Dr. Avinash kumar gupta

Vibe Rounds

Turn any clinical case write-up into a queryable knowledge graph

What this is

A repeatable prompt/runbook. Paste a case link in, get back a self-contained HTML artifact: a node-link knowledge graph of the case (symptoms → exam → investigations → diagnoses → treatment), a day-by-day timeline, and a reasoning-gap layer — a clinician’s-brain pass that adds what’s missing from the documented workup, sourced from general medical knowledge rather than the chart itself.

This file is the spec. Paste it (or its contents) to Claude along with a case URL, and say: “Run Vibe Rounds on this case: <link>“

How to run it

Give Claude the case link (a blog post, a discharge summary, a teaching case PDF — anything with a clinical narrative).
Claude fetches and reads the source.
Claude extracts the case into the structured node model below.
Claude adds the reasoning-gap layer from general clinical knowledge.
Claude builds one self-contained .html file following the design system below and delivers it as a file.

No code should be written before this extraction step — the graph is only as good as the structured read of the source case.

1. Extraction pass — build the node model

Read the case once for narrative, then extract every discrete clinical fact into one of these types. Skip nothing that’s in the chart; don’t invent anything at this stage — that comes later, and it’s a separate, visibly distinct layer.

Type	What goes here	Examples
`symptom`	Presenting complaints, history, comorbidities, risk factors	“Fever ×15d”, “K/C/O CKD”, “Chronic smoker”
`exam`	Positive and pertinent-negative findings on examination	“Shifting dullness +”, “HR 100, BP 110/100”
`invx`	Investigations ordered (not results, unless the result is the notable fact)	“RFT”, “Blood culture”
`dx`	Diagnoses — working, revised, and final — including how a label changes over the admission	“UTI (Day 1–3)”, “Hepatic encephalopathy (Day 5, revised from hypertensive)”
`rx`	Treatments, drugs, interventions, and when they were added/stopped	“Lactulose, added Day 3”

For each node, capture:

a short label (what shows on the graph)
an optional sub-label (dose, timing, qualifier)
a 1–3 sentence description written for a learner, not a copy of the chart line — explain why this fact matters, not just what it says

Then draw edges between nodes wherever the source material implies a causal or reasoning link: symptom → investigation it prompted, investigation → diagnosis it supported, diagnosis → treatment it led to, any node → an outcome. Prefer edges the case itself justifies over ones you’re inferring.

If the case has a clear chronological structure (ward-round notes, daily SOAP entries, an admission-to-discharge or admission-to-death arc), extract a parallel timeline array: one entry per distinct day/stage, each with a short label and 1–2 sentence summary. Flag any point where a diagnosis is revised — that’s usually the most instructive moment in the case.

2. Reasoning-gap pass — add the clinician’s brain

This is the layer that makes it more than a transcript. After extracting what’s documented, reread the case as a clinician would and ask: given these findings, what would a thorough workup normally chase down that never appears here?

Rules for this pass:

Every gap node must be justified by a real clinical link to a documented finding — never invent a gap unrelated to what’s in the chart. Anchor each one to the specific symptom, exam finding, or diagnosis that should have prompted it.
Draw on genuine standard-of-care reasoning: the test that would classify an undifferentiated finding (e.g. SAAG for ascites), the differential that fits the epidemiology but is never named (e.g. TB in an endemic setting), the confirmatory test for a diagnosis that’s asserted without one (e.g. ammonia for “hepatic encephalopathy”), or a reassessment that should have happened as the patient’s course changed.
Never present a gap as a criticism of the treating clinicians. Resource limits, tests pending-but-uncharted, and real-world constraints are invisible in a write-up. Frame each gap as “here’s what this finding would normally prompt,” not “the doctors missed this.”
Keep the count proportionate — roughly 8–15 gap nodes for a typical single-case write-up. More than that stops reading as curated judgment and starts reading as noise.
These nodes must be visually and textually unmistakable as a separate, non-chart layer at every point they appear (graph, panel, legend). See design system below.
Always ship a visible disclaimer near the gap layer stating it is synthesized from general knowledge, not a verified expert review of this specific case, and that it’s a learning prompt, not a clinical critique.

