The Thinking Machine at the Slit Lamp

The difference wasn't the AI. It was the prompt. The attending had configured the AI with an ophthalmic clinical persona, loaded every slit-lamp finding in a structured format, and asked for a ranked differential with supporting and refuting evidence. The resident typed a sentence. This difference — between a vague symptom query and a structured clinical prompt — is now supported by a substantial body of evidence, and it's what this guide is about.

The literature suggests three essential elements: (1) its clinical role — "You are a fellowship-trained ophthalmologist with expertise in uveitis and ocular inflammation. You reason using a systematic anatomical approach"; (2) its behavioral boundaries — "Always consider sight-threatening diagnoses first. Cite evidence when available. Flag diagnostic uncertainty explicitly"; and (3) the output structure — "Provide a ranked differential with supporting and refuting evidence, followed by a stepwise workup." Callens formalized this as the RTF (Role, Task, Format) framework, and research shows it materially changes the quality of clinical output.

Beautifully. The ophthalmic exam is already inherently structured — anatomically layered, quantified (IOP, VA, cell/flare grading), and lateralized. This maps directly to an effective prompt structure: demographics → chief complaint → timeline → pertinent history → exam by anatomical layer → quantified measurements → specific clinical question. Ayoub et al. (2026) validated this approach directly — prompts that mimic the clinician's own case presentation workflow produced human-comparable diagnostic rationales on real clinical data.

This is one of the most nuanced findings in the literature. Chain-of-thought (CoT) prompting — asking the model to reason explicitly — generally improves diagnostic accuracy. Dai et al. found that Causal Abduction CoT (generate hypotheses, then seek confirming and disconfirming evidence) worked best. In our case, you'd want the AI to reason: "Fine KPs suggest non-granulomatous inflammation → supports HLA-B27-associated uveitis, but must rule out herpetic → posterior synechia suggests significant inflammation → elevated IOP could be trabecular inflammation or steroid response..."

However, a 2025 NEJM AI study found that CoT paradoxically reduced performance on tasks requiring intuitive pattern recognition. For ophthalmology, this means: use CoT for complex diagnostic puzzles (atypical uveitis, optic neuropathy workup), but for quick pattern-matching (classic dendritic ulcer, papilledema on fundoscopy), a direct approach may work better.

This is where ophthalmology-specific prompting becomes essential. You'd follow up: "The anatomical diagnosis is anterior uveitis. Now reason through the etiological differential. Consider: the patient's age and sex, the fine (non-granulomatous) KPs, the unilateral presentation, and the posterior synechia. Rank the etiological diagnoses by likelihood and specify what systemic workup would confirm or exclude each." This two-step approach — anatomical diagnosis first, then etiological reasoning — mirrors how uveitis specialists actually think, and structured two-step prompts have been shown to outperform single-step approaches.

LLMs anchor just like humans. HLA-B27 uveitis is the most common cause of acute anterior non-granulomatous uveitis in young adults — but anchoring on it might cause the AI to underweight herpetic uveitis (which can also cause elevated IOP and requires antiviral treatment), Posner-Schlossman syndrome (glaucomatocyclitic crisis), or early sarcoidosis presenting with a non-granulomatous pattern. Practical prompting strategies: ask the AI to "list the three diagnoses most dangerous to miss in this presentation," to "identify findings that could argue against HLA-B27 uveitis," and to "consider infectious etiologies that would change management." Poulain et al. found that CoT prompting reduced anchoring bias — and Zahraei et al.'s BiasMD framework recommends explicitly prompting the model to consider how demographics might shift the differential.

This is a perfect illustration of why clinical context in prompts is non-negotiable. Pregnancy changes everything: chest X-ray requires shielding or deferral, fluorescein angiography is relatively contraindicated, many immunosuppressants are teratogenic, and even topical steroids and IOP-lowering agents require careful selection. If you omit the pregnancy from your prompt, the AI will generate a plan that could harm the patient. The correct prompt includes: "Patient is 10 weeks pregnant. Recommend a workup that avoids ionizing radiation and a treatment plan using only pregnancy-safe medications. Flag any recommendations that require obstetric consultation."

