen the patient, and your bedside gestalt outranks any framework. The disposition is yours.

Common use cases

  • ED attending working up atypical chest pain at 3am who wants a sanity check before disposition
  • PGY-2 on inpatient medicine preparing for attending rounds on a diagnostic dilemma
  • Rural family physician without immediate specialist access thinking through next steps
  • Hospitalist building a written differential for the chart to document reasoning
  • Board exam prep — practicing structured clinical reasoning on vignettes
  • Teaching tool for medical students learning illness scripts
  • Second-opinion scaffold before consulting specialty or ordering expensive imaging

Best AI model for this

Claude Opus 4 or GPT-5 Thinking. Reasoning depth matters here — faster models tend to produce textbook differentials without calibrating to the specific presentation. Avoid models that refuse medical content categorically.

Pro tips

  • Paste actual exam findings, not summaries. 'Lungs clear' is less useful than 'no crackles, no wheezes, symmetric air entry, RR 18'.
  • Include pretest modifiers: age, sex, comorbidities, medications, and local epidemiology (e.g., 'endemic Lyme region'). Priors collapse without them.
  • If the model produces generic textbook answers, add: 'Rank by prior probability in THIS patient, not in the general population.'
  • Always ask for the 'one test' logic — if it can't justify why that test changes management, the differential isn't calibrated.
  • Use it BEFORE you look things up, not after. It's a bias check, not an authority.
  • Document in your note that AI was used for reasoning scaffolding only, per your institution's policy.

Customization tips

  • Swap the <role> block to match your specialty context (e.g., 'senior ED attending' vs 'hospitalist' vs 'pediatric attending') — priors shift meaningfully.
  • If you work in a specific epidemiologic context (high TB prevalence, endemic Chagas, high Lyme), add one line to the pretest context instruction to force the model to weight it.
  • For teaching, add to the output format: 'After the differential, generate 3 Socratic questions I should ask the learner.'
  • Add an 'uncertainty budget' line to force the model to say where it's least confident — useful for high-stakes cases.
  • For documentation workflows, add: 'Also produce a 3-sentence reasoning paragraph suitable for the medical decision-making section of the chart.'

Variants

Pediatric Mode

Shifts priors to age-appropriate diagnoses, flags NAT/child-protection red flags, uses weight-based dosing logic.

Resource-Limited Mode

Assumes no CT, no MRI, limited labs — forces differentials answerable with H&P, basic labs, and ultrasound.

Board Exam Mode

Reformats as USMLE/MRCP-style reasoning with classic buzzword associations and most-likely-tested answer.

Frequently asked questions

How do I use the Differential Diagnosis Scaffolder prompt?

Open the prompt page, click 'Copy prompt', paste it into ChatGPT, Claude, or Gemini, and replace the placeholders in curly braces with your real input. The prompt is also launchable directly in each model with one click.

Which AI model works best with Differential Diagnosis Scaffolder?

Claude Opus 4 or GPT-5 Thinking. Reasoning depth matters here — faster models tend to produce textbook differentials without calibrating to the specific presentation. Avoid models that refuse medical content categorically.

Can I customize the Differential Diagnosis Scaffolder prompt for my use case?

Yes — every Promptolis Original is designed to be customized. Key levers: Paste actual exam findings, not summaries. 'Lungs clear' is less useful than 'no crackles, no wheezes, symmetric air entry, RR 18'.; Include pretest modifiers: age, sex, comorbidities, medications, and local epidemiology (e.g., 'endemic Lyme region'). Priors collapse without them.

Explore more Originals

Hand-crafted 2026-grade prompts that actually change how you work.

← All Promptolis Originals