3. Design system — Vibe Rounds visual identity

Keep this exact system across every case run through Vibe Rounds, so the brand stays recognizable case to case. Ground it in a cardiac-monitor / ward-chart aesthetic — dark, clinical, alive — not a generic dashboard.

Palette | Token | Hex | Use | |—|—|—| | --bg | #0a1210 | page background (near-black clinical green-black) | | --bg-panel | #0f1a17 | detail panel background | | --bg-card | #122019 | relationship-item cards | | --line | #22342d | borders, grid | | --ink | #d9e8e0 | primary text | | --ink-dim | #7fa093 | secondary text | | --ink-faint | #4d6b60 | tertiary / labels | | --mon-green | #4fe3a0 | monitor accent, diagnosis nodes, active states | | --amber | #e8a23d | symptom/history nodes, eyebrow text | | --red | #e2626b | reasoning-gap layer only, death/outcome markers | | --violet | #9d8ce8 | investigation nodes | | exam blue | #7fb8e8 | exam-finding nodes | | treatment coral | #e88c6b | treatment nodes |

Typography

Display / headers: IBM Plex Serif (600 weight) — gives the chart a human, editorial voice against all the mono data.
Body / data / UI: IBM Plex Mono — every label, vital, and node reads like an instrument readout.
Never substitute a generic system sans — the mono-for-data / serif-for-narrative pairing is the signature of this brand.

Signature elements

A live monitor strip across the top: a pulsing dot, key vitals as a ticker (age/sex, key vital signs, course length, outcome).
An ECG waveform (SVG polyline) bleeding into the hero header, masked to fade left.
The graph itself as swim-lane columns left→right in clinical logic order: Symptoms/History → Exam → Investigations → Reasoning Gaps → Diagnoses → Treatment. The gap column always sits between Investigations and Diagnoses — it’s literally the missing step in the reasoning chain.
Gap nodes are dashed-outline, low-fill, italic, red, prefixed with ⚠ — never the same visual weight as a documented node.
A day-by-day timeline strip below the graph, dots on a horizontal line, death/adverse outcomes marked in red.

Interaction

Click a node → highlight its direct neighbors, dim everything else, open a detail panel with description and “led to this / this led to” lists.
Graph is horizontally scrollable/pannable (cases vary in width); never compress node count to fit — widen the canvas instead.
Respect reduced-motion; keep the pulse/ECG animations subtle, not distracting from a subject this serious (see tone note below).

Tone note: this is medical content. Keep loading states, copy, and micro-interactions calm and functional — no cute jargon, no celebratory language, no gamification. The one place to spend visual energy is making the reasoning legible, not decorating the interface.

4. Branding

Every artifact produced by this runbook carries the Vibe Rounds mark:

Header eyebrow line includes: Vibe Rounds · Knowledge graph view
Footer credits the original case source (link, author/blog if known, date) and states: “Restructured by Vibe Rounds — a case-based learning tool. Reasoning-gap nodes are synthesized from general clinical knowledge, not a verified expert review. Not a clinical reference.”
Browser tab title format: Vibe Rounds — <short case description>

5. Output contract

Deliver a single self-contained .html file (inline CSS/JS, Google Fonts link only) so it can be opened, shared, or hosted with no build step. File name pattern: vibe-rounds-<short-case-slug>.html

If the person wants to build a library across many cases (not just one), say so explicitly rather than assuming — that’s a different, larger build (a merged multi-case graph with shared entities), not a variant of this single-case runbook.

Usage example

Run Vibe Rounds on this case: http://example-blog.com/2021/03/case.html

Claude should then: fetch → extract → gap-analyze → build the HTML per the design system above → present the file. No need to re-confirm the design system each time; it’s fixed by this document.