You feed them back: "The following results are now available: HLA-B27 positive, serum ACE normal, RPR non-reactive, chest X-ray deferred due to pregnancy. Update the etiological differential and recommend next steps, including whether additional workup is needed and when." This iterative approach mirrors the real clinical reasoning cycle — data accumulates, the differential narrows, and management evolves. Callens emphasized this as the strongest workflow: presentation → generation → result incorporation → refinement, rather than single-shot queries. The AI becomes a thinking partner you return to, just as you'd update a uveitis consultant with new lab values.

Several things to check here: Prednisolone acetate — generally considered safe topically in pregnancy, but you'll need a rapid taper plan to avoid steroid-response IOP elevation on top of the inflammatory IOP elevation. Cyclopentolate — reasonable for breaking the synechia, though some ophthalmologists prefer atropine for more sustained effect in significant uveitis. Timolol — this is the critical check. Beta-blockers cross the placenta and can cause fetal bradycardia. In a pregnant patient, brimonidine (with caution near term) or a topical carbonic anhydrase inhibitor may be safer — but dorzolamide is a sulfonamide, and she has a sulfa allergy. You could prompt: "Given sulfa allergy and pregnancy at 10 weeks, identify which IOP-lowering drops are safe and which are contraindicated. Explain the reasoning for each." This is exactly the kind of nuanced multi-constraint problem where the AI is useful — and where verification is non-negotiable.

You don't — unless you check. And that's the point. LLMs generate text that looks like a guideline recommendation because they've learned the pattern of what guidelines look like. But they don't retrieve from a database of real guidelines. Chelli et al. documented high rates of fabricated references in AI outputs. The specific recommendation about biologics might be a reasonable clinical approach, but the attribution to a specific AAO guideline could be entirely hallucinated. Never trust an AI-attributed guideline without verifying it against the primary source. This is especially important in ophthalmology, where practice patterns evolve rapidly and subspecialty guidelines vary.

Three powerful self-critique prompts: (1) "What diagnoses might I be missing that would change management?" — this forces consideration of herpetic uveitis (needs antivirals), masquerade syndromes (intraocular lymphoma), and drug-induced uveitis. (2) "What are the strongest arguments against this treatment plan for a pregnant patient with sulfa allergy?" — this surfaces drug safety concerns you might have missed. (3) "Identify any inconsistencies between the recommended treatment and current AAO Preferred Practice Patterns or uveitis society guidelines." — this leverages the AI's training data against its own output. Lucas et al.'s ensemble reasoning approach (generating multiple reasoning paths and checking consensus) outperformed standard approaches by 2–5%.

The clinical findings are identical, but the epidemiological context shifts dramatically. Behçet's disease — which can present with anterior uveitis and is a sight-threatening diagnosis — is far more prevalent in young men along the Silk Road region. If you don't include demographics in your prompt, the AI may not weight this appropriately. But here's the bias paradox: Poulain et al. found that LLMs sometimes over-weight demographics and under-weight clinical findings. The ideal prompt includes demographics as context but explicitly asks the AI to "weight clinical findings above demographic assumptions" and to "consider diagnoses that may be less common but clinically consistent."

Consider: reasoning depth (complex uveitis workup vs. quick refraction check), knowledge recency (does it know the latest anti-VEGF agents and DRCR.net protocols?), context window (can it handle a full surgical history and years of OCT data?), multimodal capability (can it interpret fundus photos or OCT images — and how reliably?), privacy policies (where do your patients' data go?), institutional compliance (is it approved in your hospital system?), and cost. Gaebe et al. showed that even open-source models reached 79% diagnostic accuracy with advanced prompting — so the most expensive tool isn't always necessary for every task.

There's no single right answer — that's the point. For some ophthalmologists, the revelation is that prompt design is a skill that can be learned, not an innate talent. For others, it's the evidence that the same AI that gives you "conjunctivitis" can give you a sophisticated uveitis workup — depending entirely on how you ask. For many, it's the realization that they've been using these tools like a search engine when they could be using them like a reasoning partner who knows the difference between granulomatous and non-granulomatous KPs.

Whatever your answer, the evidence is clear: the ophthalmologists who learn to prompt well will get more from AI than those who don't. And the patients of ophthalmologists who verify will be safer than the patients of those who trust blindly. The thinking machine is powerful — but it still needs the thinking doctor.