Teaching Machines to Ask Better Questions: Inside the EBM Query Generator

There's a quiet assumption baked into most "AI for medicine" tools: that the point of the AI is to answer the clinical question. The Vibe Rounds EBM Query Generator flips that assumption on its head. Its tagline says it plainly — "Rx: questions, not answers."

What it actually does

The tool is a companion module to a nine-lesson course called Evidence-Based Medicine for Techies, and the "for Techies" framing is really a simplification device rather than a gatekeeping one. The course was originally taught to a mixed room of medical students and non-medical learners, and it's deliberately built so that both groups come out the other side with an easy, intuitive grasp of EBM — jargon stripped down, concepts translated into plain and often computing-flavored language, so that someone with zero clinical background isn't lost, and a medical student or working clinician gets the same ideas reinforced in a cleaner, less jargon-heavy form than they usually get in med school. The course walks through the full arc of critical appraisal: starting with the patient, moving through PICO question formulation, appraising RCTs, systematic reviews, diagnostic test studies, prognosis and harm studies, and finally clinical practice guidelines and statistics.

The Query Generator takes the workflow taught across those lessons and turns it into something interactive. You paste in a de-identified case — a vignette, a SOAP note, a discharge summary — and pick which kind of clinical decision point the case raises: Therapy, Diagnosis, Prognosis, Harm, Cost-effectiveness, or Guideline check. For each one you select, the tool generates a structured query thread: the best-fitting study design, a full PICOT breakdown, a ready-to-run PubMed search string, and a pointer to the right critical-appraisal checklist (CASP, QUADAS-2, AMSTAR-2, and so on).

Crucially, it stops there. It does not tell you what the answer is. It hands you the scaffolding for finding the answer yourself.

A worked example

A demo run makes this concrete. The case: a woman with type 2 diabetes and newly diagnosed heart failure with reduced ejection fraction, whose own stated priority is "staying out of the hospital and being able to walk my dog without stopping to catch my breath."

From that single case, the tool spins up four parallel query threads:

• Therapy — should an SGLT2 inhibitor be added to her regimen? The tool proposes a full PICOT (adults with HFrEF and T2D; SGLT2 inhibitor vs. placebo/standard care; hospitalization, functional status, and mortality outcomes; 6–24 months), points to systematic reviews of double-blind RCTs as the ideal design, and drafts a Mesh-term-aware PubMed string ready to paste into a search bar.
• Diagnosis — can a BNP or NT-proBNP level be trusted to confirm active heart failure in someone with diabetes? Here the tool correctly identifies that diabetic nephropathy and obesity can confound natriuretic peptide readings, and routes the appraisal toward QUADAS-2.
• Prognosis — what does the literature say about 1-to-5-year outcomes for someone with her risk profile, and what does a cohort study need to control for?
• Harm — what do we actually know about the risk of adverse events like euglycemic DKA or volume depletion in older women started on SGLT2 inhibitors, and what study design is even capable of detecting rare harms?

Then, once at least one query thread exists, two "Closure" steps become available: a Stress-Test, which asks the model to argue against its own conclusions ("what's the strongest argument that this benefit is overstated?") for every thread it generated, and a Restate Patient Priority step, which pulls the conversation back to what the patient actually said she cared about — not a lab value, not a study endpoint, but walking her dog without gasping for air.

Why the constraints matter more than the features

The interesting design decisions here aren't the ones that add capability — they're the ones that remove it.

The tool runs entirely client-side. Your API key lives in your browser's local storage and is sent directly to whichever provider you choose (Claude, Gemini, ChatGPT, or another OpenAI-compatible endpoint) — never through a server the tool's author controls. That's a meaningful privacy choice for anyone even experimenting with real clinical text.

More importantly, the entire premise is Socratic rather than declarative. This is explicit in how the parent course describes itself: it's built in the spirit of AI that questions rather than answers, part of what the author calls a "Clinical Cognition Operating System." Every output carries the same disclaimer, repeated without softening: "The LLM drafts. You verify every number and citation against the source." It is positioned, unambiguously, as a clinical reasoning aid — not for patient care, for personal responsibility and learning only.

That's a genuinely different bet than most "AI copilot" framing in medicine, which tends to optimize for giving a confident-sounding answer fast. This tool optimizes for the opposite: it wants to slow you down at exactly the point where evidence-based medicine usually gets skipped — the step of turning a fuzzy clinical worry into an answerable, searchable, appraisable question.

Who it's for

The "techie" branding is a bit of a misnomer if read too literally — it describes the simplification style, not a restriction on who's welcome. PICO becomes a search schema. RCTs become A/B tests. Likelihood ratios become Bayesian updates. GRADE becomes something like a test-coverage-to-ship-decision framework. Those analogies exist to make the ideas land quickly for anyone without a clinical background, but nothing about the course or the tool excludes medical students or practicing clinicians — if anything, stripping away unnecessary jargon and rebuilding the concepts in plainer language tends to make the underlying logic of EBM clearer for medicos too, not just easier for outsiders. The Query Generator follows the same philosophy: a case is a case, and the tool doesn't ask who's typing it in. A medical student, a practicing physician, and someone with no clinical training at all can all paste in the same de-identified vignette and get the same structured PICOT, search string, and checklist pointer back — the value is in the scaffolding, not in gatekeeping who gets to use it. Paired with a tool that generates the connective tissue between a case and a checklist, that's a plausible way to get anyone — clinical or not — doing real critical appraisal rather than just reading about it.

The honest caveat: this is an educational tool built by one person, wrapped around bring-your-own-key LLM calls, and every generated search string, PICO breakdown, and checklist pointer needs the same scrutiny you'd give a first-year resident's differential. That's not a flaw in the tool — it's the tool's entire thesis.