Physician prompting techniques

A compiled field guide: how working clinicians actually use grounded medical AI (OpenEvidence, Doximity GPT, ChatGPT for Clinicians) beyond plain Q&A. Patterns paraphrased and attributed; skeletons are generalized, not verbatim. Swap the bracketed parts.

1 · Adjudication — turn Q&A into a position

Pean Settle the service debate

Don't ask for evidence — ask for a defensible position, stratified by study quality, written so two specialties can both sign it.

Summarize the highest-quality evidence on [contested practice] in [population]. Stratify by study design. Synthesize a recommendation a [specialty A] and a [specialty B] could both agree on.
IatroX Steelman the other side

After any answer, demand the strongest contradicting evidence. Grounded tools will actually find it.

Find the strongest counterargument or published evidence that would contradict the conclusion above.
IatroX What would change the answer

Surface the decision-shifting variables instead of a static verdict — the clinical version of sensitivity analysis.

List the patient factors or setting constraints that would change this recommendation, and in which direction.
IatroX Guideline delta

Ask where major guidelines disagree and why — the disagreement is usually the informative part.

How do [ACC/AHA vs ESC vs USPSTF…] differ on [question], and what drives the difference?

2 · Output contracts — dictate the shape, not just the question

IatroX Evidence table

Force tabular output with a limitation column — makes weak citations visible instead of buried in prose.

Give me a table: Recommendation | Evidence source | Year | Key limitation | Where it applies.
IatroX Minimum safe plan

Ask for the time-pressured implementation with explicit stop-triggers — converts a review article into a protocol.

If I had to implement this on a busy service, what is the minimum safe plan and what are the stop-triggers?
JMIR '25 Named formal techniques

The academic layer under the folk patterns: few-shot (show an example note/letter first), chain-of-thought ("reason stepwise before answering"), self-consistency (ask 3 ways, keep the agreement), generated knowledge (list relevant facts first, then answer), meta-prompting ("improve this prompt before running it").

3 · Administrative offload — the surprise killer app

Pean Prior-auth appeal with citations in the body

Grounded drafting turns a 30-minute letter into a 5-minute review — and the reviewer can't ignore citations that sit inside the letter.

Draft a prior authorization appeal for [patient one-liner: age, diagnosis, severity grade, failed conservative therapy] denied [procedure]. Cite peer-reviewed evidence supporting [procedure] over [alternative] in this scenario.
Pean Paperwork family

Same pattern covers FMLA/disability letters, work-restriction notes, peer-to-peer prep: state the clinical facts, name the form's audience, ask for sourced claims.

4 · Research & discovery

Pean Trials-map trigger

Simple disease term, one concept per query, then narrow by follow-up ("phase 2/3? enrolling in the US?"). Doubles as a read on where the field's research money is flowing.

Show me clinical trials for [condition] involving [drug]. → Which are phase 2/3 and enrolling in the U.S.?
Pean Manuscript scaffolding

Intro/discussion sections drafted against real literature — grounding removes the fabricated-citation failure mode that makes generic LLMs unusable here.

Draft the introduction for a manuscript on [study question]. Frame the gap, cite the foundational work, end with the study question. [N] words.
Pean ROI evidence base

For clinician-founders: pull the cost-effectiveness and implementation-outcomes literature buyers will ask about.

Compile peer-reviewed evidence on [intervention] vs [comparator] for [outcomes]. Prioritize studies with cost-effectiveness data or quantified ROI.
Pean Objection → query pipeline (tool-chaining)

The meta-technique: a general LLM plays skeptical buyer/investor and generates the hard objections; each objection becomes a grounded-tool query. General model plans, grounded model cites.

[general LLM] Act as a skeptical [CFO / reviewer]. Generate the 5–7 hardest objections to [my claim]. For each answerable with literature, write me a copy-pastable [OpenEvidence] search prompt.

5 · Platform-native habits

OpenEvidence · LiminalMD Scope tight, iterate, keep a library

Vendor guidance converges on the same three habits: specific well-scoped questions beat compound ones; refine by follow-up rather than mega-prompt; and the clinicians getting the most value keep a reusable prompt library organized by workflow (documentation, patient communication, evidence review, billing, admin). Doximity GPT ships specialty-tailored starters for exactly this reason.

Why this list matters beyond the tips. Three patterns run through everything above: clinicians impose output contracts (tables, letters, positions — not answers), they use adversarial framing (counter-evidence, objections, what-would-change), and they chain tools (general model plans, grounded model cites). This curriculum is being written on Substack and in vendor guides — not in journals — and no public benchmark tests any of it: HealthBench grades the answer to a question, not the prior-auth letter, the settled service debate, or the objection-to-query pipeline. The usage era is training its own users, and the evaluation of these workflows exists nowhere but each vendor's telemetry.
Sources
Pean — 10 Ways to Use OpenEvidence You Haven't Thought of Yet (Substack, paid — techniques paraphrased and credited, prompts generalized)
IatroX — OpenEvidence Prompting Playbook
OpenEvidence — How to Prompt Effectively (official guide)
Meskó — Prompt Engineering as an Emerging Skill for Medical Professionals (JMIR 2023)
Prompt Engineering in Clinical Practice: Tutorial for Clinicians (JMIR 2025)
LiminalMD — Doximity GPT vs OpenEvidence
Pean — Clinical AI Faceoff: ChatGPT for Clinicians vs OpenEvidence vs DoxGPT
Part of ai-lens · companion: State of Healthcare AI Benchmarks