VibeRounds: Early User Feedback on AI-Augmented Clinical Reasoning

A field report from the first weeks of testing

What is VibeRounds?

VibeRounds is a Socratic AI framework for clinical reasoning education, coined by Dr. Avinash Kumar Gupta in June 2026. The core idea is simple but deliberately unusual: instead of asking an AI to hand over a diagnosis, VibeRounds turns the AI into a Socratic attending — questioning a learner's reasoning, flagging cognitive biases, and withholding the full answer until the learner has committed to their own thinking first.

Built as a library of persona prompts and a larger "Clinical Cognition Operating System" (CCOS) — 57 reasoning modules, four pedagogical frameworks, and 200+ chainable pipelines — VibeRounds is designed to be pasted into any LLM (Claude, Gemini, ChatGPT) alongside a real or invented clinical case. The system enforces a few core constraints: learners must commit to an initial answer before receiving any hint, hints are tiered (framework → direction → partial answer), and low-effort replies get redirected rather than rewarded.

The philosophy is captured in the project's own tagline: "Not AI that answers — AI that questions."

Putting it to the test: a running feedback log

Over the course of about a week (20–27 June), a series of real and case-based clinical scenarios were run through three different LLMs — Claude, Gemini, and ChatGPT — using VibeRounds prompts. Below is a synthesis of what came back.

The good: reasoning over recall

The most consistent theme across entries is that VibeRounds succeeded at its central goal — shifting the interaction from "give me the answer" to "make me think." One early real patient-inspired case (20 Jun) was described as helping to structure clinical thinking, surface overlooked possibilities, and use gentle Socratic nudges instead of direct answers. A stroke case the next day was said to feel "like discussing with a professor," with questions that promoted critical thinking rather than recall and pushed the user to apply previously learned concepts to a real scenario.

This pattern repeated across multiple cases and models:

• A ward-round style Gemini session (21 Jun) was praised for its stepwise reasoning, investigation selection, and management-focused discussion, closely resembling bedside teaching.
• A Claude-run surgical case (23 Jun) made the user feel "like being a surgery resident" — immersive and enjoyable.
• A diabetes-focused Gemini case (24 Jun) went beyond diagnosis into perioperative management, euglycemic DKA, and medication interactions, with the user noting that case-based learning this way "bridges textbook knowledge and patient care" and improves confidence.
• A pulmonary embolism–anticoagulation case (27 Jun) was described as feeling "like solving a clinical puzzle," integrating physiology, pathology, and management across DKA, sepsis, ARDS, and PE — with Socratic questioning said to strengthen clinical reasoning "far better than memorizing isolated facts."

Several sessions also produced concrete, specific learning points users hadn't previously connected — sideroblastic anemia as an ATT complication, enoxaparin dosing and the Cockcroft–Gault formula with female correction factors, perioperative hypoglycemia and delirium management, and revisiting first-year anatomy in a clinical context. One user explicitly called a case their "favorite," citing improved patient-counselling skills as an unexpected bonus.

The mixed and critical: model and prompt sensitivity

Not every run landed. Feedback surfaced two consistent friction points:

1. Model matters as much as the prompt. A Gemini session (21 Jun) stopped after a single answer, possibly because a lighter model variant (Gemini Flash) was used — the user wanted the discussion to go "much deeper and longer." A separate Claude case the same day felt "somewhat odd," with the user wanting more cross-questioning and brainstorming than they received. Later, when a case-3 differential diagnosis exercise improved noticeably after switching from Claude to ChatGPT (23 Jun), the user themselves flagged the ambiguity: was the improvement due to the refined prompt, or simply the different LLM?

2. Prompt iteration made a visible difference. On 22 Jun, a user directly compared an older prompt version to a newer one on the same case type and preferred the newer one — citing a more complete walkthrough, histological differentiation, postoperative therapy discussion, and physiology review as more informative than the earlier version. This suggests the prompt engineering behind VibeRounds' personas is not incidental — refinements measurably changed the depth of the teaching output.

Some entries also came back with no substantive feedback at all (one Gemini case on 21 Jun), a reminder that engagement with this kind of open-ended Socratic tool varies session to session.

What this early log suggests

Across nine days and roughly fifteen logged sessions, three patterns stand out:

1. The Socratic mechanism works when it's given room to work. Users repeatedly described the experience in terms of reasoning — differential-building, management logic, guideline lookup (e.g., needing to check Wells criteria for a PE case) — rather than passive answer retrieval.
2. Output quality is not uniform across LLMs. The same persona prompt produced markedly different depth depending on which model — and which tier of that model — was used, with lighter/faster models sometimes truncating the Socratic exchange prematurely.
3. Prompt versioning is a live variable. The 22 June comparison is early evidence that VibeRounds' prompt design is iterating in a direction users notice and prefer, though isolating "better prompt" from "better model" remains an open question the project will need cleaner A/B testing to resolve going forward.

More Details (Links to sessions by 4 medical students) - https://github.com/avi33tbtt/avi33tbtt.github.io/blob/master/demo/trials/trial3/week1.md

Where to learn more

VibeRounds is openly documented, CC BY 4.0 licensed, and built around a philosophy of "a paradigm, not a prescription" — meaning every persona and module is meant to be copied, adapted, and extended rather than used as-is. The full framework, including the CCOS module library, prompt builder, and courseware, is available at avi33tbtt.github.io.

---

This article is based on an internal user feedback log spanning 20–27 June, covering real and simulated clinical cases run across Claude, Gemini, and ChatGPT.